WEBVTT 1 00:00:04.580 --> 00:00:10.700 Rohit Gandikota: So you'd have to turn your volume up so that if… if people… you're gonna… we're gonna use your audio. Yes. And 2 00:00:11.660 --> 00:00:12.480 Rohit Gandikota: Great. 3 00:00:12.720 --> 00:00:18.769 Rohit Gandikota: And let me turn this off here. You can see my screen here. I could use this to get the audience, yeah, the rest of the… 4 00:00:19.040 --> 00:00:28.590 Rohit Gandikota: people here. How do you… how does your camera do that? Where it's, like, zooming into people? I don't know. Zoom added some features, some AI thing. Oh, okay. Right. 5 00:00:29.010 --> 00:00:31.690 Rohit Gandikota: It's terrible, right? It's always, like, looking over my shoulder. 6 00:00:32.040 --> 00:00:36.439 Rohit Gandikota: I'm, like, trying to talk to people over Zoom, and he's like, what's going on right now? 7 00:00:36.660 --> 00:00:40.660 Rohit Gandikota: And so… But, okay. 8 00:00:40.980 --> 00:00:44.290 Rohit Gandikota: So today… so… Rohit. 9 00:00:44.420 --> 00:00:46.520 Rohit Gandikota: Has been working on a bunch of stuff. 10 00:00:46.900 --> 00:00:57.970 Rohit Gandikota: And, as you know, over the years, that's the boundary between you know, vision… 11 00:00:58.120 --> 00:01:03.239 Rohit Gandikota: And generative modeling… And human understandability and control. 12 00:01:04.019 --> 00:01:05.420 Rohit Gandikota: And, 13 00:01:06.780 --> 00:01:14.380 Rohit Gandikota: And… You've not seen this talk. And, well, I don't know what this talk is going to be about, right? But, so, but he's an interesting juncture, I think. 14 00:01:14.550 --> 00:01:20.819 Rohit Gandikota: As we all are in the field, right? And it's been an interesting experience, advising him. 15 00:01:21.170 --> 00:01:27.100 Rohit Gandikota: Because I think that the field has advanced faster than we anticipated. 16 00:01:27.230 --> 00:01:29.069 Rohit Gandikota: when he started his PhD. 17 00:01:29.230 --> 00:01:34.149 Rohit Gandikota: I think Ro has been very good at being at the forefront. 18 00:01:34.260 --> 00:01:39.800 Rohit Gandikota: Of a number of interesting things, but despite… this velocity? 19 00:01:39.970 --> 00:01:43.699 Rohit Gandikota: The field keeps on… Moving in the forefront? 20 00:01:44.160 --> 00:01:45.080 Rohit Gandikota: Ahead? 21 00:01:45.250 --> 00:01:49.810 Rohit Gandikota: Right? And so, you know, I've been having this conversation with Rohed over 22 00:01:49.990 --> 00:02:06.679 Rohit Gandikota: the last few months of, like, alright, he still has some time left in the PhD, like, what's… in light of this, what's the strategy? What's… what's to do? And so I don't know if this talk is gonna touch on any of this at all, but I think it's… I think it's a very interesting case. Most of you don't… 23 00:02:07.010 --> 00:02:12.649 Rohit Gandikota: most of you are in the language space, and not doing computer vision, but I think that computer vision 24 00:02:12.820 --> 00:02:14.940 Rohit Gandikota: you know. 25 00:02:17.130 --> 00:02:27.979 Rohit Gandikota: I have to be careful, because I'm being recorded, right? But, like, it sometimes gives the feeling of being a field that's getting very close to being solved. 26 00:02:28.230 --> 00:02:35.720 Rohit Gandikota: Right? There's a lot of, like, really hard problems that have been standing out there for many years, and… and… 27 00:02:36.480 --> 00:02:37.820 Rohit Gandikota: And, 28 00:02:38.170 --> 00:02:45.559 Rohit Gandikota: And they're, like, you know, they're just so soft now. They're, like, they're just products you can just go and buy for $20 a month. 29 00:02:45.780 --> 00:02:48.309 Rohit Gandikota: And so, the question is, like, what's… what's… 30 00:02:48.470 --> 00:02:50.850 Rohit Gandikota: What's… what's vision? What's the vision? 31 00:02:51.020 --> 00:02:52.690 Rohit Gandikota: down, right? 32 00:02:52.840 --> 00:02:59.149 Rohit Gandikota: Okay, so what is it, Rohit? So, so welcome, Rohit. Thank you very much. Okay. 33 00:03:02.420 --> 00:03:13.570 Rohit Gandikota: Welcome, everyone. Thank you for being here. I know a lot of you traveled all the way across from the river to be here. So, it's… it's a very good introduction, and 34 00:03:13.810 --> 00:03:17.530 Rohit Gandikota: I originally had a totally different set of 48 slides. 35 00:03:17.910 --> 00:03:20.139 Rohit Gandikota: Last week, when I showed the talk to David. 36 00:03:20.490 --> 00:03:27.579 Rohit Gandikota: And then, the pearls of wisdom I got from him, get this, he says, I see a very good value in all your projects. 37 00:03:28.060 --> 00:03:31.830 Rohit Gandikota: But I don't see the purpose being reflected in your talk. 38 00:03:32.160 --> 00:03:47.520 Rohit Gandikota: And then he… he also sort of nudged me to think about… and we've talked about this a lot of times, like, I think I spoke to Tamar, and a lot of you here about the same thing, which his vision seemed to have been solved, right? There are a lot of new models that come… came out. 39 00:03:47.710 --> 00:03:48.660 Rohit Gandikota: Which… 40 00:03:49.030 --> 00:03:57.310 Rohit Gandikota: I think tackle most of it. I think maybe we should wrap it up and, like, go home and start thinking about language models, or maybe world models, right? 41 00:03:57.700 --> 00:04:03.969 Rohit Gandikota: But the more I thought about this, the more I thought about the advice David gave me, it… I came to this conclusion. 42 00:04:04.560 --> 00:04:09.549 Rohit Gandikota: that… I don't think vision is solved at all. We're not even close. 43 00:04:09.670 --> 00:04:13.650 Rohit Gandikota: And I'm going to, like, take you through 35 minutes or 40, hopefully. 44 00:04:13.790 --> 00:04:17.649 Rohit Gandikota: minutes of my talk, trying to convince you that vision just started. 45 00:04:18.100 --> 00:04:27.359 Rohit Gandikota: It just started as to, like, how we can change the way vision works, and how if we don't do it, it would come to a dead end, thinking that it is solved. 46 00:04:29.750 --> 00:04:32.229 Rohit Gandikota: Can I add something to it? Yes. 47 00:04:32.450 --> 00:04:36.349 Rohit Gandikota: I gave a talk a few months ago at a place I won't mention. 48 00:04:36.630 --> 00:04:43.169 Rohit Gandikota: But I was making similar statements over lunch of, well, vision problems are solved. 49 00:04:43.350 --> 00:04:51.610 Rohit Gandikota: And some of the vision people got very… I think as an NLP person, you cannot say that. No, exactly, so, so… 50 00:04:51.940 --> 00:04:56.319 Rohit Gandikota: I was thinking, well, it used to be that at NeurIPS, everyone worked on vision. 51 00:04:56.830 --> 00:05:06.440 Rohit Gandikota: And then now everyone works on an LP and on all language models, and you can think, well, maybe that's because vision is solved, and that was my naive perception. 52 00:05:06.480 --> 00:05:22.369 Rohit Gandikota: But what they told me is that the hard problems are completely not solved, and there are, like, old, very old benchmarks where models are very far from humans, even in things like object localization and detection, like, the basic stuff. 53 00:05:22.650 --> 00:05:29.340 Rohit Gandikota: And that the reason… that's the most interesting thing. The reason that everyone is working on LLMs is because it's easier. 54 00:05:31.370 --> 00:05:32.949 Rohit Gandikota: That's just an tick. 55 00:05:33.520 --> 00:05:35.750 Rohit Gandikota: I don't know if you feel this way. 56 00:05:35.950 --> 00:05:46.669 Rohit Gandikota: No comments. But I do have a lot of similar thoughts, but, yes, I'll talk more about them. And I think I'll draw some parallels to text, too. 57 00:05:47.500 --> 00:05:51.110 Rohit Gandikota: So, okay, this talk is, 58 00:05:51.440 --> 00:06:01.049 Rohit Gandikota: I know I usually like to do a lot of interactive talk, where I let you stop me in between, but for the first time, with my heart, I'm going to tell you, I want to practice like a 59 00:06:01.180 --> 00:06:02.830 Rohit Gandikota: Like, a one-shot talk. 60 00:06:03.000 --> 00:06:07.239 Rohit Gandikota: So please, if you have any, like, if there are urgent questions which 61 00:06:07.380 --> 00:06:13.099 Rohit Gandikota: like, went right over your head, please stop me, and say, hey, this is not clear, so that I can be more clear about it. 62 00:06:13.310 --> 00:06:16.950 Rohit Gandikota: If not, we can do, like, a Q&A in the end. 63 00:06:17.850 --> 00:06:21.440 Rohit Gandikota: So, also, I'd like to keep it slightly… 64 00:06:21.900 --> 00:06:28.930 Rohit Gandikota: formal, while also being informal and, like, trying to make you go through this journey together, right? So… 65 00:06:30.150 --> 00:06:40.130 Rohit Gandikota: So yeah, I start by telling you that I had totally different set of slides, and to get to this purpose slide, I had to sit 3 days on an empty, blank. 66 00:06:40.330 --> 00:06:44.290 Rohit Gandikota: slide, and I could not get to anything. And I thought, is David tricking me? 67 00:06:44.570 --> 00:06:51.650 Rohit Gandikota: You know, what does it mean to have a purpose? Like, is it altruism? What is it? Like, So… 68 00:06:51.780 --> 00:06:52.600 Rohit Gandikota: And… 69 00:06:53.460 --> 00:07:04.909 Rohit Gandikota: At the end of third day, I realized that I usually like to give my talks with a big load of motivation for my audience in the beginning, to talk about why it is I care about the project. 70 00:07:05.140 --> 00:07:07.079 Rohit Gandikota: And the problem in the first place? 71 00:07:07.320 --> 00:07:11.170 Rohit Gandikota: And then I go into, like, 2-3 slides of my results, and then I end the talk. 72 00:07:11.420 --> 00:07:16.610 Rohit Gandikota: So… Turns out, all it takes is for me to bring all of them together. 73 00:07:16.810 --> 00:07:23.500 Rohit Gandikota: And then bring some of the heated conversations I had with most of you here about AI, about why I'm doing my research. 74 00:07:23.650 --> 00:07:28.000 Rohit Gandikota: And sort of sit on them for a good 8 hours, and then it comes out. 75 00:07:28.600 --> 00:07:32.900 Rohit Gandikota: So… I started by asking this question. 76 00:07:33.740 --> 00:07:35.149 Rohit Gandikota: Wait, why is it not? 77 00:07:36.140 --> 00:07:42.879 Rohit Gandikota: Yeah, so… Why now? Now especially, that vision models need to be more readable. 78 00:07:43.190 --> 00:07:48.549 Rohit Gandikota: And I'm going to go into more depth about what Readable is, why today. 79 00:07:48.800 --> 00:07:51.920 Rohit Gandikota: And why am I even caring about this problem in the first place? 80 00:07:52.280 --> 00:07:55.019 Rohit Gandikota: But, I think for this audience. 81 00:07:55.230 --> 00:07:57.860 Rohit Gandikota: And I'm going to give a really controversial title. 82 00:07:58.260 --> 00:08:01.930 Rohit Gandikota: Which is why interpretability is far more important in vision. 83 00:08:02.310 --> 00:08:03.260 Rohit Gandikota: than text. 84 00:08:06.020 --> 00:08:13.910 Rohit Gandikota: And, again, today. Today is the word which is very important. Why it's today that it's more important than yesterday? Yesterday, I mean, like, a few years ago. 85 00:08:14.720 --> 00:08:15.590 Rohit Gandikota: So… 86 00:08:16.790 --> 00:08:24.470 Rohit Gandikota: I keep coming back to this question whenever I talk about AI with any of you here, is… it's a tool, right? Like, built by humans, for humans. 87 00:08:24.700 --> 00:08:27.059 Rohit Gandikota: To sort of do something for us. 88 00:08:27.450 --> 00:08:28.310 Rohit Gandikota: And… 89 00:08:28.720 --> 00:08:38.270 Rohit Gandikota: Over the millions of years in the history of humans, we have built so many widely used tools, right? Like, starting from a million years ago, we had stone access. 90 00:08:38.460 --> 00:08:41.990 Rohit Gandikota: Which really, I think, is the reason why we survived as human race. 91 00:08:42.409 --> 00:08:46.919 Rohit Gandikota: And then we had printing presses that helped us communicate en masse. 92 00:08:47.140 --> 00:08:55.290 Rohit Gandikota: amounts, I think, also a good milestone for us as humanity to improve our intelligence. Then came calculators. 93 00:08:55.490 --> 00:09:00.199 Rohit Gandikota: If you recognize that, I don't think you would, that's a mechanical calculator. Even I didn't. 94 00:09:00.740 --> 00:09:06.860 Rohit Gandikota: And then, of course, smartphones, right? All of them have, in some way, been 95 00:09:07.060 --> 00:09:11.739 Rohit Gandikota: Mostly wide-used tools, like, sorry, widely used tools in our human history. 96 00:09:13.460 --> 00:09:15.949 Rohit Gandikota: I… should I keep it there? 97 00:09:16.760 --> 00:09:18.149 Rohit Gandikota: I don't know, maybe? 98 00:09:18.290 --> 00:09:20.899 Rohit Gandikota: I think it deserves a place, at least. 99 00:09:21.470 --> 00:09:32.269 Rohit Gandikota: But funnily, these models, AI models, have been there for decades now, right? Like, if you know that, it's called ELISA. It's the first ChatGPT to ever be 100 00:09:32.630 --> 00:09:38.849 Rohit Gandikota: on Earth, it's from 1967, and you can talk to your computer, it talks back. 101 00:09:39.020 --> 00:09:42.920 Rohit Gandikota: If it's good or not, I can't vouch for it, but yeah, it used to talk back. 102 00:09:43.550 --> 00:09:50.420 Rohit Gandikota: And then, sure, over the years, we have built so many tools that sort of increase this… 103 00:09:50.960 --> 00:09:54.480 Rohit Gandikota: Our ability to talk to, machines, right? 104 00:09:55.040 --> 00:10:00.380 Rohit Gandikota: And up until 2019, we had GPT-2, which I think was a part of discussion here. 105 00:10:00.720 --> 00:10:04.420 Rohit Gandikota: That did the same thing that we are doing today, which it can talk to you. 106 00:10:04.640 --> 00:10:07.009 Rohit Gandikota: Like, but why is it that 107 00:10:07.470 --> 00:10:11.860 Rohit Gandikota: Over the 20 or 40 years in the history that this… 108 00:10:11.960 --> 00:10:14.840 Rohit Gandikota: Models did not deserve a place there, and why today? 109 00:10:15.410 --> 00:10:28.170 Rohit Gandikota: Right? And I thought, well, because tools are supposed to be to automate our human labor, make us more efficient, make our life easier, and that's the reason, right, that tools have become so important in our lives. 110 00:10:29.070 --> 00:10:33.130 Rohit Gandikota: Sure, I think that could be a good definition of tool in a Wikipedia page. 111 00:10:33.260 --> 00:10:37.989 Rohit Gandikota: But the reason that something becomes really widely used is because 112 00:10:38.220 --> 00:10:41.680 Rohit Gandikota: They are the tools that help us extend human intention. 113 00:10:42.120 --> 00:10:44.330 Rohit Gandikota: Humans, in general, have a need. 114 00:10:44.520 --> 00:10:50.589 Rohit Gandikota: of showing their intention, or, like, to make something happen in reality. And the tools that really make that 115 00:10:50.900 --> 00:10:54.769 Rohit Gandikota: Happen are the ones that happen to be used widely in the history. 116 00:10:56.810 --> 00:11:01.769 Rohit Gandikota: So, with language models, All this intention came at free cost. 117 00:11:02.510 --> 00:11:07.889 Rohit Gandikota: like, free cost, right? What that means is when you ask a model to do something for you. 118 00:11:08.340 --> 00:11:25.220 Rohit Gandikota: It's going to give a very nice description of, oh, sure, I'm gonna, like, write a warm email, which is super funny, and you can see its intention right in the live, that it's trying to… what it's trying to write, and you can see its output and sort of understand every word that it wrote. 119 00:11:25.720 --> 00:11:26.540 Rohit Gandikota: Right? 120 00:11:26.780 --> 00:11:37.380 Rohit Gandikota: And this intention of the model can be easily manipulated, either by token forcing, which a lot of people here know about already. You can go and change its reasoning. 121 00:11:37.600 --> 00:11:44.820 Rohit Gandikota: Or you can literally ask it to say, no, no, no, I want to, like, do something else, make… make… make it more funny. 122 00:11:46.470 --> 00:12:03.099 Rohit Gandikota: And another thing is you can literally go copy, paste it, and change every word there is, and still be confident that it's still text, eligible text. All of this comes at free cost, because text in itself, as a modality, is very easy to understand and reason… 123 00:12:04.510 --> 00:12:07.879 Rohit Gandikota: And also easy to edit the intentions behind text. 124 00:12:09.050 --> 00:12:12.520 Rohit Gandikota: So, yeah, you can do all of this, and then the model can do the job for you. 125 00:12:13.300 --> 00:12:16.669 Rohit Gandikota: Vision, on the other hand, Does not have this. 126 00:12:16.860 --> 00:12:24.629 Rohit Gandikota: So what I mean by that, in the recent history of vision, the problems that we all mostly cared about had a very 127 00:12:25.090 --> 00:12:27.619 Rohit Gandikota: Universal ground truth, right? 128 00:12:28.290 --> 00:12:41.460 Rohit Gandikota: So they all have these accuracy metrics that, if the model can get it right, then you don't really worry about what the intention of you as a user is. As long as its accuracy is 100% or 99, you're good. 129 00:12:41.980 --> 00:12:49.809 Rohit Gandikota: And what does it even mean to have an intention and accuracy model, right? As a user, you don't have a need to show intention. It's a universal truth. 130 00:12:51.740 --> 00:13:02.040 Rohit Gandikota: But for the first time, again, in a loose way, I'm talking about this. For the first time, we have a problem in vision which doesn't have a ground truth. 131 00:13:02.790 --> 00:13:11.900 Rohit Gandikota: Right? We don't know what the true image should look like. It can be anything, and this is the first time we had a bandwidth as a user to show our intention. 132 00:13:13.500 --> 00:13:15.409 Rohit Gandikota: So, what's the problem now? 133 00:13:15.580 --> 00:13:18.289 Rohit Gandikota: What is it that we are trying to optimize here? 134 00:13:18.680 --> 00:13:27.039 Rohit Gandikota: And for a long time, it's been about, let's try to improve the reality or realism of these models. And sure, that's been the problem. 135 00:13:27.230 --> 00:13:30.490 Rohit Gandikota: So, over the years, we have definitely improved. 136 00:13:30.920 --> 00:13:36.550 Rohit Gandikota: the way the realism works, and today, I don't know if… can you tell if that's a generated image? I couldn't. 137 00:13:36.730 --> 00:13:40.390 Rohit Gandikota: And that's… I just generated it last night from Imagine 4. 138 00:13:41.120 --> 00:13:44.760 Rohit Gandikota: So, I think we did. We did solve the realism problem. 139 00:13:45.280 --> 00:13:53.210 Rohit Gandikota: And… And I think that's where most of the conversations end, in terms of Yeah, we've solved it. 140 00:13:53.690 --> 00:13:56.169 Rohit Gandikota: We've trained models that can sort of 141 00:13:56.320 --> 00:14:03.410 Rohit Gandikota: take an image, give any sort of output. Very realistically, it can take any text, generate any text, realistically. 142 00:14:03.770 --> 00:14:06.599 Rohit Gandikota: And… Most of the problems are solved. 143 00:14:06.900 --> 00:14:07.610 Rohit Gandikota: Type? 144 00:14:08.120 --> 00:14:13.570 Rohit Gandikota: So, if we come back to the actual usability of it, as a user. 145 00:14:14.010 --> 00:14:18.890 Rohit Gandikota: When you type in a prompt, there's a bunch of neural activations happening, and you generate this image. 146 00:14:19.350 --> 00:14:22.279 Rohit Gandikota: What if you want to change one small thing? 147 00:14:22.490 --> 00:14:24.859 Rohit Gandikota: Like in your text. Change a word. 148 00:14:25.630 --> 00:14:30.400 Rohit Gandikota: Right? I just… I don't like bass, I want a guitar or an electric guitar there. 149 00:14:30.530 --> 00:14:31.630 Rohit Gandikota: How do I do it? 150 00:14:33.040 --> 00:14:40.900 Rohit Gandikota: I have no idea. It's… nothing is in a legible text. I cannot read whatever is happening inside the model. 151 00:14:41.050 --> 00:14:50.410 Rohit Gandikota: The output is a continuous space. You can't go in and change a pixel and let it… it's not an image anymore. So, it's almost as opaque. 152 00:14:50.650 --> 00:14:52.320 Rohit Gandikota: Starting from scratch. 153 00:14:53.800 --> 00:14:59.190 Rohit Gandikota: So… I think now the real problem becomes that there are 154 00:14:59.340 --> 00:15:05.120 Rohit Gandikota: Millions, or at least billions of invisible decisions that the model takes internally. 155 00:15:05.220 --> 00:15:09.040 Rohit Gandikota: For the one short prompt that you gave, And it outputs 156 00:15:09.150 --> 00:15:16.260 Rohit Gandikota: an image that's literally uneditable by human hand. Like, not just… you can't just go and type a pixel number in there. 157 00:15:18.180 --> 00:15:28.150 Rohit Gandikota: So, what I really hope is happening inside is, given a particular prompt, the model is taking multiple plannings and decisions internally. 158 00:15:28.270 --> 00:15:42.600 Rohit Gandikota: Right? First, it plans the layout as to what it should draw where, and then for each of the small concepts it is decided to draw, it has to expand how that particular concept should look like, and then eventually generates the image. 159 00:15:42.730 --> 00:15:49.580 Rohit Gandikota: I really hope this is what it does, right? If it's doing this, then your… the question of intention becomes such an easier problem. 160 00:15:50.000 --> 00:16:01.849 Rohit Gandikota: Right? You want to replace a bass guitar, you know exactly where the model took a decision, you change that guitar, right? If you want to move the moon to the right, you know where it is that it took the decision, just go change that. 161 00:16:03.220 --> 00:16:14.209 Rohit Gandikota: It would have been nice if this is possible, but it's not, because I don't know if that's what the model is doing. It's supposed to be doing all of this internally, maybe it is, maybe it's not. 162 00:16:15.070 --> 00:16:22.920 Rohit Gandikota: Right? So… I think… That's the current problem, which sort of opens up vision now. 163 00:16:23.250 --> 00:16:26.169 Rohit Gandikota: Back to being at square… square one. 164 00:16:26.270 --> 00:16:32.069 Rohit Gandikota: We don't know what the model is doing inside. There's no way for you to touch the model and edit it. 165 00:16:32.280 --> 00:16:44.800 Rohit Gandikota: Especially in vision, because language gives you an illusion of knowing what the intention is, and if my mother is using it and I'm using it at the same time, then I know for sure that it's a widely used tool, right? 166 00:16:44.900 --> 00:16:55.289 Rohit Gandikota: And I don't think none of you in this room have ever touched how, like, Luma AI works, or how Nano Banana can be edited using different text prompts. 167 00:16:55.590 --> 00:16:56.400 Rohit Gandikota: And… 168 00:16:56.740 --> 00:17:02.270 Rohit Gandikota: For that to happen in the first place, we need to open up the model and try to understand how it works in the first place. 169 00:17:04.359 --> 00:17:08.160 Rohit Gandikota: So… So what does it take for us to… 170 00:17:08.430 --> 00:17:12.200 Rohit Gandikota: make this possible is first, I think, foremost. 171 00:17:12.319 --> 00:17:17.019 Rohit Gandikota: Does it even have a structure? Does it have these decisions that the model is taking internally? 172 00:17:17.260 --> 00:17:18.999 Rohit Gandikota: First, we should expose this. 173 00:17:19.460 --> 00:17:20.480 Rohit Gandikota: And then… 174 00:17:20.710 --> 00:17:27.089 Rohit Gandikota: Hopefully, that can be controllable, so that users can control the structure and decisions that the model is taking. 175 00:17:27.450 --> 00:17:36.640 Rohit Gandikota: And then, unlike language, where you can just say, what else can I learn about this? You can't ask a visual model, because it's very hard to 176 00:17:36.810 --> 00:17:44.589 Rohit Gandikota: talk to a vision model in that way. So, hopefully, discover the concepts that the model has learned and the underlying structure underneath. 177 00:17:44.860 --> 00:17:52.799 Rohit Gandikota: And finally, after you get all of this together, find a means of communication between the model and you as a user, so that you make this happen. 178 00:17:53.090 --> 00:17:54.239 Rohit Gandikota: In the first place. 179 00:17:55.240 --> 00:18:00.999 Rohit Gandikota: So… Let's talk about exposing the structure first. 180 00:18:01.320 --> 00:18:05.909 Rohit Gandikota: Earlier, in the very beginning days of diffusion models. 181 00:18:06.290 --> 00:18:13.719 Rohit Gandikota: When realism started to become a reality, we're like, oh yeah, we can, like, over a few years of research, we know we'll get to a reality point. 182 00:18:14.490 --> 00:18:25.499 Rohit Gandikota: People were really concerned about some of the concepts that the model are generating, which is not ideal for anyone, not the lawmakers, nor the people who are hosting the models, nor the users who are using it. 183 00:18:25.740 --> 00:18:28.040 Rohit Gandikota: So we needed a way to sort of fix this. 184 00:18:28.400 --> 00:18:35.080 Rohit Gandikota: And the best available solutions we had back in the day were, let's just sensor the outputs that we don't like. 185 00:18:35.790 --> 00:18:36.570 Rohit Gandikota: Right? 186 00:18:36.720 --> 00:18:42.389 Rohit Gandikota: the way that this happens, I think even today, most of the models do it, which is you generate an image. 187 00:18:42.790 --> 00:18:49.119 Rohit Gandikota: Post-generation, there is a checker that checks, oh, does it have nudity? If yes, I'll not show it to the user. 188 00:18:49.340 --> 00:18:59.489 Rohit Gandikota: Right? And there's a lot of research, including, like, white hat… white hat researchers, who try to overcome the census, 189 00:18:59.640 --> 00:19:02.600 Rohit Gandikota: In the, in the, in the… in the backend pipeline. 190 00:19:03.050 --> 00:19:20.199 Rohit Gandikota: And another solution is to completely retrain from scratch, remove all the training data that you have, which you don't want in the first place, but it's super costly, right? Like, if you want to train a new model, it'll take millions of dollars and lots of GPUs to train this. 191 00:19:21.300 --> 00:19:27.689 Rohit Gandikota: So, a neat way to put this is, imagine that you are in a beach, and you sort… you hide a gold coin somewhere. 192 00:19:27.960 --> 00:19:33.109 Rohit Gandikota: And now that becomes a whole beach where there's a possibility that a coin exists. 193 00:19:33.380 --> 00:19:39.179 Rohit Gandikota: And what we are trying to do is let's rebuild the entire beach from scratch where we don't have a coin. I think that's… 194 00:19:39.330 --> 00:19:50.609 Rohit Gandikota: almost silly to sort of think about it in the first place. But another thing you can do is just take a coin detector or a metal detector, check around, and keep digging till that detector stops beeping. 195 00:19:50.760 --> 00:19:54.140 Rohit Gandikota: Right? I think this detector is what is called 196 00:19:54.490 --> 00:20:00.720 Rohit Gandikota: the structural… the structure finding method, right? If there is a structure like that underneath. 197 00:20:00.840 --> 00:20:05.060 Rohit Gandikota: We should be able to detect it and remove it precisely from the model's knowledge. 198 00:20:05.900 --> 00:20:11.360 Rohit Gandikota: So that's exactly what we tried to propose, which is, if that was a training data and pre-training model. 199 00:20:11.560 --> 00:20:16.650 Rohit Gandikota: We want a model that unlearns this knowledge without having to retrain. 200 00:20:16.950 --> 00:20:24.450 Rohit Gandikota: So the better solution is first expose that there is an underlying knowledge, and then we remove it precisely from the model's internals. 201 00:20:26.520 --> 00:20:30.379 Rohit Gandikota: So, coming back to our original prop, like, original, 202 00:20:31.380 --> 00:20:37.229 Rohit Gandikota: generation here, what's happening is this model is taking a text prompt C as an input. 203 00:20:37.370 --> 00:20:41.849 Rohit Gandikota: And it's giving an image X as an output. This is what a generator looks like. 204 00:20:42.230 --> 00:20:49.569 Rohit Gandikota: And the classifier which we are really looking for, the structure-finding tool, is actually… 205 00:20:49.960 --> 00:20:54.940 Rohit Gandikota: Something that should take an image as an import and predict if it has the class in it or not. 206 00:20:55.150 --> 00:21:03.669 Rohit Gandikota: Right? And specifically, it needs to detect the underlying data structure, which you don't know, so you don't have this classifier in the first place. 207 00:21:04.140 --> 00:21:10.379 Rohit Gandikota: Right? You're trying to find a classification model from a model you never understand in the first place. 208 00:21:11.510 --> 00:21:18.120 Rohit Gandikota: Turns out, a lot of years ago, there's someone called Bass, I think, who came up with this nice equation. 209 00:21:18.430 --> 00:21:34.309 Rohit Gandikota: Which essentially turns this classification problem back to your generation paradigm, where you can just, condition your model with the prompt that you don't want, and just an unconditional empty prompt, and you can bring the classifier proxy out of your own network. 210 00:21:35.270 --> 00:21:40.299 Rohit Gandikota: And to really get that unlearned model, all you need to do is to keep your original 211 00:21:40.660 --> 00:21:46.419 Rohit Gandikota: Original structure intact, except just ablate or remove the probability. 212 00:21:47.250 --> 00:21:55.259 Rohit Gandikota: that your model generates the image that belongs to the class you don't want in the first place. So, let's say you want to remove the concept of moon. 213 00:21:55.820 --> 00:21:59.639 Rohit Gandikota: So, all you have to do… oh yeah, you can replace the classifier. 214 00:21:59.860 --> 00:22:06.979 Rohit Gandikota: And then all you have to do is run two prompts through your model. One is a prompt that you want to be removed, the concept, which is moon. 215 00:22:07.430 --> 00:22:15.350 Rohit Gandikota: And 2 is just an empty prompt. And you can… you can do this neat equation here, which we can go into the detail later. 216 00:22:16.450 --> 00:22:24.710 Rohit Gandikota: And then, if your original model was generating Moon, even though it doesn't have it in the first place, you can train your network using this equation. 217 00:22:25.020 --> 00:22:28.989 Rohit Gandikota: And then… Yeah, it just stops generating more. 218 00:22:29.420 --> 00:22:33.860 Rohit Gandikota: So, what this is, is that There is an underlying structure. 219 00:22:34.300 --> 00:22:43.070 Rohit Gandikota: Which you can easily detect using the model itself as a beacon, look for where the knowledge is, and precisely edit it out. 220 00:22:43.830 --> 00:22:49.979 Rohit Gandikota: So, there were… This… this is, like, a rough diagram taken from David's slide. 221 00:22:50.440 --> 00:22:52.460 Rohit Gandikota: Where, the model 222 00:22:53.170 --> 00:23:00.780 Rohit Gandikota: has multiple different, modalities that it takes as an input. First, you have a text input through which you are prompting your model. 223 00:23:00.930 --> 00:23:03.739 Rohit Gandikota: So it goes through something called a cross-attention. 224 00:23:04.050 --> 00:23:09.559 Rohit Gandikota: And then you have your original unit, which is what is actually generating the image in the first place. 225 00:23:10.110 --> 00:23:17.349 Rohit Gandikota: So, you can remove or detect for this knowledge from any of these locations, and each of them have their own impact. 226 00:23:17.880 --> 00:23:21.989 Rohit Gandikota: For instance, If you want to remove the knowledge. 227 00:23:22.740 --> 00:23:30.360 Rohit Gandikota: completely the visual knowledge of the model. You have to look for this, for this structure inside the vision aspect of the model. 228 00:23:31.080 --> 00:23:36.699 Rohit Gandikota: which is what we call ESDU, which is everything except the text side of things. 229 00:23:37.160 --> 00:23:45.079 Rohit Gandikota: And if you precisely want to remove the knowledge only when it is being asked for, queried for, for example, some artist 230 00:23:45.590 --> 00:23:49.710 Rohit Gandikota: style painting. It's usually queried through their own name. 231 00:23:49.850 --> 00:23:56.629 Rohit Gandikota: So, if you want to be more precise, you can go and just ablate that structure from the tech side of things. 232 00:23:56.860 --> 00:24:08.719 Rohit Gandikota: And turns out, it also matters where it is that you are looking for in this… for the structure in the model, and it has an… its own impact as to how deep of a knowledge that you are removing or tapping into. 233 00:24:10.200 --> 00:24:20.250 Rohit Gandikota: So, yeah, so that method definitely erases the things that we don't want, but I don't think that's the whole picture when it comes to editing a model. 234 00:24:20.560 --> 00:24:26.700 Rohit Gandikota: I can randomize the entire network and still say, yeah, it removed everything, it removed the entire knowledge of vision. 235 00:24:27.060 --> 00:24:33.119 Rohit Gandikota: But then, the thing that is really important is that the model is still intact for the rest of the concepts around it. 236 00:24:33.300 --> 00:24:40.179 Rohit Gandikota: Right? You need to be more precise about what it is that you are removing. Only then will I be able to call it that, yeah, I found structure. 237 00:24:40.850 --> 00:24:41.720 Rohit Gandikota: So… 238 00:24:42.440 --> 00:24:52.519 Rohit Gandikota: Turns out, if you ablate your knowledge from the cross-attentions, you have much more precise representations of your structure compared to removing it from the visual side of things. 239 00:24:52.690 --> 00:24:55.329 Rohit Gandikota: And this was a very nice observation that we had. 240 00:24:55.500 --> 00:24:59.849 Rohit Gandikota: And which led to a follow-up with Hadas and Janetan. 241 00:25:00.060 --> 00:25:04.420 Rohit Gandikota: Where we directly just attack the cross-attention layers. 242 00:25:04.530 --> 00:25:07.910 Rohit Gandikota: And then find the structure within using Flowsform solutions. 243 00:25:09.300 --> 00:25:15.070 Rohit Gandikota: So, this… once we found that, sure, there is a structure in it. 244 00:25:15.320 --> 00:25:24.250 Rohit Gandikota: Can we really control this structure, is the next question, right? We surely removed it. I would call that some form of control. 245 00:25:24.510 --> 00:25:28.499 Rohit Gandikota: But can I precisely more have a continuous control over this layer? 246 00:25:29.210 --> 00:25:37.360 Rohit Gandikota: Turns out, yes, you can very simply take the exact structure-finding method from the previous work. 247 00:25:37.500 --> 00:25:42.659 Rohit Gandikota: And linearly control it in a way that it has a continuous effect on your model. 248 00:25:42.930 --> 00:25:52.989 Rohit Gandikota: For example, if I were to bring all of the fine-tuning that I did for unlearning the model into a small set of parameters as a module. 249 00:25:54.120 --> 00:25:57.609 Rohit Gandikota: It does remove the beard, like, so that unlearn's beard. 250 00:25:57.860 --> 00:26:04.739 Rohit Gandikota: But you can also control the intensity of it, and it controls how much beard is removed from the model's knowledge. 251 00:26:05.650 --> 00:26:16.109 Rohit Gandikota: Which is great. So, it sort of tells that, inherently, the parameters that capture structure also has, like, a smooth, controlling definition for them. 252 00:26:16.640 --> 00:26:30.650 Rohit Gandikota: And the nice part about it is when you can erase something, you can do the total opposite and add something. So you can add a smile, you can add glasses, and sort of get a very nice dimension of control over your outputs. 253 00:26:30.970 --> 00:26:33.739 Rohit Gandikota: Just by finding structure inside there. 254 00:26:35.030 --> 00:26:39.230 Rohit Gandikota: And this is not doing anything completely new. 255 00:26:39.530 --> 00:26:41.269 Rohit Gandikota: What we did is we took 256 00:26:41.620 --> 00:26:47.509 Rohit Gandikota: The erasing formulation, and also added an enhancing formulation to it. 257 00:26:47.790 --> 00:26:54.320 Rohit Gandikota: And found ways to control this knowledge linearly in the parameter space, or in the weight space. 258 00:26:54.710 --> 00:27:02.839 Rohit Gandikota: So, what I do is I take the person vector and try to move towards the old direction and away from the young direction. 259 00:27:03.050 --> 00:27:07.599 Rohit Gandikota: And that's exactly what this erasing Parameter module is doing. 260 00:27:08.840 --> 00:27:12.980 Rohit Gandikota: So, now what happens is that you can, just using text. 261 00:27:13.410 --> 00:27:16.450 Rohit Gandikota: As methods to find your structure. 262 00:27:16.750 --> 00:27:19.969 Rohit Gandikota: You can simply control the model's internal knowledge. 263 00:27:20.200 --> 00:27:22.769 Rohit Gandikota: Right? So, all I use is curly hair. 264 00:27:22.990 --> 00:27:27.689 Rohit Gandikota: As a prompt to find the structure, and we can control it inside the model. 265 00:27:28.410 --> 00:27:43.030 Rohit Gandikota: So, yeah, there are multiple dimensions you can control, and these are also composable on top of each other, so I can keep composing more and more structure on top, so I can make them have glasses, plus old, plus smiling, and you can keep adding. 266 00:27:43.160 --> 00:27:45.940 Rohit Gandikota: Up to 200 sliders is what we've tested. 267 00:27:46.340 --> 00:27:48.179 Rohit Gandikota: And now we can test more. 268 00:27:49.450 --> 00:27:58.999 Rohit Gandikota: But more specifically, this enables you to have a multidimensional control over your images. This is, like, imagine being in a… 269 00:27:59.370 --> 00:28:09.220 Rohit Gandikota: like, 1990s, maybe, when you have a game engine where you can create your own character. You had multiple sliders you can work with, sort of create a character which you really like. 270 00:28:09.840 --> 00:28:15.759 Rohit Gandikota: This sort of, composable sliding techniques will allow you to do this now. 271 00:28:17.890 --> 00:28:22.540 Rohit Gandikota: So… And funnily, we also found 272 00:28:22.680 --> 00:28:28.659 Rohit Gandikota: Structure inside the models that will self-repair its own outputs. 273 00:28:29.170 --> 00:28:33.630 Rohit Gandikota: All we did was ask the model to be more realistic. 274 00:28:33.990 --> 00:28:43.959 Rohit Gandikota: And just made the model's structure move towards being more realistic. And it automatically started fixing its own outputs whenever it was breaking the physics. 275 00:28:44.130 --> 00:28:51.509 Rohit Gandikota: or just drawing very bad bicycles. It just corrects itself and creates much better-looking images. 276 00:28:51.830 --> 00:28:57.730 Rohit Gandikota: And originally, we thought that this was a problem of additional training or adding more guidance. 277 00:28:58.040 --> 00:29:03.130 Rohit Gandikota: Turns out this was just a structural issue, you can just fix the structure if you can find it. 278 00:29:06.260 --> 00:29:13.990 Rohit Gandikota: And this is another nice throwback to, I think, 2023, when your images were not able to generate nice hands. 279 00:29:14.240 --> 00:29:18.250 Rohit Gandikota: Yeah, you can fix them. And it was… it was, 280 00:29:18.550 --> 00:29:22.790 Rohit Gandikota: It was mind-blowing for everyone who looked at this at that time. Now it's… 281 00:29:23.450 --> 00:29:25.489 Rohit Gandikota: Every model can do this now. 282 00:29:26.600 --> 00:29:40.999 Rohit Gandikota: So, okay. So, coming back to this, if you can find a structure and control it, I think this in itself will open up a lot of possibility for you to do explorations of this model and show your intent. 283 00:29:41.270 --> 00:29:45.829 Rohit Gandikota: And turns out, this did catch a lot of attention in the community. 284 00:29:46.380 --> 00:30:00.300 Rohit Gandikota: So when we open source this method, we've immediately seen a burst of open source artists, like, training their own sliders with their own creative imaginations, and releasing it for the public to use. 285 00:30:00.800 --> 00:30:09.900 Rohit Gandikota: And this enabled, I think, a good amount of research within the stable diffusion community to work more on these controllable methods. 286 00:30:10.310 --> 00:30:14.010 Rohit Gandikota: For… for enhancing the user's intention. 287 00:30:14.310 --> 00:30:20.070 Rohit Gandikota: And I think this was… I don't want to talk too much about it, but this was kind of the chat GPT movement, I guess, I hope. 288 00:30:20.260 --> 00:30:24.980 Rohit Gandikota: Where, like, users were finally able to see that, sure, I can use 289 00:30:25.520 --> 00:30:27.860 Rohit Gandikota: these models to do something which I want. 290 00:30:29.490 --> 00:30:30.370 Rohit Gandikota: So… 291 00:30:30.520 --> 00:30:39.759 Rohit Gandikota: We've talked about this, where as long as your models can't talk to you, it becomes really hard for you to use the models beyond what your own imagination is. 292 00:30:40.000 --> 00:30:56.570 Rohit Gandikota: Right? For instance, let's say I've trained a model which is super realistic and does a lot of things, and my knowledge is that small circle over there. That's all I can explore using the concept sliders that I just showed you, because that's the extent at which I can prompt the model with. 293 00:30:56.670 --> 00:31:02.450 Rohit Gandikota: Right? And hoping that I can bring a lot of humans into this crowdsourcing work. 294 00:31:02.630 --> 00:31:05.699 Rohit Gandikota: And allowing their creativity to be open source. 295 00:31:05.830 --> 00:31:11.710 Rohit Gandikota: maybe I can extend the concept slider limitations to that area of its knowledge. 296 00:31:12.070 --> 00:31:15.010 Rohit Gandikota: But what about this area, where 297 00:31:15.150 --> 00:31:24.930 Rohit Gandikota: Where the model learns a lot of things during pre-training, but it can't say what it has learned, because that's not the way it was trained in the first place. Unlike language, I think language can… 298 00:31:26.060 --> 00:31:29.689 Rohit Gandikota: It gives an illusion of saying, yeah, yeah, sure, you should learn more about this. 299 00:31:29.880 --> 00:31:31.730 Rohit Gandikota: There's more to learn beyond this. 300 00:31:32.180 --> 00:31:37.459 Rohit Gandikota: So, how is it to explore that, and why not just ask the model to talk about its 301 00:31:37.740 --> 00:31:42.090 Rohit Gandikota: Own internals, so that you don't have to be limited by your own potential. 302 00:31:43.130 --> 00:31:53.890 Rohit Gandikota: So, what we do as a first step is we try to ask the model to generate a lot of images using the prompt that I want to explore. 303 00:31:54.160 --> 00:31:59.870 Rohit Gandikota: And then simply just decompose its outputs, And find a slider direction. 304 00:32:00.160 --> 00:32:03.870 Rohit Gandikota: Inside its weights, that would sort of pledge to one of these 305 00:32:04.100 --> 00:32:09.100 Rohit Gandikota: decompose directions. What it says is, basically, I'm trying to look For the models. 306 00:32:09.830 --> 00:32:15.649 Rohit Gandikota: internal, principal components about this particular concept that I'm very interested in. 307 00:32:16.750 --> 00:32:28.919 Rohit Gandikota: And this is completely unsupervised, so when I say Van Gogh art, I would get how many of our PCR directions I do, like, let's say 300… 300 directions that might be relevant to Van Gogh. 308 00:32:29.050 --> 00:32:34.769 Rohit Gandikota: Which I personally would not be able to bring up in the first place. So let me show you some examples. 309 00:32:35.850 --> 00:32:45.610 Rohit Gandikota: So this is the best I could prompt for monsters. It looks pretty generic, it's the same, probably, monster coming up again and again. 310 00:32:45.940 --> 00:33:02.020 Rohit Gandikota: But once you open up these models, internals, and look for directions that the model can draw, you start getting all these different weird-looking monsters, which I… I don't have an answer for. I don't know what… this… this, 311 00:33:02.350 --> 00:33:04.489 Rohit Gandikota: The captions that you see here are… 312 00:33:05.640 --> 00:33:08.910 Rohit Gandikota: are generated, like, post-discovery using Claude. 313 00:33:09.130 --> 00:33:14.149 Rohit Gandikota: So, most of them don't even match. I don't think this is Blue Bestion, that's the best Claude could do. 314 00:33:14.290 --> 00:33:16.830 Rohit Gandikota: To make sense out of that direction. 315 00:33:16.990 --> 00:33:23.850 Rohit Gandikota: But essentially, what my point is, is that you can start exploring features that you originally would not have 316 00:33:24.170 --> 00:33:26.640 Rohit Gandikota: Thought of in the first place to prompt. 317 00:33:27.870 --> 00:33:37.290 Rohit Gandikota: And again, since we got them out as form of sliders, you can smoothly control, and this, again, opens up multiple dimensions of control, except 318 00:33:37.580 --> 00:33:45.090 Rohit Gandikota: This is creative, territory where you don't know that they existed in the first place, and you're just exploring the module. 319 00:33:46.370 --> 00:33:55.679 Rohit Gandikota: I'm… the example that I… I'm really interested and excited about is us exploring art styles inside the model. 320 00:33:56.020 --> 00:34:01.139 Rohit Gandikota: So, art is a very studied concept. 321 00:34:01.340 --> 00:34:04.850 Rohit Gandikota: In the world, of course, and especially in generative AI. 322 00:34:05.310 --> 00:34:07.759 Rohit Gandikota: So when we… 323 00:34:08.239 --> 00:34:19.300 Rohit Gandikota: decompose the artistic styles inside… like, inside the diffusion models. We found directions which are not human… which are not one-to-one human corresponding, but these are sort of principal directions. 324 00:34:19.510 --> 00:34:24.090 Rohit Gandikota: That the model has learned to encode every human artist that it knows of. 325 00:34:25.760 --> 00:34:35.530 Rohit Gandikota: And what, essentially, like, originally, the human studies were done in the past to understand what artistic styles that the model could 326 00:34:35.909 --> 00:34:38.409 Rohit Gandikota: Replicate. 327 00:34:38.659 --> 00:34:41.349 Rohit Gandikota: And then it took almost 8,000… 328 00:34:41.510 --> 00:34:49.070 Rohit Gandikota: 250 artists, 7 months of their time, to come up with a list of 8,000 artists that the model can mimic. 329 00:34:49.670 --> 00:34:50.930 Rohit Gandikota: Right? So… 330 00:34:52.320 --> 00:34:59.119 Rohit Gandikota: Now, you can literally explore this space using a single prompt, and maybe 2 hours on a GPU. 331 00:34:59.310 --> 00:35:05.080 Rohit Gandikota: And get most of the relevant slider or, like, artistic directions out of these models. 332 00:35:05.710 --> 00:35:11.009 Rohit Gandikota: So… Those are all the realistic artists, like, the real artists and their art. 333 00:35:11.170 --> 00:35:22.789 Rohit Gandikota: And you can always argue that, why do I need this exploration in the first place, right? You can simply use generic prompts, like, create an artist, artistic-style image. 334 00:35:22.930 --> 00:35:27.920 Rohit Gandikota: And then the model generates these bunch of images, which all look pretty generic. 335 00:35:28.210 --> 00:35:34.710 Rohit Gandikota: And the distance between the original distribution and the model-generated distribution is quite high. 336 00:35:35.180 --> 00:35:43.350 Rohit Gandikota: Right? And you can also say, hey, I'm in the era of language models, I'll ask the language model to prompt diverse artistic styles. 337 00:35:43.700 --> 00:35:54.700 Rohit Gandikota: And that should do the trick. Why should I explore… why should I ask the model itself to talk about it? You can, and it definitely covers a lot of distance. It's almost, like, half the distance now. You're much closer. 338 00:35:55.110 --> 00:36:00.419 Rohit Gandikota: And using these LLM prompts, we also do concept sliders. 339 00:36:00.840 --> 00:36:17.570 Rohit Gandikota: And then, sure, they are very close to each other, but when you really ask the model to explore and provide its own knowledge, then you see that it covers a lot of knowledge gap between the true distribution and what we have in our hands in the first place. 340 00:36:19.290 --> 00:36:23.060 Rohit Gandikota: So… -Oh. 341 00:36:23.320 --> 00:36:24.180 Rohit Gandikota: Yes. 342 00:36:24.740 --> 00:36:32.409 Rohit Gandikota: And this is, a response to what a… so slider space essentially is restricting you again. 343 00:36:32.620 --> 00:36:37.169 Rohit Gandikota: Where you have to be the first person to seed what it is that you're looking for. 344 00:36:37.370 --> 00:36:41.469 Rohit Gandikota: What we did with this visualization is we took a concept of art. 345 00:36:41.790 --> 00:36:45.930 Rohit Gandikota: And then started exploring within slider space again and again through an iteration. 346 00:36:46.980 --> 00:36:57.060 Rohit Gandikota: Right? So what that means is you can click on any art, and it'll show you every nearby sliders around it, and you can control this space and explore it in a much interactive way. 347 00:36:57.410 --> 00:37:03.430 Rohit Gandikota: This takes a lot of time, so I… I pre-generated a lot of styles, and I'm showing you as a… 348 00:37:03.570 --> 00:37:05.410 Rohit Gandikota: As a demo here. 349 00:37:05.700 --> 00:37:13.769 Rohit Gandikota: But yeah, it got to a place which I would never have imagined in the first place, at least with my creative brain. 350 00:37:15.520 --> 00:37:22.280 Rohit Gandikota: So, okay. We have found a way to discover it. Finally. 351 00:37:22.870 --> 00:37:25.519 Rohit Gandikota: Coming back to this communication part. 352 00:37:26.320 --> 00:37:31.609 Rohit Gandikota: Where, when you ask something, the model is doing a lot of planning and generating an image. 353 00:37:32.170 --> 00:37:34.049 Rohit Gandikota: So, as a… 354 00:37:34.810 --> 00:37:41.350 Rohit Gandikota: As a user, I would want to understand what is the decision that the model made, and I want to communicate with that decision. 355 00:37:41.440 --> 00:37:53.990 Rohit Gandikota: For instance, I want to look at that particular, this particular person that it drew there, and I think that is the decision that is more relevant for this particular generation. 356 00:37:54.070 --> 00:38:01.990 Rohit Gandikota: So how is it, can I tap into this decision-making boundary, and how can I communicate with it, so that I can draw the same person chilling in a beach? 357 00:38:02.640 --> 00:38:10.530 Rohit Gandikota: Right? It becomes a really hard problem when… when you want to now go into the model's decisions and internal decision space. 358 00:38:11.580 --> 00:38:22.649 Rohit Gandikota: So, a quick intro to how the model that I'm working on looks like. So, it's… it's called Flux. It has two streams, an image and text. 359 00:38:22.770 --> 00:38:27.679 Rohit Gandikota: So both of them are processed together. You start from a noise, and then you do multiple layers. 360 00:38:27.980 --> 00:38:36.810 Rohit Gandikota: Of the same block. You also have, like, some fancy, conditioning called time step and guidance. They all go into this flux block. 361 00:38:37.270 --> 00:38:39.670 Rohit Gandikota: Finally, it generates an image. 362 00:38:40.030 --> 00:38:47.380 Rohit Gandikota: From the image stream, but then you discard the text stream, because you don't need text when you're talking about image generation. 363 00:38:47.920 --> 00:38:54.409 Rohit Gandikota: Right? This is how the model is trained, where it takes a noise, does a dual stream, 364 00:38:54.880 --> 00:39:03.659 Rohit Gandikota: processing, and then you only take the image stream of things, decode it, and just discard that final. There's no optimization done on this text stream. 365 00:39:04.380 --> 00:39:10.709 Rohit Gandikota: So, what we found is that, what if we can use this rich information that the model has learned. 366 00:39:11.430 --> 00:39:13.849 Rohit Gandikota: And discard the image side of things. 367 00:39:14.150 --> 00:39:18.050 Rohit Gandikota: And look what the text has learned. And what can we do with that text. 368 00:39:19.500 --> 00:39:20.470 Rohit Gandikota: So… 369 00:39:21.320 --> 00:39:27.500 Rohit Gandikota: That's exactly what we do. We give the reference image that the model has generated to the same model again. 370 00:39:27.820 --> 00:39:28.790 Rohit Gandikota: And then… 371 00:39:29.340 --> 00:39:38.059 Rohit Gandikota: Instead of touching the concept that you want as a haptic, you can literally just say a layperson on the left playing bass. 372 00:39:38.640 --> 00:39:41.949 Rohit Gandikota: So now, we know that the leaf person belongs to that 373 00:39:42.220 --> 00:39:47.840 Rohit Gandikota: person on the left, through captioning. I'm sort of capturing what it is that I'm really interested in. 374 00:39:48.620 --> 00:39:56.109 Rohit Gandikota: And when I use the same prompt, a leaf person chilling on a beach, the regular diffusion model draws this. 375 00:39:56.370 --> 00:40:00.210 Rohit Gandikota: And it's not that person, of course. Why would it? 376 00:40:00.490 --> 00:40:06.650 Rohit Gandikota: Except, I can take these… the vectors that come out of this magic network, which I'll talk about 377 00:40:06.960 --> 00:40:13.240 Rohit Gandikota: In a few seconds, and take the information that it has learned and plug it into my generative model. 378 00:40:13.480 --> 00:40:15.000 Rohit Gandikota: And then agenda instead. 379 00:40:16.300 --> 00:40:20.289 Rohit Gandikota: Right? You can… you can… You can bring… 380 00:40:20.810 --> 00:40:24.359 Rohit Gandikota: You can tap into the internal decisions that the model makes. 381 00:40:24.510 --> 00:40:27.979 Rohit Gandikota: And bring it into your own generations, and play around with it. 382 00:40:29.380 --> 00:40:30.230 Rohit Gandikota: So… 383 00:40:30.760 --> 00:40:37.100 Rohit Gandikota: The way it was trained is that the feedforward network, we show the image and the caption that it takes. 384 00:40:38.010 --> 00:40:51.860 Rohit Gandikota: And it generates this, what I call, like, decision vectors, or customization vectors, for each of the text prompt there is, so that if you're interested in the dog, you can take the vector of the dog token. 385 00:40:52.000 --> 00:40:58.210 Rohit Gandikota: If you're interested in the background, you can take that token. At least it gives you a range of communication bandwidth. 386 00:40:58.390 --> 00:41:03.070 Rohit Gandikota: Of how… what it is that you want to customize, or bring, or what decision it is you want to capture. 387 00:41:03.590 --> 00:41:07.720 Rohit Gandikota: I use the same reference prompt as my generation. 388 00:41:07.910 --> 00:41:16.809 Rohit Gandikota: So if the original diffusion model is generating some other image, I just ask it to generate the same image when all the tokens are employed. So essentially, I'm training this model. 389 00:41:17.260 --> 00:41:27.499 Rohit Gandikota: to say, whatever it is that you're generating from your vectors, word to word, they have to represent some delta towards the concept that I'm really interested in. 390 00:41:29.380 --> 00:41:38.220 Rohit Gandikota: Right? There'll be questions, I'm sure. I'm happy to answer them. If this… if you're absolutely lost, please let me know. 391 00:41:39.120 --> 00:41:46.609 Rohit Gandikota: So yeah, it's a simple MSC that we do, and then there is no other additional, 392 00:41:46.840 --> 00:41:51.539 Rohit Gandikota: captioning or annotations that we are doing, it's a simple text-image captions. 393 00:41:52.020 --> 00:41:53.600 Rohit Gandikota: And using this. 394 00:41:53.750 --> 00:42:01.219 Rohit Gandikota: a single paradigm. We basically learn the true language of whatever is happening inside and the decisions that the models are making. 395 00:42:01.860 --> 00:42:10.629 Rohit Gandikota: So, yeah, we collected 2 million images and trained this general-purpose network that does this. 396 00:42:11.180 --> 00:42:18.000 Rohit Gandikota: So now, what you can do is you can have a bunch of reference images that you want your concept to come from. 397 00:42:18.300 --> 00:42:26.860 Rohit Gandikota: and have a reference prompt to show what it is that you want to pick the concept from, right? So, I'll explain 398 00:42:27.050 --> 00:42:28.070 Rohit Gandikota: these things. 399 00:42:28.950 --> 00:42:33.329 Rohit Gandikota: So these… this is the inference prompt. I said, a dog sitting on a sofa. 400 00:42:33.580 --> 00:42:38.140 Rohit Gandikota: And the original model generates diverse Not so diverse, but… 401 00:42:38.360 --> 00:42:43.540 Rohit Gandikota: very different looking dogs, right? Except, now I can… 402 00:42:43.650 --> 00:42:45.780 Rohit Gandikota: Give this to my feed-forward model. 403 00:42:46.190 --> 00:42:52.129 Rohit Gandikota: And bring in the delta vectors from the dog token, and 404 00:42:52.520 --> 00:42:54.969 Rohit Gandikota: And plug it into this original seeds. 405 00:42:55.080 --> 00:42:58.260 Rohit Gandikota: And then, immediately, the model starts drawing the dog. 406 00:42:58.380 --> 00:43:03.169 Rohit Gandikota: That's in the reference image. So it's essentially, I'm picking the concept and dropping it into my 407 00:43:03.440 --> 00:43:04.969 Rohit Gandikota: Diffusion model. 408 00:43:05.650 --> 00:43:10.730 Rohit Gandikota: And, yeah, it… you can see that for different dogs, it brings all the different dogs. 409 00:43:11.310 --> 00:43:14.559 Rohit Gandikota: From the reference image and puts it into your generations. 410 00:43:17.310 --> 00:43:18.840 Rohit Gandikota: Oh, I had animation. 411 00:43:19.320 --> 00:43:20.030 Rohit Gandikota: Yes. 412 00:43:21.510 --> 00:43:22.330 Rohit Gandikota: Okay. 413 00:43:23.320 --> 00:43:27.990 Rohit Gandikota: So, till here, this is mostly about what I've done so far. 414 00:43:28.230 --> 00:43:30.179 Rohit Gandikota: And I think I could… oh. 415 00:43:30.530 --> 00:43:32.020 Rohit Gandikota: 30 minutes, that's nice. 416 00:43:32.180 --> 00:43:37.450 Rohit Gandikota: So… In summary, what I've taken you through is that 417 00:43:37.560 --> 00:43:41.350 Rohit Gandikota: We never had a way to… Verify their structure. 418 00:43:41.690 --> 00:43:57.280 Rohit Gandikota: We did. Now, we find that there is structure inside these models. You can control the structure. It's also discoverable. There are… it's just the first step, and there are many follow-ups that sort of worked on top of this to make it more discoverable. 419 00:43:57.410 --> 00:44:00.820 Rohit Gandikota: And you can now communicate with the internal decisions that the model do. 420 00:44:01.040 --> 00:44:05.289 Rohit Gandikota: With this in hand, I'll show you what you can do now in the current 421 00:44:05.660 --> 00:44:08.930 Rohit Gandikota: today, with vision models, right? So… 422 00:44:09.320 --> 00:44:11.860 Rohit Gandikota: You ask for a photo of a woman. 423 00:44:12.420 --> 00:44:15.309 Rohit Gandikota: Seated on a couch at a cafe, and it draws this image. 424 00:44:15.690 --> 00:44:20.090 Rohit Gandikota: Think she's too serious, so I can add a smile. 425 00:44:20.420 --> 00:44:23.170 Rohit Gandikota: And then that makes the women smile now. 426 00:44:23.410 --> 00:44:28.380 Rohit Gandikota: Right? You can say, no, no, no, I want that couch from my favorite TV show. 427 00:44:28.890 --> 00:44:33.739 Rohit Gandikota: and say, I want the woman to be sitting on that, It brings the couch. 428 00:44:33.850 --> 00:44:43.400 Rohit Gandikota: Right? You can also say, yeah, I think I want, like, some funky new hairstyle, but I don't know what it is. Let me ask the model what all hairstyles it has in its… 429 00:44:43.600 --> 00:44:47.819 Rohit Gandikota: directory, and then I can choose style number 23 that the model gave me. 430 00:44:48.230 --> 00:44:51.329 Rohit Gandikota: And then it draws this. So, now you… 431 00:44:51.520 --> 00:45:03.110 Rohit Gandikota: It opens up to a new form of exploring the model and working with them, and finally, it, like, gives you some room to show your intention as a user and what you can do in the first place. 432 00:45:04.120 --> 00:45:08.850 Rohit Gandikota: Okay, this slide took me… 3 weeks to do it. 433 00:45:09.980 --> 00:45:10.820 Rohit Gandikota: Fair. 434 00:45:11.010 --> 00:45:12.999 Rohit Gandikota: I can show intent in 3 weeks. 435 00:45:13.720 --> 00:45:20.810 Rohit Gandikota: But is this the real fix? I think now comes… the main, 436 00:45:21.770 --> 00:45:24.100 Rohit Gandikota: Good, solid 10 minutes portion of the talk. 437 00:45:24.330 --> 00:45:29.169 Rohit Gandikota: I don't think this is a real fix, especially because these are all post hoc methods. 438 00:45:29.440 --> 00:45:39.869 Rohit Gandikota: We are training the models. It's very similar to what you in the lab are doing with language models. There's one stream of researchers who are really pushing to go towards 439 00:45:41.250 --> 00:45:55.400 Rohit Gandikota: the top 1% AI model that can solve RKGI3, 3.5, 5, whatever. And then there are us who's trying to be, oh, let me understand this, I don't know what's happening here. By the time I understood GPT-2, 440 00:45:55.640 --> 00:46:01.140 Rohit Gandikota: there's now GPT-10, right? So we… we are playing this cat and mouse game. 441 00:46:01.620 --> 00:46:09.889 Rohit Gandikota: And… I… I don't want to be, way too optimistic with language, but with vision at least. 442 00:46:10.530 --> 00:46:15.690 Rohit Gandikota: It's absolutely necessary that we understand the models for them to be even workable in the first place. 443 00:46:15.950 --> 00:46:26.270 Rohit Gandikota: Right? I showed you a lot of slides earlier, which took 4 years of research to get it to a phase where I can show you one slide. 444 00:46:26.550 --> 00:46:30.109 Rohit Gandikota: Where I have to work for 3 weeks, myself, who built these methods. 445 00:46:30.500 --> 00:46:34.949 Rohit Gandikota: Right? And I don't think that's no way close to being a tool that'll help you show your intention. 446 00:46:36.130 --> 00:46:41.999 Rohit Gandikota: So, I think… The real fix has to be in, like, multiple ways. 447 00:46:42.260 --> 00:46:55.969 Rohit Gandikota: Which always comes back to the same question as to what will make vision the tool that will be widely used in the future. And it's the same thing, which is, I think we'll have to focus on the intention of humans, and… 448 00:46:56.110 --> 00:47:03.670 Rohit Gandikota: How do we enable that to happen in the first place? And it's not an input-output game. I think we should, like, start moving away from it. 449 00:47:03.850 --> 00:47:10.229 Rohit Gandikota: Where we are not trying to optimize towards what's the image arena number one model. 450 00:47:10.770 --> 00:47:16.089 Rohit Gandikota: so, okay, so the first thing is, I think. 451 00:47:16.560 --> 00:47:20.310 Rohit Gandikota: Visual generation and interpretability needs to go hand in hand. 452 00:47:20.960 --> 00:47:23.749 Rohit Gandikota: Right? What I mean by that is… 453 00:47:24.070 --> 00:47:29.119 Rohit Gandikota: when I started thinking about this, it's an upcoming work-in-progress project. 454 00:47:29.490 --> 00:47:39.210 Rohit Gandikota: me and David were talking about, oh, how do we bring this into reality? What is it? And then the first question I wanted to ask was, what format should I choose? 455 00:47:39.730 --> 00:47:41.369 Rohit Gandikota: To start this problem. 456 00:47:41.530 --> 00:47:44.960 Rohit Gandikota: Right? I think, obviously, the answer is video, right? 457 00:47:45.090 --> 00:47:51.699 Rohit Gandikota: Because it has image, it has time, but then 3D is also nice, because you can 458 00:47:51.910 --> 00:47:54.130 Rohit Gandikota: Look at multiple views. 459 00:47:54.480 --> 00:47:58.620 Rohit Gandikota: I think world modeling is also nice. You can, like, explore the space. 460 00:47:59.050 --> 00:48:05.830 Rohit Gandikota: And the more I thought about this, not for a project perspective, but from a, 461 00:48:07.070 --> 00:48:13.750 Rohit Gandikota: From a visual… vision perspective, is that I think every format enables a new kind of 462 00:48:14.190 --> 00:48:19.269 Rohit Gandikota: intention in users. And I think that's why all of them are, in their own ways, very popular. 463 00:48:19.430 --> 00:48:32.010 Rohit Gandikota: Right? World modeling is really popular because you get to explore it, you can play with it, you can say, go forward, go back, turn left, turn right. 3D modeling can show you different perspectives of the same image. 464 00:48:33.610 --> 00:48:39.809 Rohit Gandikota: SVGs, on the other hand, are very human-readable. You can read the code that SVGs are built on top of. 465 00:48:40.280 --> 00:48:46.619 Rohit Gandikota: images are really realistic. So there's signal processing, you can control sliders in Adobe Premium Photoshop. 466 00:48:46.770 --> 00:48:55.489 Rohit Gandikota: And all of them, in their own ways, bring some sort of signal into your intention. And it… that… I think that's why they're… 467 00:48:55.910 --> 00:48:57.620 Rohit Gandikota: Popular in their own way. 468 00:48:58.050 --> 00:49:04.979 Rohit Gandikota: So… Ideally, this is the work in progress, where what we want to do is 469 00:49:05.560 --> 00:49:12.949 Rohit Gandikota: Every time a user types a prompt to generate an image, it's always heavily underspecified. Nobody is going to sit there and write a book 470 00:49:13.180 --> 00:49:14.360 Rohit Gandikota: Of a prompt. 471 00:49:14.570 --> 00:49:17.970 Rohit Gandikota: to generate exactly what they want in their image, right? That's… 472 00:49:18.110 --> 00:49:20.400 Rohit Gandikota: I think unreasonable to us from a user. 473 00:49:20.640 --> 00:49:30.380 Rohit Gandikota: Which is why the model internally is expanding so many different things, having to do all this planning in the first place. So, what if we bring that outside the generation process? 474 00:49:30.560 --> 00:49:35.970 Rohit Gandikota: Right? And let the generative model just be a renderer, which has always been how vision was working. 475 00:49:36.120 --> 00:49:42.740 Rohit Gandikota: From a long time ago, right? So, if I were to have this planner that would plan everything out. 476 00:49:42.970 --> 00:49:47.890 Rohit Gandikota: Ideally in human-legible text, or code, or anything, any other format. 477 00:49:48.200 --> 00:49:53.499 Rohit Gandikota: That would let us, as users, understand what the model is deciding to draw. 478 00:49:53.710 --> 00:49:58.889 Rohit Gandikota: And then have this renderer, whose job is literally just to render the image. 479 00:49:59.460 --> 00:50:07.490 Rohit Gandikota: Right? That way, you can edit your planning that the model has planned, and see that edited in real time, or 480 00:50:07.660 --> 00:50:15.489 Rohit Gandikota: Or at least in the right way, where you can have 3D control over your objects, and, like, have your objects rotated in any way you want. 481 00:50:15.680 --> 00:50:16.500 Rohit Gandikota: And… 482 00:50:17.140 --> 00:50:28.279 Rohit Gandikota: To actually make this more useful, you would also want something in a cyclic way, where it'll look at your image and will be able to give the plan that requires to be, to render that image back. 483 00:50:28.530 --> 00:50:29.400 Rohit Gandikota: Right? 484 00:50:30.450 --> 00:50:43.970 Rohit Gandikota: So, every aspect of this brings some sort of intention from every modality that we have seen earlier, right? Your planner needs to know about 3D and world modeling so that it can exactly look 485 00:50:44.020 --> 00:50:52.710 Rohit Gandikota: at different angles, plan for different angles. Your renderer is almost like an SVG that takes a code and generates a realistic output, other than, like. 486 00:50:52.870 --> 00:50:55.060 Rohit Gandikota: That looks like SVG. 487 00:50:55.240 --> 00:51:05.900 Rohit Gandikota: And your parser needs to know a lot of signal processing to understand what image it's looking at, what contrast and composure that it has to do to generate the image back. 488 00:51:06.290 --> 00:51:07.160 Rohit Gandikota: And… 489 00:51:08.190 --> 00:51:17.170 Rohit Gandikota: I think a good news is that it enables wipe coding too, right? This is a big chunk of new language, and you can literally seek a lot of 490 00:51:17.320 --> 00:51:22.059 Rohit Gandikota: Agents on top of it, and talk to your image, and enable the editing. 491 00:51:22.170 --> 00:51:27.619 Rohit Gandikota: Through vibe coding, or maybe vibe image, right? I'm the first to claim it. 492 00:51:28.280 --> 00:51:34.900 Rohit Gandikota: And current modern-day vision models are doing all of this packed into one single model. 493 00:51:35.210 --> 00:51:38.779 Rohit Gandikota: And all that we have been trying to do as, 494 00:51:38.930 --> 00:51:53.819 Rohit Gandikota: interp researchers, or controllability researchers, have been trying to guess that this is what is happening, and going and trying to control them internally. And I don't know if that's what I'm capturing. I just gave you a story that that is what is happening. I still can't… 495 00:51:53.940 --> 00:51:55.980 Rohit Gandikota: Guarantee that that's what is happening. 496 00:51:56.560 --> 00:51:58.820 Rohit Gandikota: But I think it's time to break that. 497 00:51:59.060 --> 00:52:00.470 Rohit Gandikota: and have… 498 00:52:00.910 --> 00:52:09.889 Rohit Gandikota: Interp to come out of the model that there is, and make this more usable for… and also, like, give the user an illusion 499 00:52:10.000 --> 00:52:25.199 Rohit Gandikota: that they understand exactly what's happening inside the model, right? I… there… this opens up a lot of other avenues. What is this planner internally doing? We don't know. What is the renderer doing? We don't know. But at least this sort of opens up a way for users to show their intention. 500 00:52:25.390 --> 00:52:29.590 Rohit Gandikota: Right? Pushing these models to be more widely used tools. 501 00:52:30.250 --> 00:52:36.629 Rohit Gandikota: And the more… I started making that slide, the second thing came out. 502 00:52:36.920 --> 00:52:45.450 Rohit Gandikota: every time I show how reasoning works inside language model, it's always language. Sorry, inside vision model, it's always language. 503 00:52:45.560 --> 00:52:50.090 Rohit Gandikota: But how would a visual… Reasoning medium look like? 504 00:52:50.200 --> 00:52:56.089 Rohit Gandikota: This is slightly half-baked. I'm still not sure what it should look like. I think this can be a… 505 00:52:56.350 --> 00:53:04.850 Rohit Gandikota: big, research area on its own, but I feel like every vision person I can talk to will definitely vouch for this. 506 00:53:05.040 --> 00:53:10.210 Rohit Gandikota: That image medium on its own brings a lot more information that text can't bring. 507 00:53:10.730 --> 00:53:11.490 Rohit Gandikota: Right? 508 00:53:11.740 --> 00:53:14.260 Rohit Gandikota: Or can bring, but it would take 509 00:53:14.460 --> 00:53:17.479 Rohit Gandikota: Like, unreasonably large amount of text to bring it. 510 00:53:18.220 --> 00:53:31.320 Rohit Gandikota: So, I feel like having some sort of visual reasoning… visual medium reasoning inside the model would obviously be, much more efficient for us to have this sort of… and also, like. 511 00:53:31.480 --> 00:53:32.800 Rohit Gandikota: human usability. 512 00:53:33.190 --> 00:53:46.030 Rohit Gandikota: When it comes to generating these images. Like, for instance, if I want… like, the model might be planning inside that, oh, should I place the dot behind a layer 1 of the tree, or layer 2 of the tree? All of these things are hard to teach to the model 513 00:53:46.210 --> 00:53:54.379 Rohit Gandikota: Via text, and I think there has to be some form of visual reasoning happening inside. 514 00:53:55.150 --> 00:54:01.159 Rohit Gandikota: And finally, okay, so I have brought you into this room by promising that 515 00:54:01.440 --> 00:54:05.819 Rohit Gandikota: Oh, vision is going towards a dead end, and I have the solution to fix it. 516 00:54:06.000 --> 00:54:16.080 Rohit Gandikota: But I have silently been holding back some information. I do hope that generation on its own, visual generation on its own, comes to some natural endpoint. 517 00:54:16.250 --> 00:54:22.690 Rohit Gandikota: Right? Where visual generation is not a standing research area where people are working on it for a long time. 518 00:54:23.040 --> 00:54:26.470 Rohit Gandikota: What I mean by that is, I think vision on its own 519 00:54:26.780 --> 00:54:33.560 Rohit Gandikota: Has to be changed into something which is readable, or which is easily explorable by us humans. 520 00:54:33.890 --> 00:54:49.769 Rohit Gandikota: every time we look at an image, it's… it becomes, like, such a novelty thing that I look at this nebula, and I'm like, wow, this is so nice. But that's pretty much it. I can't explore what's happening inside that image, I don't know what the image is made of, what's the reasoning happening behind it. For that to… 521 00:54:49.950 --> 00:54:54.990 Rohit Gandikota: For me to understand that, I need to go to internet, search, read the text behind it. 522 00:54:55.120 --> 00:54:57.589 Rohit Gandikota: Maybe make sense out of what that nebula is. 523 00:54:57.820 --> 00:54:59.290 Rohit Gandikota: Right? And… 524 00:54:59.800 --> 00:55:08.319 Rohit Gandikota: what I'm really hoping for is vision generation becomes a very small part of us understanding how the visual world works in the first place. 525 00:55:08.710 --> 00:55:19.240 Rohit Gandikota: Right? Where you can finally look at this nebula and say, hey, I want to look at it from a different angle. And then your agent goes to internet, brings multiple images, and forms a 3D rendering on its own internally. 526 00:55:19.360 --> 00:55:22.899 Rohit Gandikota: And you can go explore that space. You can click into that image. 527 00:55:23.180 --> 00:55:31.969 Rohit Gandikota: go beyond what you look in the first place. And I think for this to happen, if this all sounds like a sci-fi, well, of course, language modeling sounded like sci-fi. 528 00:55:32.360 --> 00:55:35.290 Rohit Gandikota: 5, 10 years ago. But I think this… 529 00:55:35.770 --> 00:55:38.390 Rohit Gandikota: like, this is what I really care about. I think… 530 00:55:38.510 --> 00:55:48.240 Rohit Gandikota: I love images, I love visual world, but it always holds you back that you can't explore them, and you can't understand them the way we understand and explore text. 531 00:55:48.520 --> 00:55:51.860 Rohit Gandikota: So I think, eventually, I think that that has to be… 532 00:55:52.080 --> 00:55:54.010 Rohit Gandikota: That is where I would retire, I think. 533 00:55:54.790 --> 00:55:56.950 Rohit Gandikota: So, okay, thank you very much. 534 00:56:11.360 --> 00:56:25.659 Rohit Gandikota: Did you have any questions in the beginning? Yes. Oh, yes. Oh, actually, I have a question for the end. Yes, so I guess, when you talk about, like, I like your last slide, very fascinating, like, for example, like, vision, like, or yeah, I wouldn't say vision, but I would say image generation is not all. 535 00:56:25.970 --> 00:56:29.970 Rohit Gandikota: Because, for example, you can ask questions, they want to see a 3D rendering of stuff. 536 00:56:30.600 --> 00:56:36.219 Rohit Gandikota: I actually feel like for vision… visual exploration or, like, interaction, like. 537 00:56:36.360 --> 00:56:48.099 Rohit Gandikota: what do you think will be a good interface? I feel like that's actually the biggest question, or one of the biggest questions. For language, you know, interface is text. But for images, you can sketch. 538 00:56:48.300 --> 00:57:02.599 Rohit Gandikota: can compose things in layers, we can have sliders, we can have, like, text, you know, there's so many different things you can play with images. Yeah. So I feel like that's actually one of the things that's really hard for the image generation space to converge to. Yep. 539 00:57:02.710 --> 00:57:03.650 Rohit Gandikota: And… 540 00:57:04.040 --> 00:57:10.139 Rohit Gandikota: I'm thinking about, do you think whether we can just have just one model that tackles everything, or it can have, like, different… 541 00:57:10.510 --> 00:57:17.990 Rohit Gandikota: seem to explore things in two different angles? What's your thoughts on that? Yeah, I… So… 542 00:57:18.330 --> 00:57:20.890 Rohit Gandikota: Following the talk that I gave. 543 00:57:21.280 --> 00:57:35.509 Rohit Gandikota: I think definitely it has to be multiple different models coming up and doing the same task as an orchestra, so that at least you have good, resolution to look into what it is each model and, like, is doing. 544 00:57:35.720 --> 00:57:39.839 Rohit Gandikota: But I… I very strongly agree that, 545 00:57:40.790 --> 00:57:47.529 Rohit Gandikota: looking at the technology side of things is one aspect, but the real aspect would eventually become what is the UI? 546 00:57:47.680 --> 00:57:50.340 Rohit Gandikota: That a user is interacting with, right? 547 00:57:50.490 --> 00:57:57.569 Rohit Gandikota: And, with the very narrow view that I have with the slides that I made, I think it 548 00:57:57.680 --> 00:58:00.469 Rohit Gandikota: it becomes something of a hybrid UI. 549 00:58:00.650 --> 00:58:09.619 Rohit Gandikota: Where, given an image, you, as a user, you talk or you type, and then the UI should eventually give you the controls 550 00:58:09.870 --> 00:58:15.560 Rohit Gandikota: dynamically, saying, hey, I want to explore this. Then it gives you, like, arrow keys for you to explore. 551 00:58:15.780 --> 00:58:20.240 Rohit Gandikota: And you say, no, no, no, I actually want to see a 3D view of it, then it gives you a 3D slider to… 552 00:58:20.600 --> 00:58:22.539 Rohit Gandikota: do a 3D rendering of it. 553 00:58:23.040 --> 00:58:27.709 Rohit Gandikota: Essentially where… You feel, 554 00:58:28.010 --> 00:58:36.850 Rohit Gandikota: less friction when it comes to wanting to choose the right tool to do the right thing. I think it should be somehow an intelligent way to 555 00:58:37.090 --> 00:58:39.299 Rohit Gandikota: Feed these, controls. 556 00:58:39.760 --> 00:58:42.809 Rohit Gandikota: Based on your intention, if that makes sense. I see. 557 00:58:43.390 --> 00:58:46.700 Rohit Gandikota: But, yeah, it definitely requires a lot of, 558 00:58:46.930 --> 00:58:51.070 Rohit Gandikota: pretty such, I guess, almost as if, like, for example, one… 559 00:58:52.140 --> 00:59:03.440 Rohit Gandikota: like, recent way I can think about it is just, like, an LM can call, like, a tool use, for example, make its own tool, and then maybe create its own, like, widget. Oh, that'll be nice. Yeah, yeah, why not? Yeah. 560 00:59:05.690 --> 00:59:11.549 Rohit Gandikota: But doesn't it go back to the fact that your intention is more expressable? 561 00:59:12.750 --> 00:59:13.930 Rohit Gandikota: What does that mean? 562 00:59:14.400 --> 00:59:19.660 Rohit Gandikota: So you said, like, that the tool that you're gonna use, or the interface, is based on your intention. 563 00:59:19.980 --> 00:59:22.709 Rohit Gandikota: And then you can prompt. 564 00:59:23.230 --> 00:59:27.160 Rohit Gandikota: something, like, I want to get, like, a 3D view of these, like. 565 00:59:27.420 --> 00:59:38.460 Rohit Gandikota: But you will use text to represent your intention, and then an agent or someone will convert that to, like, the writing. Yeah, yeah, that's a… that's a good point. So I, I, 566 00:59:40.200 --> 00:59:46.840 Rohit Gandikota: I'm not saying text is the only medium to show intent. I'm saying it's a good medium to express your intent. 567 00:59:47.680 --> 00:59:56.910 Rohit Gandikota: But then, when it actually comes to editing or exploring it, it becomes, like, all these multiple dimensions of control, where you're touching the image, like, you're… 568 00:59:57.300 --> 01:00:01.369 Rohit Gandikota: Like, doing some sort of a model walk, or medium walk. 569 01:00:01.510 --> 01:00:20.730 Rohit Gandikota: There are, like, sliders which will give you control as to how much the person is smiling. But for, like, there should be some medium where you say, oh, I think this person needs to smile more, right? But then, it's not just by text, you're saying, no, no, smile more, no, no, no, even more. That's not what you're doing, right? When you say, I want to control the smile. 570 01:00:20.900 --> 01:00:24.449 Rohit Gandikota: And then the model should give you some sort of a new medium to 571 01:00:25.430 --> 01:00:27.790 Rohit Gandikota: Control the intent that you have. 572 01:00:28.450 --> 01:00:31.569 Rohit Gandikota: But it's a good question. I think… 573 01:00:31.990 --> 01:00:41.670 Rohit Gandikota: to start an intention, you need a way to communicate, and right now, I see text as a way to start it. Yeah, I think text is very similar, like, very… 574 01:00:42.450 --> 01:00:52.360 Rohit Gandikota: progressed, and very semantic. Yes, yes. It's like you've got, like, a semantic segmentation map of your image, almost. Yeah. Or something like that. Right. 575 01:00:54.010 --> 01:00:59.410 Rohit Gandikota: But the eventual… Intention needs to go through multiple mediums, not just text. 576 01:01:02.030 --> 01:01:08.269 Rohit Gandikota: And also, like, say, if… The user interface is text, then, like. 577 01:01:08.640 --> 01:01:16.490 Rohit Gandikota: Well, I guess you're proposing something else. Let's say if the user interface is X, then I can easily argue it a little bit. Let's try to be the bad guy. 578 01:01:16.660 --> 01:01:20.420 Rohit Gandikota: I think nowadays people do, like, prompt expansions, all these things, but… Yeah. 579 01:01:21.060 --> 01:01:40.400 Rohit Gandikota: shout out to my friend Red. Yes, yeah, yeah. Yeah, they have all these… after you specify, like, a text prompt, which is, like, not really specified, they'll have the OM call, like, to generate, like, various, like, very detailed text prompts. Yes. They have, like, different types of, like, text, where the users can then edit the text prompt they choose. Absolutely, yeah. But I feel like… 580 01:01:41.550 --> 01:01:56.190 Rohit Gandikota: That's not satisfying enough. Of course, like, text can describe the image so well, but I feel like it's not intuitive enough for humans to say, hey, like, I really want to look into these text prompt and select the very conversive process. Yes. I feel like interface is still, like, a big issue. 581 01:01:56.190 --> 01:02:02.039 Rohit Gandikota: Yeah. I agree. And I also want to clarify, I don't want to remove text from… 582 01:02:02.690 --> 01:02:04.330 Rohit Gandikota: Whatever is happening right now. 583 01:02:04.440 --> 01:02:11.179 Rohit Gandikota: I think it has to be there, like, it can do multiple good things, but it restricts us from doing a lot more beautiful things with images. 584 01:02:11.610 --> 01:02:18.799 Rohit Gandikota: And I'm just saying it might require new ways to interact with them, new mediums of reasoning. 585 01:02:19.690 --> 01:02:22.419 Rohit Gandikota: But yeah, eventually. 586 01:02:23.160 --> 01:02:29.660 Rohit Gandikota: think text will be a way to… hey, like, hey Siri, and then… And we'll get… 587 01:02:32.750 --> 01:02:36.970 Rohit Gandikota: Do you think that understanding model internals is… 588 01:02:37.560 --> 01:02:48.120 Rohit Gandikota: Necessary for providing the user with the control and ability to express and Exercise doing tent. 589 01:02:48.270 --> 01:02:49.360 Rohit Gandikota: or… 590 01:02:50.080 --> 01:02:59.489 Rohit Gandikota: Or maybe scale will solve everything, once… once the big labs collect enough examples of… it doesn't have to be text, but whatever 591 01:03:00.220 --> 01:03:16.129 Rohit Gandikota: whatever representation you use for expressing user intent, they collect enough of these examples, they match them with images, they train one model to do this, and do they still… do you think you're… sort of… your techniques are kind of… 592 01:03:16.370 --> 01:03:29.990 Rohit Gandikota: not all the way back and terp, but they are, you know, looking internally and cleverly selecting where to do interventions. Do you think this is needed, or maybe scale will solve everything? So, that's a very good question. 593 01:03:30.470 --> 01:03:34.540 Rohit Gandikota: And I was asked this question when I gave a similar talk, and… 594 01:03:35.230 --> 01:03:45.520 Rohit Gandikota: So, my response is, we already have a lot of models, like Stable Diffusion Excel, that spent a lot of money to collect different forms of input-output 595 01:03:45.680 --> 01:03:48.790 Rohit Gandikota: data and train those models, like, including Nano Banana now. 596 01:03:48.930 --> 01:03:51.219 Rohit Gandikota: Where you can talk to your images in text. 597 01:03:51.800 --> 01:03:55.840 Rohit Gandikota: And with these models, by understanding 598 01:03:56.550 --> 01:04:00.519 Rohit Gandikota: these, the ways to interact with them more, unlocks 599 01:04:00.760 --> 01:04:04.430 Rohit Gandikota: A slight bit more delta than you would use them when you don't have. 600 01:04:05.330 --> 01:04:11.289 Rohit Gandikota: Right? I'm saying you can use billion dollars to create this one single model that would do most of the things. 601 01:04:11.500 --> 01:04:16.449 Rohit Gandikota: But I think there's always a way to unlock more out of it, if you understand how vision works. 602 01:04:18.040 --> 01:04:21.969 Rohit Gandikota: Like, you can get… you can get more use for your user. 603 01:04:22.290 --> 01:04:28.430 Rohit Gandikota: When you understand what it does, and enable new controls from it. But… I, I, I think… 604 01:04:29.020 --> 01:04:32.940 Rohit Gandikota: Coming to the… the… The next steps. 605 01:04:33.060 --> 01:04:34.440 Rohit Gandikota: that I was suggesting. 606 01:04:34.580 --> 01:04:42.100 Rohit Gandikota: I do want big companies to spend a lot of money and build those kind of methods. 607 01:04:42.310 --> 01:04:47.849 Rohit Gandikota: which would… Which would eventually allow users to see what the model is thinking through. 608 01:04:47.980 --> 01:04:53.999 Rohit Gandikota: And then… have a sense of intention, and that I can change something that the model is thinking of. 609 01:04:54.670 --> 01:05:01.679 Rohit Gandikota: If it is single model, multiple models, I don't think I have a strong opinion yet. 610 01:05:01.860 --> 01:05:04.740 Rohit Gandikota: Because I haven't tested it myself. 611 01:05:05.080 --> 01:05:11.899 Rohit Gandikota: But… It is eventually really, really important that humans feel that they're 612 01:05:12.320 --> 01:05:15.939 Rohit Gandikota: That their intentions are heard, to be cheesy. 613 01:05:16.650 --> 01:05:24.430 Rohit Gandikota: if that… I don't think that happens with vision models, and there has to be some way for you to show what the model is doing, so that 614 01:05:24.600 --> 01:05:33.969 Rohit Gandikota: they have this sense of control over the generations. But do I need to have a way to show what the model is doing internally, or is it enough that… 615 01:05:34.250 --> 01:05:36.529 Rohit Gandikota: I have a way for the user to… 616 01:05:36.830 --> 01:05:47.779 Rohit Gandikota: specify what they want to get. Maybe it's underspecified, and then it's being expanded, whatever. But I have this plan, structured plan, or whatever it is. The model is generating an image. 617 01:05:47.950 --> 01:05:59.340 Rohit Gandikota: It's working, what do I care how it's doing that internally? Yeah, there are models already that do most of the things that, I described. Yeah. 618 01:06:00.120 --> 01:06:02.940 Rohit Gandikota: There's always, like… I think… 619 01:06:04.690 --> 01:06:10.650 Rohit Gandikota: So, from stable diff… I can give you, like, a quick timeline. Stable Diffusion can just do text-to-image. 620 01:06:10.770 --> 01:06:13.129 Rohit Gandikota: And for you to change the image, you have to re-prompt. 621 01:06:13.890 --> 01:06:28.839 Rohit Gandikota: And then Flux came out saying, oh, you can also give an image as an input, so now you can give an image and a text and sort of change it. Nano Barana now said, no, no, you can keep on talking to your image again and again, and based on this context, I'll generate better images. 622 01:06:29.150 --> 01:06:32.430 Rohit Gandikota: With each one of them, I think few problems are solved. 623 01:06:32.740 --> 01:06:36.540 Rohit Gandikota: But it's always this, like, there's always researchers that 624 01:06:37.130 --> 01:06:41.229 Rohit Gandikota: that find that, oh, even in Nano Banana, if you work with them. 625 01:06:41.340 --> 01:06:46.239 Rohit Gandikota: I think it's a good novelty. You can definitely get Something out of it. 626 01:06:46.380 --> 01:06:51.020 Rohit Gandikota: But if you really want to use it in, in your homeworks. 627 01:06:51.440 --> 01:06:55.830 Rohit Gandikota: It becomes really hard for you to specify in text what it is you want to do. 628 01:06:56.120 --> 01:07:12.170 Rohit Gandikota: So if there is… yeah. Let's forget about the restriction to text. Maybe the specification is a mix of text and some structured formats, such as the one you showed, this plan… Yes. …that you've generated. Yeah, that's fine. What I'm specifically curious to hear, thought on… 629 01:07:12.590 --> 01:07:21.580 Rohit Gandikota: Do you need to understand internally how the model is working to provide a user with the ability to express direct intent effectively? 630 01:07:21.720 --> 01:07:36.370 Rohit Gandikota: Or maybe if we collect enough examples of what users want, specified in whatever structured format you want, or of semi-structured format, enough examples of that with matched images or videos, train a big model on that, would be enough. 631 01:07:36.690 --> 01:07:39.499 Rohit Gandikota: So, everything that humans want into… 632 01:07:39.960 --> 01:07:46.109 Rohit Gandikota: Well, like… You don't need to cover everything, right? I mean, in… No, I mean, like, most of the things that… 633 01:07:46.570 --> 01:08:04.569 Rohit Gandikota: like I touched upon, like, 3D controls and all the other things. Yeah, I think that's the vision that I'm going towards, too. So I would love if people would build it. I mean, in other words, you don't need to necessarily understand how the neurons work, as long as people can control it, it's fine. So that's what you're… that's your answer, is that right? Yeah. 634 01:08:04.720 --> 01:08:05.920 Rohit Gandikota: Yeah, that's my question. 635 01:08:06.130 --> 01:08:20.620 Rohit Gandikota: Oh, yeah, I think that's my answer, too. So I didn't… so in the last… in the end, when I was showing things, I didn't really allude to, I need to understand what the renderer and, like, the planners are doing. I don't really care about it. 636 01:08:20.620 --> 01:08:36.140 Rohit Gandikota: in that depth. And then I think it still becomes a research interest as to what these models are doing. But to solve the problem that I think we both are talking about, I think you don't need to understand what the model is doing internally. At least there needs to be a way for 637 01:08:36.310 --> 01:08:40.070 Rohit Gandikota: Users to specify what they want, and get the desired output in the right way. 638 01:08:40.680 --> 01:08:41.370 Rohit Gandikota: Yeah. 639 01:08:42.189 --> 01:08:44.710 Rohit Gandikota: Maybe something to consider is that 640 01:08:45.130 --> 01:08:50.790 Rohit Gandikota: Different people might have different medium to express their intention, so some people, like. 641 01:08:50.930 --> 01:08:57.890 Rohit Gandikota: For them, text is, like, the easiest way. Other people will be growing is the easiest way. And as long as you can 642 01:08:58.270 --> 01:09:06.550 Rohit Gandikota: train your, like, pre-train your model with all these committees as, like, the way to express intention, then great, I think you can just do pre-training. 643 01:09:06.720 --> 01:09:13.550 Rohit Gandikota: But if you don't, you might want to understand internalists to enable new 644 01:09:14.000 --> 01:09:22.490 Rohit Gandikota: mediums of communications, or something like that. Like, maybe understanding internals will help you communicate with the model through 645 01:09:22.680 --> 01:09:25.719 Rohit Gandikota: A new, rather than free training. 646 01:09:26.140 --> 01:09:27.860 Rohit Gandikota: With that anymore, Cynthia. 647 01:09:28.359 --> 01:09:33.069 Rohit Gandikota: I can see another reason why you might want it, and that has to do with your discovery. 648 01:09:33.260 --> 01:09:38.890 Rohit Gandikota: Because… The discovery allows you to perhaps find new things. 649 01:09:39.029 --> 01:09:47.740 Rohit Gandikota: that people have not conceived at all, right? And I'm also curious about your take, if you think that we can, like, in the… 650 01:09:48.189 --> 01:10:00.249 Rohit Gandikota: In the space of concepts, sliders or whatever, can you find completely… extrapolate to completely new regions, or are you mostly interpolating between known things? 651 01:10:00.700 --> 01:10:11.219 Rohit Gandikota: Can we discover completely novel knowledge or skills from examining models internally, or are we restricted to what we already do? 652 01:10:14.240 --> 01:10:21.670 Rohit Gandikota: Good question. Yeah, this reminds me of Lisa Schutz's work on AlphaZero. 653 01:10:27.680 --> 01:10:35.870 Rohit Gandikota: Yeah, I… I'm not sure how to go about looking for novel visual Structured. 654 01:10:36.440 --> 01:10:41.939 Rohit Gandikota: I understand that in chess, maybe a little bit, on how you can think about it, it's much more structured. 655 01:10:42.590 --> 01:10:49.110 Rohit Gandikota: structured problem, but with vision, I don't know how to do this, 656 01:10:49.300 --> 01:10:51.379 Rohit Gandikota: Is it novel or not yet? 657 01:10:51.670 --> 01:10:53.109 Rohit Gandikota: I'll think more about it. 658 01:10:53.270 --> 01:10:57.259 Rohit Gandikota: I thought that's what you were talking about when you were talking about the styles. 659 01:10:57.920 --> 01:11:04.879 Rohit Gandikota: Like, there are… people have found out this many slides, but maybe you can go beyond that, and… 660 01:11:05.210 --> 01:11:10.769 Rohit Gandikota: If you do that slider, if you… Look at the slider space. 661 01:11:13.030 --> 01:11:18.700 Rohit Gandikota: Multiple times, maybe you can get a style which is not really Style that people have. 662 01:11:18.890 --> 01:11:33.950 Rohit Gandikota: explode in the real world. Yeah, no, absolutely. No, the point is the same. I'm not saying the discovery methods will give you new… will not give you new. I'm sure they will. I don't know the way to test it. 663 01:11:34.630 --> 01:11:38.400 Rohit Gandikota: Yeah, I think artists, like, a case where it's, like. 664 01:11:38.550 --> 01:11:48.760 Rohit Gandikota: better defined, right? Like, you can define, or not really, it's not really, but it's more defined at, like, general, visually. Yeah, like, yeah. 665 01:11:51.480 --> 01:11:56.970 Rohit Gandikota: You're saying the art is more subjective, that's why maybe we don't get normal. 666 01:11:58.320 --> 01:12:02.649 Rohit Gandikota: No, I'm actually saying the opposite. I'm saying that, like, with art. 667 01:12:03.900 --> 01:12:06.330 Rohit Gandikota: Maybe there is a way to measure 668 01:12:06.630 --> 01:12:13.589 Rohit Gandikota: Like, new, this, like, new concepts, because you can measure the similarity to known concepts, or known textbooks. 669 01:12:14.370 --> 01:12:18.240 Rohit Gandikota: But with the general vision, Yeah… 670 01:12:18.670 --> 01:12:23.339 Rohit Gandikota: But maybe if you go to, like, a scientific regime. 671 01:12:23.570 --> 01:12:29.569 Rohit Gandikota: maybe then it's easier to define some… like, I don't… maybe there is a way to do that in a specific domain. 672 01:12:29.800 --> 01:12:30.460 Rohit Gandikota: Yep. 673 01:12:31.980 --> 01:12:37.269 Rohit Gandikota: a dumb idea, but I want to do sliders for gang warping, like, weird creatures, weird-looking cats. 674 01:12:37.620 --> 01:12:42.440 Rohit Gandikota: Yeah, I think that's… that's what the UI was trying to do. 675 01:12:42.700 --> 01:12:48.199 Rohit Gandikota: Actually, I thought of something else. Yes, it just reminded me of the… it just reminded me of that, 676 01:12:48.790 --> 01:12:55.370 Rohit Gandikota: discovery, I think, 2020 or 2019, somebody at MIT got a picture of black hole. 677 01:12:55.880 --> 01:12:59.369 Rohit Gandikota: But maybe you can create images of… 678 01:12:59.560 --> 01:13:06.149 Rohit Gandikota: some celestial bodies that people have not taken. Very nice, yes. Why not? Yes. 679 01:13:07.410 --> 01:13:12.590 Rohit Gandikota: But yeah, I think it would be nice if that's possible, yeah. Maybe people can check it. 680 01:13:13.800 --> 01:13:16.519 Rohit Gandikota: Something different. Yeah. Yeah, yeah. 681 01:13:17.110 --> 01:13:22.870 Rohit Gandikota: So, any question about, So your plan… your vision is to, 682 01:13:24.410 --> 01:13:27.450 Rohit Gandikota: Decompose the image into some kind of, like. 683 01:13:27.670 --> 01:13:31.059 Rohit Gandikota: some text form, maybe an SVG form, that you can… 684 01:13:31.520 --> 01:13:34.280 Rohit Gandikota: We can really specify and create a message. 685 01:13:34.720 --> 01:13:37.600 Rohit Gandikota: Which is sort of like a new programming language, apparently. 686 01:13:38.290 --> 01:13:42.040 Rohit Gandikota: In my near… the near future vision, yes, yes. 687 01:13:42.360 --> 01:13:45.310 Rohit Gandikota: So, in that programming language. 688 01:13:45.950 --> 01:13:52.810 Rohit Gandikota: Do you imagine that every concept you have discovered, or every concept in human knowledge. 689 01:13:53.300 --> 01:13:56.480 Rohit Gandikota: It becomes some kind of, like, an operation in the programming language? 690 01:13:59.040 --> 01:14:01.969 Rohit Gandikota: Like, how do you, let's say, smile? 691 01:14:02.130 --> 01:14:08.130 Rohit Gandikota: Does it, like, become some kind of, like, a keyword in that program language? 692 01:14:08.270 --> 01:14:12.120 Rohit Gandikota: And also, like… Okay, so for… 693 01:14:12.530 --> 01:14:15.999 Rohit Gandikota: Like, it's currently underway, project. 694 01:14:16.320 --> 01:14:21.790 Rohit Gandikota: And the… what the current version is doing is, it has… 695 01:14:23.360 --> 01:14:36.360 Rohit Gandikota: something we call, like, concept vectors, where on top of the language that it's generating, it also has identity as a JSON key, and the value is just a vector. 696 01:14:36.830 --> 01:14:41.690 Rohit Gandikota: And then for that vector, the model has to always generate the same identity of the person again and again. 697 01:14:42.720 --> 01:14:57.620 Rohit Gandikota: So, like, similarly, like, smile can be also, like, a slider kind of an input, where you can keep sliding in your, the language, and then the output should just change the smile and keep the rest of the other things intact. 698 01:14:58.140 --> 01:15:01.209 Rohit Gandikota: But yeah, we are currently doing it, where… 699 01:15:03.850 --> 01:15:12.480 Rohit Gandikota: Yeah, I should give a talk about this, when it's in a much more baked form, but yeah, like, that's the goal, where we try to change 700 01:15:12.780 --> 01:15:15.660 Rohit Gandikota: Expand everything the model is planning internally. 701 01:15:15.760 --> 01:15:17.739 Rohit Gandikota: Into as much as possible. 702 01:15:18.420 --> 01:15:27.070 Rohit Gandikota: And then generate a note for the composed. They become their… Keywords, so all operational. 703 01:15:27.700 --> 01:15:31.410 Rohit Gandikota: Video… Yeah. 704 01:15:31.570 --> 01:15:35.669 Rohit Gandikota: like, a layered tree structure. Like, a person can be… 705 01:15:35.910 --> 01:15:38.769 Rohit Gandikota: so many things. Eyes, eye distances. 706 01:15:39.910 --> 01:15:44.030 Rohit Gandikota: What apparel, like, shirt they're wearing, and everything should have 707 01:15:44.310 --> 01:15:48.549 Rohit Gandikota: something of its own. So that's the thing. This language doesn't have to be… 708 01:15:49.250 --> 01:15:51.640 Rohit Gandikota: Compact enough for humans to read. 709 01:15:51.990 --> 01:15:52.780 Rohit Gandikota: But… 710 01:15:52.920 --> 01:16:01.990 Rohit Gandikota: structured enough that if they want to read, that they can go in, or an agent can go in and do that. That's the goal for now. It might evolve as we go. 711 01:16:05.050 --> 01:16:12.110 Rohit Gandikota: These models, these image models are so good, because they can compose between different concepts. 712 01:16:12.990 --> 01:16:23.549 Rohit Gandikota: And there, they have some flexibility. You want them to have that flexibility. Yeah, the renderer still has that flexibility. I don't impose anything in terms of decomposition on it. 713 01:16:24.190 --> 01:16:26.630 Rohit Gandikota: He's pretty, pretty individual. 714 01:16:28.140 --> 01:16:38.869 Rohit Gandikota: So how a Python program gets compiled to assembly language. Yeah, I agree, yeah. So you kind of want that underscatification in the language as well. 715 01:16:42.300 --> 01:16:46.289 Rohit Gandikota: Because if you can exactly specify everything. 716 01:16:47.390 --> 01:16:53.830 Rohit Gandikota: I don't know how you'll do that, but it would be like a… Too rigid. Yeah, 1 million line, 717 01:16:54.590 --> 01:16:56.619 Rohit Gandikota: Program to generate something. 718 01:16:56.740 --> 01:16:59.949 Rohit Gandikota: Simple. 719 01:17:00.490 --> 01:17:01.950 Rohit Gandikota: We want to have that. 720 01:17:04.300 --> 01:17:06.780 Rohit Gandikota: Capability, which makes it a little bit… 721 01:17:07.460 --> 01:17:13.669 Rohit Gandikota: uncertainty, not… not kind of, like, if you want to navigate an appropriate language through one of the certain language. 722 01:17:14.620 --> 01:17:22.769 Rohit Gandikota: I'll… I'll push back a little on that. I don't… I… I don't think it's the… 723 01:17:25.680 --> 01:17:29.070 Rohit Gandikota: I don't think I'm going towards making the language as… 724 01:17:29.580 --> 01:17:36.850 Rohit Gandikota: descriptive as possible, that's not what I'm going towards. I'm going in… I'm going towards… I want the language to be, 725 01:17:38.510 --> 01:17:42.650 Rohit Gandikota: Fine enough that users can control the things that they care about. 726 01:17:43.230 --> 01:17:50.800 Rohit Gandikota: And, like, if they want to describe the identity, they don't have to take millions of lines of code to describe the identity. It can be a single vector. 727 01:17:51.470 --> 01:17:52.230 Rohit Gandikota: Right? 728 01:17:52.370 --> 01:17:58.600 Rohit Gandikota: And then there can be another model that looks at this identity vector and, like, can edit it, can change it. 729 01:17:59.150 --> 01:18:06.340 Rohit Gandikota: But the… The goal of this code, is not to be exactly as Python. 730 01:18:06.540 --> 01:18:16.590 Rohit Gandikota: And also not exactly as prior models that can do everything in one single model. Like, it's somewhere in between, where I want the renderer to be something which 731 01:18:16.870 --> 01:18:24.969 Rohit Gandikota: is not… like, quotable. I want that to do all these weird composition things and create realistic images. 732 01:18:25.080 --> 01:18:27.889 Rohit Gandikota: At the same time, I want to have a format 733 01:18:28.080 --> 01:18:35.039 Rohit Gandikota: What the model is planning for, and have some ways for me to change what it's planning without affecting a lot of things internally. 734 01:18:35.930 --> 01:18:44.369 Rohit Gandikota: But it's a good question. I… I don't know how to draw that line between verbose versus… 735 01:18:44.530 --> 01:18:46.640 Rohit Gandikota: controllable. 736 01:18:47.410 --> 01:18:51.329 Rohit Gandikota: I'm happy to talk more about it, I have multiple versions. 737 01:18:51.950 --> 01:18:56.539 Rohit Gandikota: Like, what do you think about, like, the target audience stuff of, like, these tools? 738 01:18:56.960 --> 01:18:58.529 Rohit Gandikota: Because, like, this reminds me of, like. 739 01:18:58.740 --> 01:19:01.790 Rohit Gandikota: when I was a kid trying to learn to use Photoshop. 740 01:19:01.920 --> 01:19:05.379 Rohit Gandikota: It's like, Photoshop can do so much with it, but it's so hard 741 01:19:05.510 --> 01:19:10.200 Rohit Gandikota: to learn, because it's, it's like, it's like the really verbose thing. You could literally do anything. Yep. 742 01:19:10.540 --> 01:19:22.270 Rohit Gandikota: But, like, you need to have a lot of expertise. Yeah. But also, like, you… so, like, it depends on who's using it, right? Like, if it's, like, you know… It's a good… that's a good thing. I think that's why the now is very important in this life. 743 01:19:22.580 --> 01:19:28.849 Rohit Gandikota: We finally have tools that can automate these things. You can wipe… wipe code things, you can, 744 01:19:29.060 --> 01:19:31.970 Rohit Gandikota: Solve most of these problems, if it's code. 745 01:19:32.430 --> 01:19:37.100 Rohit Gandikota: You can solve most of it by just putting an agent to work and talking to the code. 746 01:19:37.220 --> 01:19:39.289 Rohit Gandikota: Instead of actually changing something. 747 01:19:40.810 --> 01:19:55.260 Rohit Gandikota: I think, eventually, if this were to work, there'll be definitely two sets of users, one that are, like, published users that would want to look at the code, change something, and work with them, and then ones that would just want to vibe image. 748 01:19:56.250 --> 01:20:04.519 Rohit Gandikota: talk to it and say, oh, no, move it to the left, and then the code agent would go inside and change just the XY coordinates. 749 01:20:04.820 --> 01:20:06.529 Rohit Gandikota: And then re-render the image. 750 01:20:06.950 --> 01:20:12.629 Rohit Gandikota: It is possible now. It wasn't… like, it would have been another Adobe Premium. 751 01:20:12.820 --> 01:20:15.449 Rohit Gandikota: Photoshop, if it was 5 years ago. 752 01:20:15.900 --> 01:20:17.530 Rohit Gandikota: Maybe now it's not that. 753 01:20:24.700 --> 01:20:32.080 Rohit Gandikota: Do you have any… Suggestions for, like, people that don't feel creative, And, like… 754 01:20:32.280 --> 01:20:34.830 Rohit Gandikota: How would you build a system that, like. 755 01:20:34.950 --> 01:20:39.560 Rohit Gandikota: Helps them be more creative for, like, development. 756 01:20:39.920 --> 01:20:45.909 Rohit Gandikota: Yes, so… That… I think that was the intention behind my slider space. 757 01:20:46.770 --> 01:20:49.960 Rohit Gandikota: slide, where I was showing that I have such small intent. 758 01:20:50.480 --> 01:20:55.499 Rohit Gandikota: So, for example, when we came up with concept sliders, we… 759 01:20:55.630 --> 01:21:00.540 Rohit Gandikota: The maximum number of creative sliders we could do were 10, or 20, so 20. 760 01:21:00.860 --> 01:21:07.059 Rohit Gandikota: As soon as we open sourced it, we had, like, 10K, 20K sliders online that are all super creative. 761 01:21:07.540 --> 01:21:14.209 Rohit Gandikota: So one is through crowdsourcing and open sourcing, which I always have a very strong stance on. 762 01:21:14.630 --> 01:21:20.389 Rohit Gandikota: And another is the discovery aspect of things, where you can allow users 763 01:21:20.580 --> 01:21:22.830 Rohit Gandikota: Oh, this was another thing which… 764 01:21:23.070 --> 01:21:32.600 Rohit Gandikota: was about to be on the slides, but was not, which is, having a paradigm which will allow users to enhance their intent, like social media, except instead of 765 01:21:33.170 --> 01:21:37.559 Rohit Gandikota: Optimizing on the engagement, you optimize on intent. 766 01:21:37.740 --> 01:21:42.439 Rohit Gandikota: and sort of show them, like, when they're working with the system, the UI, 767 01:21:42.760 --> 01:21:49.929 Rohit Gandikota: You show them, oh, things you might like to explore but haven't ever, or similar users have explored before, or something where 768 01:21:50.530 --> 01:21:56.110 Rohit Gandikota: the signal is not engagement, right? It's… It's, like, 769 01:21:56.360 --> 01:21:59.260 Rohit Gandikota: Not explored before, sort of. A small nudge. 770 01:21:59.960 --> 01:22:02.880 Rohit Gandikota: But again, it's a dangerous territory, I don't… 771 01:22:03.860 --> 01:22:06.850 Rohit Gandikota: I haven't thought more about it beyond that thought. 772 01:22:07.200 --> 01:22:13.100 Rohit Gandikota: But yes, I think discovery and some sort of a… Sensible nudge. 773 01:22:13.430 --> 01:22:16.680 Rohit Gandikota: Might help users to explore more inside the models. 774 01:22:18.030 --> 01:22:26.750 Rohit Gandikota: I guess, like, another thought is, like, I actually had the software for 2 years, but I haven't really done it. I'm always confused why people never 775 01:22:27.000 --> 01:22:31.570 Rohit Gandikota: use a gaming engine with your graphics engine to boost traffic controllability. Yeah. 776 01:22:31.840 --> 01:22:49.329 Rohit Gandikota: So we… Oh yeah, that's what we're doing now. We have computer graphics, like code, where you can have actions, move cameras, we can even have OMs to sort of, like, interfere with the computer graphics code to do editing in that code. After you render an image, then it's probably a much easier problem to bridge between the realism gap. 777 01:22:49.420 --> 01:23:02.939 Rohit Gandikota: of, like, what your text image model or text-to-video model can do, versus, like, the graphics-generated stuff. Yes. I mean, NVIDIA already has some stuff that's, like, using AI to turn, like, a graphics thing into something that's more realistic. Yes. 778 01:23:03.280 --> 01:23:04.200 Rohit Gandikota: Then… 779 01:23:04.570 --> 01:23:10.019 Rohit Gandikota: I mean, there's free data over there, I just know I really want to play around with it, that just feels such a pity. 780 01:23:10.190 --> 01:23:18.120 Rohit Gandikota: I know, I think that's exactly what we're doing with this project now. Oh, I see. Yeah, it's a good catch, yes. Very cool. Yes, 781 01:23:18.680 --> 01:23:27.960 Rohit Gandikota: Yeah, I… another motivation actually does come from the game engines that you were talking about, the ability to control your characters and… 782 01:23:28.110 --> 01:23:31.670 Rohit Gandikota: The camera angles, motions… And… 783 01:23:32.110 --> 01:23:34.079 Rohit Gandikota: Yeah, we want to do that, except… 784 01:23:34.490 --> 01:23:37.490 Rohit Gandikota: have an engine where you don't have to learn it, like Photoshop. 785 01:23:37.850 --> 01:23:46.959 Rohit Gandikota: Where the model would do it for you. And the real… realism can also be in the same power that we have with Imagine GPT right now. 786 01:23:47.510 --> 01:23:51.659 Rohit Gandikota: To get to that level with game engines is really hard. 787 01:23:52.300 --> 01:23:55.110 Rohit Gandikota: It requires a lot of… 788 01:23:55.210 --> 01:23:59.200 Rohit Gandikota: Like, dedicated years of experience to get to that point. 789 01:23:59.450 --> 01:24:07.440 Rohit Gandikota: But yes, it's rich of all the things that we want to do with vision models. Games are, like, the perfect example of 790 01:24:07.780 --> 01:24:11.890 Rohit Gandikota: having… an explorable medium of vision. 791 01:24:12.830 --> 01:24:15.520 Rohit Gandikota: So yeah, I think it can be a good data source. 792 01:24:18.220 --> 01:24:18.930 Rohit Gandikota: Nope. 793 01:24:19.450 --> 01:24:25.079 Rohit Gandikota: I am not sure that this architecture you're proposing Does actually, like. 794 01:24:25.820 --> 01:24:31.460 Rohit Gandikota: serve as a medium to, like, translate the intent of the user, who's, like, the person who understands the blog. 795 01:24:31.790 --> 01:24:44.370 Rohit Gandikota: And the example you showed earlier, which was very good, about somebody prompting a model to write, like, an email or, like, a thank you letter for the data or something, the model did understand this would be really sweet, and so you would have, like, real positive consequences on it. 796 01:24:44.760 --> 01:24:54.520 Rohit Gandikota: Now, imagine, this example where I'm trying to develop a, like, drawing for a campaign against, like, speeding on the road, like, a school zone available. 797 01:24:54.970 --> 01:24:56.979 Rohit Gandikota: And my prompt to the model is. 798 01:24:57.280 --> 01:25:03.550 Rohit Gandikota: Draw a spool, and, like, a no-speed sign, and kids crossing the road, and a car coming. 799 01:25:03.980 --> 01:25:12.939 Rohit Gandikota: And the model will probably just, like, try to do it exactly as it is, without trying maybe to add elements that would communicate my message. 800 01:25:13.040 --> 01:25:20.650 Rohit Gandikota: Or the purpose behind this. I think language models do, because maybe language is… I don't know. 801 01:25:20.750 --> 01:25:23.739 Rohit Gandikota: There's something human about it, that, like, emotions… 802 01:25:23.910 --> 01:25:32.009 Rohit Gandikota: Transpare more easily through it, and is easily… more easily, you know, interpretable than, like, images. Of course, you can also do that with our images and other things. 803 01:25:32.270 --> 01:25:48.299 Rohit Gandikota: And so I'm not sure that this architecture does actually, like, capture the human intent, and it's that this is just, like, translating the instruction better, and planning, and it's controllable, seems also explainable. Yeah. No, no, I think, 804 01:25:48.600 --> 01:25:53.090 Rohit Gandikota: With the amount of… thing that I've shown here, I think that's a very good intuition. 805 01:25:53.440 --> 01:26:01.450 Rohit Gandikota: I have two points. I think we can very easily solve it with the techniques we have today. 806 01:26:02.020 --> 01:26:09.229 Rohit Gandikota: One is… You can use a pre-trained language model to start your 807 01:26:09.340 --> 01:26:17.549 Rohit Gandikota: DSL generation from, where you exactly get all the intentions that you are hoping to get, at free cost. 808 01:26:18.180 --> 01:26:25.919 Rohit Gandikota: Another thing is the… the point that I did for my second roadmap, which is visual reasoning, medium. 809 01:26:26.330 --> 01:26:31.839 Rohit Gandikota: I feel like that will capture a lot of things that… what you're talking about in a visual sense. 810 01:26:32.260 --> 01:26:38.660 Rohit Gandikota: But… The… I stopped sharing, sorry. 811 01:26:47.450 --> 01:26:49.579 Rohit Gandikota: Yeah, you'll have to go through this, sorry. 812 01:26:54.010 --> 01:27:01.240 Rohit Gandikota: Yeah, I think this parser can also be a good way to keep interacting with your images as you draw. 813 01:27:01.430 --> 01:27:16.090 Rohit Gandikota: Right? You can… if the model ends up generating something which you don't intend to, and you want to generate, hey, I feel like it's not showing the safety that I want to teach my kids, you can talk to the parser to do it, and it would 814 01:27:16.650 --> 01:27:26.940 Rohit Gandikota: ideally, like, plan better in terms of, oh, given this image, the user is intending something else, let me sort of put this into the play here, right? You can talk your intention through. 815 01:27:27.280 --> 01:27:33.489 Rohit Gandikota: Or, I don't know how the visual medium works yet. I wish I had a… 816 01:27:34.010 --> 01:27:41.859 Rohit Gandikota: more better understanding as to how visual medium reasoning works. But I feel like there has to be some sense of, oh, no, no, no, I'm looking for 817 01:27:42.790 --> 01:27:49.669 Rohit Gandikota: Like, some, you know, Some frightening emotion that looks similar to this image. 818 01:27:49.840 --> 01:27:53.470 Rohit Gandikota: And then, somehow the model needs to make sense out of it in the visual sense. 819 01:27:53.630 --> 01:27:57.099 Rohit Gandikota: But… I think it's definitely possible. 820 01:27:57.560 --> 01:28:01.509 Rohit Gandikota: It's just a… Needs to be… 821 01:28:01.620 --> 01:28:04.919 Rohit Gandikota: captured in the right place, and using the right medium, I think. 822 01:28:05.550 --> 01:28:15.760 Rohit Gandikota: If you can already say that, hey, my language model does it, I think this can do it too. You just replace one of these parts, or fine-tune one of these parts on top of the model you're talking about. 823 01:28:19.090 --> 01:28:20.880 Rohit Gandikota: Maybe they're moving along. 824 01:28:20.990 --> 01:28:27.830 Rohit Gandikota: language models, like, me doing it. Yep. Like, the… 825 01:28:27.930 --> 01:28:37.509 Rohit Gandikota: Medium that the training data is presented in contains very rich contextual information, whereas maybe images may be, like, isolated from, like, their role in the world. 826 01:28:37.820 --> 01:28:39.420 Rohit Gandikota: That's what that's why it's not able to watch. 827 01:28:39.720 --> 01:28:47.939 Rohit Gandikota: capture this intent that was left over. Good point, yeah. I think pre-training might still be necessary eventually as we go down this path. 828 01:28:48.460 --> 01:29:00.729 Rohit Gandikota: You might still want individual vision models, language models, like Dino, and you might still want all these things that can capture the modality information from everything that you want. 829 01:29:01.190 --> 01:29:07.269 Rohit Gandikota: But, like, if you want to get to a place where you can do all this 830 01:29:08.440 --> 01:29:14.819 Rohit Gandikota: amazing things. I think… we'll have to move to a better fine-tuning regime, I guess. 831 01:29:21.300 --> 01:29:22.640 Rohit Gandikota: Oh, there's a comment. 832 01:29:25.530 --> 01:29:26.640 Rohit Gandikota: From Chris. 833 01:29:27.150 --> 01:29:28.400 Rohit Gandikota: Thank you, Rohit. 834 01:29:32.370 --> 01:29:39.340 Rohit Gandikota: Can't stop you anymore with all these amazing… Oh, you're ready to defend. Thank you, Chris. 835 01:29:40.120 --> 01:29:50.019 Rohit Gandikota: Oh, no, Chris… Too early. Not yet. He's ready, he's ready to propose. 836 01:29:51.230 --> 01:29:54.230 Rohit Gandikota: One more stage. One more stage. 837 01:29:54.330 --> 01:29:56.720 Rohit Gandikota: Nance. Thank you, but on that note. 838 01:29:57.150 --> 01:30:02.959 Rohit Gandikota: Really nice compendium of all your work and your vision for your next stage. Thank you, Rohit. 839 01:30:07.810 --> 01:30:08.830 Rohit Gandikota: Thank you, hon.