E37: The Mind-Reading Revolution with Dr. Tanishq Mathew Abraham (Part 1 of 2)

In this episode, Nathan sits down with Tanishq Mathew Abraham, 19-year-old UC Davis grad and one of the youngest people in the world to receive a Ph. D, with a degree in biomedical engineering. Tanishq is the founder of the Medical AI Research Center (MedARC), and with his teammates, recently published a paper: Reconstructions of the Mind's Eye, which encompasses their breakthrough research on reconstructing visual perceptions from fMRI scans into images. In this episode, Nathan and Tanishq talk about the technology behind the fMRI-to-image project, developing the model, and future applications for this research. Part 2 with Tanishq will be released as the next episode. We're hiring across the board at Turpentine and for Erik's personal team on other projects he's incubating. He's hiring a Chief of Staff, EA, Head of Special Projects, Investment Associate, and more. For a list of JDs, check out: eriktorenberg.com. RECOMMENDED PODCAST: The HR industry is at a crossroads. What will it take to construct the next generation of incredible businesses – and where can people leaders have the most business impact? Hosts Nolan Church and Kelli Dragovich have been through it all, the highs and the lows – IPOs, layoffs, executive turnover, board meetings, culture changes, and more. With a lineup of industry vets and experts, Nolan and Kelli break down the nitty-gritty details, trade offs, and dynamics of constructing high performing companies. Through unfiltered conversations that can only happen between seasoned practitioners, Kelli and Nolan dive deep into the kind of leadership-level strategy that often happens behind closed doors. Check out the first episode with the architect of Netflix’s culture deck Patty McCord. https://link.chtbl.com/hrheretics The Cognitive Revolution is a part of the Turpentine podcast network. To learn more: Turpentine.co TIMESTAMPS: (00:00) Episode Preview (05:43) The MindEye Project (09:06) Resemblance between AI reconstruction of mind's eye and visual presented (10:00) What is a voxel and which regions of the brain were studied? (10:23) What would the raw data of a voxel be? (11:44) Is there a time dimension to voxels? (15:00) Sponsor: Omneky (17:50) Goals for the MindEye project (25:57) What is the starting point of the model? (31:15) Aligning the model: reconstruction vs retrieval (40:34) Would doing a full end-to-end training be fine for the reconstruction? (42:15) The role of a limited data set (43:09) Training separate models per subject (45:07) Generalizability with a limited dataset (47:20) Mapping from one high-dimensional space to another (50:47) Stable Diffusion VAE encoding (1:00:50) How long does it take to train the model? (1:03:14) How similar or different are the subjects and their individual models? (1:05:59) The future of this research: custom models for your brain? (1:07:34) How much does this research contribute to brain research and wearables? (1:11:15) Fuzzing data and future research applications LINKS: MedARC: medarc.ai MindEye Paper: https://www.researchgate.net/publication/371136623_Reconstructing_the_Mind's_Eye_fMRI-to-Image_with_Contrastive_Learning_and_Diffusion_Priors MP3 of this episode: https://chrt.fm/track/993DGA/traffic.megaphone.fm/RINTP1584997572.mp3?updated=1687271014 TWITTER: @iScienceLuvr (Tanishq) @MedARC_AI (MedARC) @CogRev_Podcast @labenz (Nathan) @eriktorenberg (Erik) SPONSOR: Thank you Omneky (www.omneky.com) for sponsoring The Cognitive Revolution. Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work, customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off. MUSIC CREDIT: MusicLM

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  I think this idea of of mapping one latent space to another is a very powerful idea. I think it's always best to try to take advantage of that as much as possible. And the real, I guess, innovation these days is to be able to use these multimodal spaces as well. And being able to map, you know, different things to these to these multimodal spaces. That's kind of a really exciting area.

• 0:21Nathan Labenz?

  Hello and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs and builders working on the frontier of artificial intelligence. Each week we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life and society in the coming years. I'm Nathan Labenz, joined by my co-host, Erik Torenberg. Hello and welcome back to The Cognitive Revolution. Our guest today, and also for our next episode, is Tanishq Matthew Abraham. Tanishq is a remarkable talent. He started college at seven years old, gave his first TED Talk shortly thereafter,

• 1:02Nathan Labenz?

  and has recently completed his PhD, just as he turned 20 years old. He is also the CEO of MedArc, an organization he founded with the goal of translating AI progress to real-world medical applications. If you listen to this show, you know that much of the incredible AI progress we've recently seen has been unlocked by a strikingly consistent recipe. A clever, easy-to-score objective, called a loss function, allows the transformer architecture to be scaled up and trained with unsupervised, often web-scale data. This approach has repeatedly advanced the state of the art

• 1:44Nathan Labenz?

  in text generation, image understanding, image generation, and so much more. Against this backdrop, Tanishq has recently published two remarkable papers with two different sets of co-authors, which show that there is still plenty of opportunity to produce breakthrough results, even with relatively small data sets and modest compute budgets, by using more thoughtfully designed, problem-specific architectures.

• 2:13Nathan Labenz?

  The first, called "Reconstructing the Mind's Eye", which we'll cover in today's episode, shows that AI can quite literally read minds. Leveraging pre-trained clip and stable diffusion models, Tanishq and his collaborators are able to take FMRI scan data, collected while a person viewed a particular image, and reconstruct that same image with remarkably high fidelity. Our conversation doesn't depend on visual aids, but to properly appreciate this result, you absolutely must look at the headline graphic of this paper. Follow the link in the show notes and I promise you will be impressed. For me, this conversation was not only a chance to learn

• 2:57Nathan Labenz?

  about some of the latest techniques that leverage the power of pre-trained foundation models in small data environments, but also a chance to reflect on the big picture. This result, like so many others we've seen recently, would have felt like science fiction just a couple of years ago. Now, with the pace of AI progress so relentless, they often seem to come and go with minimal fanfare. And just as important, I think Tanishq's work demonstrates the potential for AI to help advance neuroscience in general. Decoding the brain's activity in a non-invasive way holds the promise of helping us understand our own human cognition like never before, potentially bringing an entirely new meaning

• 3:40Nathan Labenz?

  to the phrase "cognitive revolution". As always, if you're finding value in the show, I would really appreciate it if you'd take a moment to share it with your friends. We recently hit a new high as the number 15 rated show on Apple's Technology Podcast Chart. And while I'll always prioritize depth over reach, it is extremely gratifying to know that this project is helping others make sense of the rapidly evolving AI landscape.

• 4:07Nathan Labenz?

  So now, without further ado, I hope you enjoy this conversation with Tanishq Matthew Abraham. Tanishq Matthew Abraham, welcome to the Cognitive Revolution.

• 4:17Speaker 00

  Thank you for having me.

• 4:18Nathan Labenz?

  Very excited to have you. You are prolific all of a sudden, if not, if not always. For folks who don't know you, you are a child prodigy, I think, fair to say, who's now grown into a young adult AI researcher, who is hitting on the level of, from what I can tell, a tenured professor. So extremely impressive young career that you have and a couple of papers that have come out in just the last week or so, both of which are super interesting and for a whole bunch of reasons. One about reading the state of the brain and translating what is detected in the brain into a reconstruction of what the person saw, which is pretty amazing. And there's some incredibly striking visuals there.

• 5:03Nathan Labenz?

  And then one also that is about taking raw images of tissues and predicting what those would look like if they were treated in various ways to then allow like a lab technician to read, you know, what information they contain and make diagnoses. So super fascinating stuff, really where the rubber hits the road kind of research. This is not, not pie in the sky at all, but like very sort of applied, practical, hands-on stuff. And I think it's just fantastic. So great work for starters. And I'm looking forward to getting into all the details.

• 5:37Speaker 00

  Yeah, sounds great.

• 5:39Nathan Labenz?

  So let's talk about the first paper first. This is the one that has, I think, blown up a little bit more on social media. It's called "Reconstructing the Mind's Eye: FMRI to Image with Contrastive Learning and Diffusion Priors." Maybe just for starters, I was surprised to learn, if I understand correctly, there is a FMRI data set that was already out there in the public for you to build on. So maybe just start off by kind of telling us what you began with as you undertook this project.

• 6:08Speaker 00

  Yeah, it seems like, I think the field of neuroscience also tends to have somewhat of a kind of open. Approach to research in terms of releasing data sets and you know there are these sorts of databases with you know neuroscience data sets like EEG data sets, fMRI data sets. So luckily, you know, there has been some of these data sets have been released where you know, they they they take, you know, a subject, ask them to look at, you know, some images and they measure their, the fMRI signal at the same time. And so this was, so there have been a few data sets over the years that have done similar sorts of things. The data set that we used was released, I think maybe it was 2020, 2021,

• 6:52Speaker 00

  so pretty recently, I think. So it was, it's a pretty recent data set. It's a, it's a very high resolution data set. So it's used to uses a, so, so most of the data sets are using like, I think magnet that is used for the MRI. It's I think like 1.5 Tesla, 3 Tesla, something around this range. Whereas the one that this one is using is 7 Tesla. So it's a much more powerful magnet. So it actually results in a much more powerful and higher resolution fMRI signal. And so you have a better signal that you can also measure as well. So this was created by another team that, that conducted that, that research. And, you know, they, they, they measured the, the, the signal for many different subjects.

• 7:32Speaker 00

  I think about eight subjects in total. And, you know, the subjects went through hours of, of looking at images. Maybe they look at each image for a few seconds and then the signal would be measured and then there'd be a break of another few seconds and then they measure again for another image. So they have that sort of process. And, and then, yeah, they, they collected all of that data and they released it publicly. And so, yeah, actually, I wasn't originally aware of this data set. It was only after meeting my collaborator here, who's the, the first author of the paper, who, who told me about this data set. Actually, we had started working on some similar ideas for this project. You know, about a year ago or so, and we were actually looking at a different data set at the time. We were really happy to see this data set, this new data set that, you know,

• 8:17Speaker 00

  is much higher resolution and much better signal quality. And yeah, it was, you know, it was really designed for these sorts of, these sorts to research these sorts of questions. So yeah, it's really a great data set. And, you know, it's great to see that the neuroscience community also has this sort of open research spirit. So I'm really happy about that. And, you know, it just kind of makes me more excited to work in that field as well.

• 8:39Nathan Labenz?

  Yeah, well, the interaction between AI and many, if not all, aspects of biology is certainly going to be fascinating. And this is one that people are just so struck by because you can look ultimately at, this is what we showed the person, and this is what we were able to reconstruct based on a reading of their brain. And, you know, that first figure in this paper is incredibly striking, where it's just like, wow, that is not a, you know, similar, I mean, it is a similar image, but that's not just a somewhat similar image. That is a very similar image that really looks like what you, you know, what you showed the person. So we can put a link, and we will put a link into the show

• 9:21Nathan Labenz?

  notes to the paper. But go check out the, you know, that first visual, pause right here and go look at it, because it really does ground what we're talking about. The striking resemblance of the reconstructions is just incredible.

• 9:33Speaker 00

  The reason why it's doing so well is both because of a combination of this very high quality data set, as well as, you know, the latest AI advances that we're taking advantage of. And I think these two, because of these two factors, that's why we're able to see such such great results. And so, you know, this wasn't, this isn't something that was possible before, but only because of these two factors now, we're able to see such amazing results.

• 9:55Nathan Labenz?

  Cool. So let's talk just a little bit more about the FMRI side and kind of the nature of that data set. So reading through the paper, there's 15,000 voxels. A voxel, as I understand, is a 3D space in and of itself, right? So basically, 15,000 little cubes, if you will, you know, each corresponding to a physical region in the brain. Is this all kind of like on the back of the head in the visual cortex? Like, how much of the brain is sort of under consideration here?

• 10:26Speaker 00

  Yeah, that's approximately right. We took a, we, the data set actually has the FMRI signal of the entire brain. But we used a subset of the voxels that correspond to visual perception. So on, yeah, and this was actually something that the original data set folks actually prepared. They actually prepared a subset of their data with focused on visual perception. And so that is the subset that that we used in this in the study.

• 10:50Nathan Labenz?

  If we were to look at the sort of raw data of one voxel, like what would that contain?

• 10:58Speaker 00

  I think it's just be kind of a value associated with that particular voxel for that particular the signal. So the signal is kind of measuring the FMI signal is actually measuring kind of it's dependent on the blood oxygenation level. So basically, you know, as as a particular, you know, region in the brain is using, you know, there's when there's a lot of activity. It's going to be using up a lot of blood oxygen. So you kind of see that sort of change in blood oxygenation level. And so you have a value that's associated with that particular voxel that is indicative of the sort of blood oxygenation and the usage of blood oxygen in that in that area of the brain.

• 11:32Nathan Labenz?

  Gotcha. Okay, so it's essentially activity or energy. You know, it's just it's a raw kind of scaler that says this is how intense activity was in this particular place. Is there a time dimension to it as well? Or are you kind of just staring as the user at an image, you know, until there's some sort of, you know, steady state or whatever that is then kind of averaged? How does that work?

• 11:55Speaker 00

  Yeah, kind of like that. There is isn't much of a time dimension. FMI isn't doesn't have as much high temporal resolution, especially because, you know, you do have this sort of blood oxygenation and that process is a little bit slower. And so it's it's much harder to get very, you know, at kind of high accuracy in terms of the time resolution. So it's more of kind of a, yeah, a single kind of steady state kind of value. So that's why, you know, you have the patient looking at it for a few seconds first. And then you take that measurement for over a course of a few seconds. And then, yeah, that's kind of the measurement that you're using. Yeah.

• 12:29Nathan Labenz?

  So essentially, the input is, you know, when we think about an AI model that you're now going to develop on this data set. As kind of an input-output device, you know, I always kind of fall back to that framework. The input then is 15,000 numbers that correspond to intensity of activity in each, you know, one of 15,000 little regions of the the brain that are all kind of in that back visual cortex area. That's strikingly not that much information, you know, 15,000 numbers in the grand scheme of things feels small. Does that feel small to you?

• 13:08Speaker 00

  Yeah, it is. Yeah, it definitely is kind of striking in terms of how much information is present in the fMRI. And again, like, this isn't looking at individual neurons or anything like this. This is kind of a a region. It's about one to two, like, I guess, two, maybe two millimeters, two by two by two millimeter cube, kind of like, basically, like, yeah, it's a two by two by two cube millimeters kind of cube of of area or volume that you're looking at for that particular voxel. So it's not like a very, you know, it's not even that fine-grained. It's it's more fine-grained than maybe some of the other technologies like EEG or some of these other technologies. But like, it's not like a lot of these sorts of systems where you have these sorts of invasive, you know, measurements of specific neurons or things like this.

• 13:53Speaker 00

  It's not looking at specific neurons. So, you know, it's it's kind of interesting that, you know, you would expect you would need the actual neuron neuron signal, the actual electrical activity, and, you know, at that very fine-grained level to be able to get the, you know, to be able to get that response, to be able to accurately predict the the reconstruction. But it turns out that's not absolutely necessary. And I think a lot of it has to do with just the way the the brain is organized and, you know, certain, you know, there are certain regions that respond to certain features in an image, for example. So just being able to know, like, what what regions are being, what regions are being activated and responding to to the image, that kind of gives you an overall idea of maybe the sorts of features that are there in an image already.

• 14:35Speaker 00

  So I think, yeah, a lot of it has to do with kind of the organization of of the brain and the visual perception of the brain. And, yeah, and that's also kind of partly why these sorts of these sorts of approaches are interesting from a from my research perspective is to also better understand how that perception works as well. And you know, yeah, what's being activated? What's uh, yeah, what kind of signals are there?

• 14:59Speaker 02

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 15:03Speaker 00

  So, uh, yeah, overall, yeah, it sometimes can be, when you think about it, it's a little bit surprising. I mean, I think it's just surprising overall that you can, you know, do this sort of reading of brain activity in the first place. You know, whether or not it's with F M I or any other thing, it's that's like, that's still kind of pretty impressive. And doing it in a non-invasive manner, it's like, that's pretty surprising. Um, but yeah, of course, you know, you still are getting very, you still have to get that very high quality data and you know, you know, you have, you have doctor, you have subjects that are, you know, in this MRI machine, sitting there for extended period of time. And so it's still a very, uh, in, you know, kind of involved process anyway. But you know, it's a, it's a great kind of first step and still useful for hopefully research applications. Yeah.

• 15:49Nathan Labenz?

  I mean, there's, the learning is going to go both ways here for sure. I think that's another, a super key point and probably key theme is like the feedback cycles and dynamics here. Research wise is, you know, we're just scratching the surface of that. Again, just grounding for a second. So you said kind of two millimeter cube is one voxel. So if I were to do a little, you know, mental math on that, I'm like, okay, five by five by five is 125. So there's, there would be in a cubic centimeter, there would be 125 of these voxels. And we're on our way to 15,000. So that would be basically 100 cubic centimeters.

• 16:30Nathan Labenz?

  So if I'm thinking about that as like the cortex, it's maybe like a 10 by 10 by one deep kind of segment of, of tissue is what we're really looking at here.

• 16:40Speaker 00

  Yeah, I think that sounds about right.

• 16:42Nathan Labenz?

  We've got that kind of slab of tissue. We've divided it up into these two millimeter cube voxels. Each one has a number that corresponds to its intensity. That is again, where we started. And so that's the input now to the things that you are training AIs to do. I guess there's a couple of big themes here that, that jump back to me. One, the

• 17:09Nathan Labenz?

  Just the volume of data, not that high, right? You've got only so many images, only a few patients. And this is true of your other paper as well, right? It's a low data environment and that creates some interesting challenges. Tell me about the outputs. You have multiple different approaches to create different outputs that you use for different purposes. So maybe just run those down for us.

• 17:29Speaker 00

  The final goal was to, of course, reconstruct the image. We want a final image there. The approach that we wanted to take was kind of, how can we leverage sort of the existing

• 17:43Speaker 00

  image generation models and image representation models that exist already to be able to perform this reconstruction as well as this additional task that we talk about, which is retrieval. And so in order to do that, the goal is to basically get some sort of clip embedding. So this is again, like sort of clip models that OpenAI has developed and released

• 18:09Speaker 00

  where it's kind of a sort of joint representation of an image and text. And so it's in this sort of, you have this sort of representation space. And the idea is that lots of, a lot of these image generation models already taken some sort of clip embedding. So if we're able to predict the clip embedding from the FMI signal, then we can use that clip embedding for image generation and get a reconstruction that way. So that's kind of, at a high level, the approach. There are a few different other aspects that allow this to work. So the idea is that, okay, we have this sort of neural network and a basic MLP, multi-layer perceptron.

• 18:52Speaker 00

  That's like, you know, a basic neural network architecture. And so we take this MLP to predict the embedding.

• 19:02Speaker 00

  And the problem is that the embedding that we get at the beginning isn't aligned. So it isn't so so basically, when you when you have these embeddings, like you can get you can get them so that the sort of, I guess, cosine similarities, the similarities between the embeddings kind of all match up the way you want them to. But the actual representations aren't kind of aligned with the other representations. It's not just a problem with, you know, in this case, F M R I embeddings that we're working with, but it's actually an issue that happens with even just within the regular clip, where you have image to text, image and text embeddings. So for example, in the OpenAI DALL-E work that they have, they wanted to take text and then they wanted to, you know, convert that, of course,

• 19:47Speaker 00

  to a nice image generation. So the problem is that they they did something similar where they actually took text and then they got the clip embedding for the text. And, you know, that's that's they just get that because that's how the clip model is already trained. But then that clip text embedding isn't aligned again to the to the image embedding. So of course, the clip text embedding, well, you know, for example, if it's, let's say it's a picture of a dog and that's the text. And then if it's the actual image of the dog, those those embeddings have high similarity to each other, the text embedding and image embedding, but they're not of similar values. So that's the issue. They're not really of similar values. They're not really matching. So you cannot take, like, for example, a clip text embedding

• 20:31Speaker 00

  and pass it in as a clip image embedding to an image generation model that expects clip image embeddings. You need to somehow convert from the clip text embedding to a clip image embedding. So the DALL-E paper introduces Diffusion Prior, which is another model that converts the clip text embedding into a clip image embedding and aligns them. And so that's what we did here, where we predicted our clip embedding from the F M R I signal. And then we also had a diffusion prior that would take that predicted F M R I embedding and align it with the clip image embedding. And now we have a clip image embedding that's aligned, and that can be passed into a pre-trained image that image generation model. So that's the reconstruction pipeline.

• 21:15Speaker 00

  And then we also have this retrieval pipeline where in this case you don't really need to have the alignment because here all you're looking at is similarity. So you can just say, okay, if you have some sort of fMRI embedding, what images are those similar to based on, again, on cosine similarity of the clip embedding? So you can get the similarity to a bunch of images and you can say these are the ones that it's most similar to. And that we also did those experiments as well. And we can see that like if we used images from the data set that we worked with, then the ones that it's most similar to are the actual images that, you know, the subject actually saw in that case. Or we can also extend it to a very large data set like Lion 5B. And, you know, you can do, again, some sort of similarity and get similar images.

• 22:00Speaker 00

  And those similar images would look very similar to the actual image. And you can kind of treat this as a generation pipeline without actually using, you know, a generation model because, you know, you're getting these sorts of, you know, similar images that, you know, could correspond very, very closely to the original image. And again, the nice thing about what we show in this paper is that you can get actually very fine-grained information. So, you know, if you had, the example that we have in our paper, which we think is a really nice example, is that there are actually a lot of zebra images in the data set that they were working with to collect from some of the subjects. And so if we have one zebra image that a subject looked at and you try to predict

• 22:43Speaker 00

  the fMRI signal and you look at the retrieval of the images, you can actually retrieve the exact same zebra image based on the fMRI embedding that we predict with our model. And so this indicates that, you know, it's not just, you know, oh, this is a zebra image. You can get, this is the exact zebra image that the patient was looking at. So there's actually that fine-grained level of information that's also present. So that's kind of the two sort of approaches that we have here. We have the sort of reconstruction, we have this retrieval. There's one other aspect that is also worth pointing out is that because with reconstruction, you're having this, yeah, you have this clip embedding and that's being passed into this image generation model, you get some nice image.

• 23:24Speaker 00

  But it would be also nice if, you know, the sort of positions of the objects are similar, or maybe the color is similar, these sorts of more low-level information. So we also have a low-level pipeline that also kind of helps predict.

• 23:41Speaker 00

  That that can help with that low level information. And so the idea there is you actually just have a simple model that takes in the FMI data and predicts the a signal in the like a VAE representation. So like with stable diffusion, you know, it has this sort of VAE, which is a which is a variational autoencoder. So this is just kind of a a lower dimensional representation of the image. And so if we can predict that and then we actually just then decode that VAE representation that was predicted from the FMI signal. And that's kind of a very blurry image, but it has sort of some of the features that are important, like some of the position of the objects or the colors, things like this, and we can use that as a starting point for our diffusion process.

• 24:26Speaker 00

  So because our standard pre-trained image generation is usually some sort of diffusion process, we can use some of these sort of like image to image processes that people have been using and use this low level image as a starting point for the high level image generation to occur. And so that way you can get the high level image generation with the low level information. So that's kind of this combination of this sort of low level pipeline and this high level pipeline. And this allows us to get, you know, very high metrics also on the low level metrics. So, you know, people will do things like look at the, yeah, just kind of again, the comparison of the original image and the and the and the output image. And of course, if if you have very similar semantic information in terms of like, okay,

• 25:10Speaker 00

  this is a zebra, but the zebra is not in the same position. If you're going to, you know, directly compare those images, they're gonna it's not going to be very high, high, the the results won't be very good in terms of that sort of low level comparison. But here we can show that our low level comparison actually is much better than previous approaches as well. So, yeah, that's a that's a lot of information, I guess. But those are the different pipelines. And, yeah, of course, feel free to ask questions about the specific pipelines and the specific details. But, yeah, hopefully, hopefully that makes sense.

• 25:37Nathan Labenz?

  So am I right, first of all, to say that kind of the the simplest first step is taking a 15,000 voxels, each a number, and mapping that to a clip embedding?

• 25:54Nathan Labenz?

  And that's kind of the starting point regardless of then whether you're going to go to the retrieval or to the reconstruction.

• 26:00Speaker 00

  Yes, that's correct. Yes.

• 26:01Nathan Labenz?

  So I guess first of all, how big is that? How big is a clip embedding? Like, how many numbers is that at the end of the prediction?

• 26:08Speaker 00

  It depends. Yeah, because we did try different clip models. I don't remember which one we used finally. But it's the, basically, you have a sort of a vector for each. If in the case of text, it would be each kind of word or token, for example. And then those would be concatenated together. Or in the case of an image, you know, for example, if you're using the sort of transformer, you have like image tokens as well. So you have, you have, you have this sort of array of vectors that you have for each image or each sentence or whatever. And then typically, what people will do is they like, for example, like, average across that and just get a single vector that they work with. We don't actually do that.

• 26:50Speaker 00

  Because it turns out if you actually do average, you know, if you average across all the tokens, all the, you know, all the words in the sentence or the entire image or whatever, you actually lose information that way. And if you actually keep that information and try to predict, basically, we're predicting, you know, the clip embedding for the whole image and not just that kind of global representation, but actually the different parts of the image as well. Again, that allows you to get some more spatial information, some more kind of low-level, accurate information that way, too. But it's basically some sort of a vector for each of these tokens. And then, you know, you have that. So overall, it's, yeah, I mean, I can quickly look up the dimension if you. It's probably mentioned in the diagram. Yes. Yeah.

• 27:35Nathan Labenz?

  257 by 768, if I'm reading the diagram.

• 27:39Speaker 00

  Yeah, so to clarify what that means, basically, you have, for each token, you have a 768 vector. So, for example, if you had a clip. So, yeah, again, clip image embeddings are a little bit different because, basically, the image is divided into multiple patches. And then you have, then for each patch is treated as a sort of token. So then you have a 768 vector for each patch in the image. And they have 256 patches. So there's 256 patches. So there's 256 of those 768 vectors. And also, there's one global patch to global representation for the entire image. So that's another, you know, another 768 vector.

• 28:23Speaker 00

  So that's how you get 257 by 768. And so instead of just predicting, you know, a single one by 768 vector for the entire image, we actually predict that 257 by 768 for, you know, each patch plus the full global representation. So maybe hopefully that makes more sense in terms of what's being predicted here for the, for the clip image embedding.

• 28:46Nathan Labenz?

  Yeah, interesting. So that's just, I mean, it's a dramatic increase in the number of numbers, right? I mean, you're going from 15,000 to something like, whatever, 150,000, maybe 200,000 numbers. So that's always kind of interesting in and of itself. Now, is that, so you're predicting there the image embedding?

• 29:10Speaker 00

  We're trying to predict the image embedding. But again, when we predict it, it's, you know, predicting that this is sort of, you know, we have these different losses, like a catastrophic loss and these, all these different losses. So again, when we predict it, it doesn't, it's not necessarily aligned from the very beginning to the image embedding. So it's not going to have the exact same values as the image embedding from the, you know, the, when we start out. And so that's why there's this additional process for the reconstruction. But you can use that predicted image embedding for retrieval, because the similarities will work out perfectly fine anyway, without needing to do this alignment process. So it's, it's just, you know, we predict these image embedding.

• 29:54Speaker 00

  So that's why in the paper, we don't call it necessarily image embeddings, the direct, direct prediction from the MLP. I think we call them like FMRI embeddings or something like this. But it's not, you know, the direct output from the MLP isn't going to match exactly with the clip image embeddings. But, you know, that's kind of what it's trying to be, what we're trying to predict there. Yeah.

• 30:14Nathan Labenz?

  And that's enough to do the retrieval out of like the lion 5 billion data set, often down to even the single one image out of 5 billion, correct?

• 30:26Speaker 00

  Yeah, that's basically correct. Yeah, basically. And so. Again, yeah, we have this sort of contrastive loss that we're training the model, the MLP with. And so that allows, again, for the cosine similarities to work out because we're doing this sort of contrastive learning similar to how the original clip models were trained. And so then we can get the sort of embeddings that work well with this sort of retrieval task. Yeah.

• 30:49Nathan Labenz?

  The next step is doing this sort of aligning. And I'm a little confused there as to was there something preventing you from doing that in an end to end training sort of way? Or was it just a matter of kind of, we knew we wanted to do this in pieces anyway. So this was kind of a an efficiency. Like, why is that two steps instead of one end to end?

• 31:10Speaker 00

  So it turns out to be a bit of actually a sort of trade off between retrieval performance and reconstruction performance. It's hard, if you do like end to end, it would be hard to get the best performance for both, both, both tasks. Dividing it up into these two processes, these two pipelines. This allows us to get, you know, we can, we have something that we can do well for retrieval task. And then we can get something that does well for, for image reconstruction. So that's kind of the, the main motivating factor there for, for dividing it up. And again, we have some ablation studies in the, in the paper that kind of demonstrates this, this, this concept of the sort of trade off between reconstruction and retrieval.

• 31:50Nathan Labenz?

  Is there an intuition for that? I'm a little drawing a blank as to why I would expect, you know, that they would, I currently don't expect. So you have to help me develop the intuition for why that would be the case.

• 32:02Speaker 00

  I think it's potentially because there are some, yeah, like what may do, like in terms of getting that nice cosine similarity that sort of, again, you know, making sure these sorts of embeddings are similar to some images and different to others. I guess that's kind of different, I guess maybe different information is needed or different sort of, yeah, representations may be needed for that approach versus if you want to just take something to predict the, the final image. I guess there may be some differences. Yeah, I mean, it's even for us, and maybe we don't, I'm not sure if we have the, the best intuition as well for what's going on there. But yeah, I guess there are some, you know, there's a different information that may be needed for these two different tasks, I guess.

• 32:45Speaker 00

  Yeah, it's a, it's an interesting question, why, why the, the reconstruction and the retrieval have, yeah, this sort of trade off.

• 32:53Nathan Labenz?

  But the, so if I, again, understand correctly, the, it is a sequential.

• 32:59Nathan Labenz?

  Process even in the inference step, right? Like, maybe I have this wrong, but what I understood was

• 33:08Nathan Labenz?

  voxel data comes in. It's first then mapped onto the clip embedding space in such a way where it'll have this, you know, using the contrastive learning. So it'll have this high similarity so that you can perform your database search with like a vector database. For folks who want to go deeper on that concept, we have a whole episode on vector databases. With, with Anton Terenikov, who is from Chroma, one of the founders of Chroma. And they have a powerful demonstration of their technology built on this same data set, you know, that that really flexes the scalability of it. So go listen to that if you want to think about, if you want to learn more about the the nature of kind of doing these vector database searches.

• 33:53Nathan Labenz?

  But, okay, so you've got that. That's enough then to power the database search. And then you're doing another transformation sequentially, right? Into the, the form that would then feed into the diffusion model.

• 34:06Speaker 00

  Yes, that's, that's, that's pretty much the case, yes.

• 34:10Nathan Labenz?

  So all the information is there the whole time, like that intermediate step. It contains all the information by definition, right? That then gets kind of reformulated. So is it maybe the case that just like the, the constraints that you're working with are different across these tasks? Like, maybe it's not sort of a fundamental thing and more like, well, what the diffusion model needs is just different from what, you know, the sort of you're training not necessarily for informational content, but like for a specific representation that is required for to use some frozen existing system.

• 34:48Speaker 00

  Exactly. Yeah, that was one of the things as I mentioned, because the idea is that you have a diffusion model that's taking in, you know, specific clip embeddings. And so The problem is that those embeddings that we produce, again, they're trained with this contrastive loss. Basically, you can either train with the contrastive loss, or you can train with like something like an MSC loss, which is more like, yeah, you can directly compare, you know, you can get the values to match exactly. But so if the contrastive loss, that's good for that retrieval sort of task. So when we train with the contrast of loss, it's good for that sort of retrieval stuff. And so we're using that for the retrieval task. But then the problem is, again, those representations may not necessarily match a regular clip image embedding.

• 35:32Speaker 00

  Again, it's useful for doing this sort of cosine similarity, getting similarities to existing embeddings, but they don't match those sort of existing embeddings directly. And yeah, again, there was actually, we have a good example of this sort of alignment process in the appendix of our paper. And so if there's this sort of UMAP depiction of the clip image embeddings, as well as the FMI embeddings, you can see that without any sort of, without our sort of alignment diffusion prior process, you could see that the embeddings are kind of in two separate clusters almost. And so, you know, they don't, they don't really line up together.

• 36:16Speaker 00

  And then we need them to line up together in order for us to be able to pass it into a image generation model that was trained with those clip image embeddings. Because, yeah, you want it to match those clip image embeddings, because that's what the diffusion model knows to work with. So we're just trying to make, yeah, the information is there, but we're just trying to put it in a way that the model is used, that the pre-trained frozen models are used to. And I guess that also has to do with the different objectives, the sort of contrastive objective that we're training with. You know, it's easy to get, you can train that, you can get these good representations for the retrieval task, but it doesn't give you representations that match exactly the clip image embeddings

• 36:59Speaker 00

  that we, that we need for passing into the pre-trained models.

• 37:04Nathan Labenz?

  So the final version that actually can be passed into the image generation model, that is the thing that is directly analogous to the image embedding. And I see how that's represented in the figure. So is then maybe the surprising thing or the thing that like needs understanding is like all that seems okay. Now I'm like, well, wait, so in addition to that, there's like a fact that if you want to maximize your retrieval performance, you probably could just use those. If you had trained the whole thing end to end, you could use the the genuine like aligned image embeddings to power retrieval, but it doesn't work as well.

• 37:44Nathan Labenz?

  And what's revealed is there is actually a better way to represent these these voxels for retrieval purposes that is kind of its own thing. It's like a, it's kind of a different subspace in this space that somehow it kind of just gradient descent its way into, and we really don't even know what that is. That's like not really interpretable as of now.

• 38:09Speaker 00

  Well, I think this is the same question for clip in general, because this is a similar problem with clip, image embeddings and text embeddings. They are not lined up in the same way. So I think this may be just kind of a general question in terms of, yeah, what what's going on in terms of the clip, image embeddings and text embeddings. Uh, yeah, it but it seems like when you train these sorts of models with contrast, this sort of contrastive objective, that doesn't really incentivize them to actually map to the same exact region, the same exact space. They map to kind of different spaces that have the appropriate, I guess, cosine similarities and appropriate relationships, because the contrastive learning is just kind of just, you know, trying to maximize similarity

• 38:53Speaker 00

  to similar things and, you know, minimize similarity to to things that are not similar. But I guess that doesn't, but that doesn't necessarily mean that they have to be in the exact same space, I guess. So I think there's there's sort of that question even in case of clip, image and text embeddings as well. Uh, and again, that's why in the original DALL-E work, they did something similar where they took the the clip, text embedding, and then they had to incorporate this additional. Diffusion prior to convert it into a clip image embedding. They couldn't just pass in their text embedding into a diffusion model that took image embeddings. So it's a similar sort of problem. And I think it's just kind of a question for the contrastive learning community, I guess, or something like this.

• 39:37Speaker 00

  Yeah. I think it also has to do with the sort of like, when you work in these high dimensional spaces, this is what, this is something that starts to, to, to show up. So I think that's the, the other thing. It's like, you know, we're working with this very high dimensionality data in terms of the clip, clip embeddings. So that's something that there's some, something about, about working with this high dimension data that, that leads to these sorts of properties.

• 40:04Nathan Labenz?

  If you had done the full end to end training, that would presumably be fine for the reconstruction.

• 40:11Speaker 00

  Yeah, I mean, the goal was obviously to get best reconstructions. We, we tried different things out. I think we kind of stumbled upon, upon this sort of retrieval task. I think, yeah, because the issue is that if you want to go from, again, the problem is like, if you're going from the FMI signal to the clip image embedding, I guess it has to depend. Yeah, it depends on the source of losses that you use. Cause like, if, because we trained it with a contrastive objective, like theoretically it should work, but maybe more practically it might be difficult to just directly go to the, the clip image embedding. I think that like, because you can't, so you can't train it necessarily maybe with a contrastive objective, because that is meant, that, that's going to, you know, cause it not to be, it won't necessarily be aligned that way.

• 40:54Speaker 00

  But like theoretically, that, that should be a perfectly fine pipeline. But I think practically there, there can be some difficulties getting such a model to train well and to, to get, get an accurate model at the end. So that's, that would be the caveat I would mention.

• 41:09Nathan Labenz?

  Cause one thing that's really interesting about both of these papers that you put out in the last short period is again, the low data, but also kind of the non like transformer architecture, right? I mean, you've got something here that jumped out to me like, if I understand correctly, there's no attention mechanism in this.

• 41:29Nathan Labenz?

  Architecture, at least in the models you've created.

• 41:32Speaker 00

  Yes, the model that we create is basically the MLP. And so, yeah, there isn't really. Yeah, it's just the MLP. And then, of course, the clip will have, of course, attention in there. And of course, the diffusion models could have as well, but the MLP not really, yeah.

• 41:48Nathan Labenz?

  How big of a role then does the limited data set size play? Because you have basically just under 1,000 images per individual. And you end up training models for each individual. So you're really just training on a thousand pairs of this is the image that the person saw. And the this is the 15,000 numbers for the 15,000 voxels as measured for that image.

• 42:19Nathan Labenz?

  That's it, right? Just 1,000 of those pairs?

• 42:21Speaker 00

  The the 90, I don't know if you're talking about like the 982, if that's what you were referring to in terms of under 1,000. That was specifically a test set.

• 42:29Speaker 00

  There are, I think, several thousand training samples that are used during training. And that was with the retrieval. I think that's the value you may be talking about. I'm not 100% sure if that's what you were referring. So that, yeah, that was that was just the retrieval experiments. And so basically, we have so we have multiple subjects, first of all. So that's another thing worth mentioning is that we have multiple subjects. So that's multiple different people that are looking at these images. And the the FMI signal is measured, the FMI signal is different per subject. And so we actually train separate models per subject. So the way we do that is that so we have the subject, we have subject one. And the data for subject one is divided into a training set and a test set.

• 43:14Speaker 00

  And so, you know, there may be a few several thousand samples from subject one. That's used for training. And then, you know, you have a test set again. And so, yeah, that's what we have. And then we then you have to do similar process for subject two and all the remaining subjects. And then another thing to note is that, yeah, so we have to subject one.

• 43:33Speaker 00

  We did all the model development on just subject one. And so like, yeah, any, any sort of experiments were done. They were all done with subject one. And then for the rest of the subjects, we just did the same. Whatever worked best for subject one, we did the same thing for the rest of the subjects. And that's what, and that's what we also reported. And then the other thing, of course, is that when we're looking at, you know, the data sets, you know, the images that the subject looks at during the training, in the training set is different from the images that they look at. That's present in the, in the test set, of course. So just again, to make sure that we are, you know, not overfitting or anything like this. You know, we want to make sure they're two separate sets. But yeah, there are, you know, several thousand samples, you know,

• 44:18Speaker 00

  for, you know, the, the training set. I think the exact information, all the exact data is, or the exact information is in the appendix of the, of the paper. But yeah, I mean, it's still, still not that much, of course.

• 44:31Nathan Labenz?

  Yeah, it's small, right? I mean, that's, we're looking at, you know, five billion images in the, in the Lion dataset. And now we're talking, you know, basically one, one millionth of that. So it's surprising that you're able to get this much generalization out of seemingly such a small set, right? I mean, these images are just all kinds of different random scenes and, you know, subject matter. And, you know, it's a pretty diverse set of imagery that you're drawing from.

• 45:03Speaker 00

  I, I think part, part, part of the reason is because we are leveraging these pre-trained models. So of course we're leveraging the, the representation space of the clip models. And so the, the MLP just needs to learn how to map something that is similar to the clip image embedding space. So, I mean, again, even with the sort of, even before the diffusion prior alignment step, it's still mapping to some space that should have kind of similar properties of the, of the, of the clip image embedding space. And in terms of how the images are laid out in that space, it should also be similar. That's, you know, it's, it's, it's still have some sort of similar properties of the clip image embedding space. So I think that really helps.

• 45:45Speaker 00

  It's just like you're kind of already mapping to something that is a very well-established representation. So you have that sort of well-established representation that you're mapping to. And, but again, also like, We don't know how well it's going to generalize out of natural scenes, you know, whatever we were training with. And it's very much possible that it doesn't generalize out of those specific kind of images. And yeah, again, it also has to do with the sort of image reconstruction, the sort of image generation models that we work with. They also may have some particular data sets there that they are trained with, and those are the sorts of images they're best able to produce. So yeah, there are some sort

• 46:24Speaker 00

  of limitations in terms of the generalizability. And I guess, yeah, we haven't studied that as much. But the fact that we're able to train with such little data and get really good results is also partly because I think of this really rich representation space that we're working with. Yeah.

• 46:46Nathan Labenz?

  One of my prior Touchdown episodes that I always keep going back to was with the authors of BLIP out of Salesforce Research in Singapore. And I use some of their models in my own product development work. And that was kind of my first trip down the rabbit hole of these kind of, you know, bridging or mapping from one high dimensional space to another. And since then, it's like this just kind of comes up everywhere. It seems that to a first approximation, it seems like every space is kind of bridgeable or mappable to another space if you're clever enough to figure out, you know, where to put the pylons. Does that seem right to you? Like, it seems like all these things are kind of

• 47:31Nathan Labenz?

  learning similar world models.

• 47:34Speaker 00

  I think so. I think there is some, yeah, I think this idea of of mapping from one space to another, you know, one latent space to another is a very powerful idea. And yeah, I think that that, yeah, I think it's always best to try to take advantage of that as much as possible. And To also, yeah, to take advantage of these pre-existing latent spaces, especially in the cases that you may not have enough data. I think this is really a great opportunity. Again, I mean, that's, I mean, this has been true, I think, for a long time. I think more, the more recent thing is the idea of these multimodal spaces. Of course, the general idea of using pre-trained latent spaces isn't new. I mean, that's kind of what transfer learning in a nutshell kind of is,

• 48:17Speaker 00

  using the, the pre-existing latent space of these pre-trained models. But, you know, of course, a lot of them are more focused for specific domains. I think now the real advantage, the real, I guess, innovation these days is to be able to use these multimodal spaces as well and being able to, to map, you know, different things to these, to these multimodal spaces. I think that's, that's kind of a really exciting area that, um, yeah, it's interesting to see lots of developments in that. I think there was like some papers that recently it was like mapping anything to anything. I think it's from Facebook, Meta, I think, that they published a paper like this. So it's again, a very, very powerful idea. And, uh, yeah, I guess there is some sort of like base representation

• 49:02Speaker 00

  of something that can, that can be utilized for various tasks. But yeah, I guess it also depends, like, there are representations that are useful for certain things and that same representation may not be useful for other things. So it's also about like, there are different representations for a particular image, for example, and maybe some representations are better than another, depending on the task. So there's also a lot to explore in the sorts of like, how is those representations trained and, you know, what, what tasks are those useful for? Again, it comes, you know, maybe different sort of clip models, blip models, all these different models that are out there and they're trained with different, uh, different objectives and they lead to different representations. And that's something I think isn't, isn't, isn't fully explored yet.

• 49:46Speaker 00

  And there's a lot of opportunity there to explore, you know, the differences between these different representations and, and how they can be used for different applications. So lots to explore there.

• 49:55Nathan Labenz?

  Yeah, so we're not even fully done with the architecture. We alluded to kind of the last part, um, earlier. But since you mentioned like these, you know, different pre-trained models and the different spaces that they have. You've got another kind of clever dimension to this as well, right? So that is the, and I'd love to maybe understand a little bit, again, trying to develop some intuition. With the clip embeddings, you know, that is originally a joint space of language and text. And so you point this out in the paper that the image embeddings are kind of inherently more about the captionable part of the image, if you will. This is something I've explored in a lot of different ways too.

• 50:40Nathan Labenz?

  But what notably things that are not included in the captionable, you know, qualities of an image would be like, how beautiful does the image look, you know? Is it appealing? You know, those things are typically not represented in captions. They're not really captured in this clip representation all that much. So what you've really kind of communicated is some sort of semantic understanding at that point. But like, this is what the image would contain. You also then have this other angle and then they converge at the end. But then you also have the, the stable diffusion VAE encoding. And I want to develop a little bit more intuition for that as well.

• 51:21Nathan Labenz?

  But that has kind of the other half, which is like, what does this thing look like? Kind of compositionally, like what colors are where? And you have this kind of intermediate state that is like a blurry thing where it's like, I don't necessarily know what that is, but I do kind of know where the color, you know, what colors are where and like, there's some blobs here. Yeah. Well, I mean, what more would you tell us about that side of the pipeline?

• 51:45Speaker 00

  I mean, that's pretty much a pretty good summary of it. I think, yeah, the idea is just to map the FMI signal to some sort of VAE latent representation. And that latent representation is coming, this is coming from the stable diffusion autoencoder. And that is able to contain a lot of this more low level information. And, and we're able to then output some sort of blurry, some sort of blurry image. And that is used as a starting point for the diffusion models that are then. Producing the final image given the fMRI clip embedding, the clip embedding that is produced. So yeah, this is doing this sort of image to image process.

• 52:30Speaker 00

  You know, people have explored this for various AI art applications already where you can kind of start out with some image often like, for example, you can start with a sketch or something like this and then you take that sketch and then you can use like stable diffusion, for example, to just produce this beautiful image based on the sketch. But again, the idea is that, you know, the the output image has a lot of that sort of the colors or the or the the sort of structures, the positions, very similar to that original sketch. And so that's a similar idea is that we have that starting image. That's almost like a sketch of the, you know, the final image. And we are then, of course, taking our diffusion model, which is given the

• 53:13Speaker 00

  semantic information coming from our clip embeddings. And it is producing a nice final result that matches the sort of maybe colors or the sort of spatial position structures that was there in the original very blurry kind of sketch of the image that was produced by our low level pipeline. So I mean, it's again, very blurry. So it just helps a bit. It helps a bit and gives you better low level information, better low level results. But yeah, it's it's it's good to have additional information where you can kind of see. You have to kind of look for it and you can see you'll start to see that's kind of apparent that, okay, maybe that zebra is in the right place or that person is kind of in the right place. And, you know, it's you see a little bit of that kind of matching up a bit when you look compared to the original image.

• 53:57Speaker 00

  And so, yeah, it kind of just provides an extra factor, that extra, yeah, improvement in the in the in the final image, I'd say, yeah. But again, it's not completely necessary. Like if you just want, if you just cared about the semantic information, you don't need to have a low level pipeline. It's not a necessary part of the You can just start from noise. Yes, exactly. You just start, you start your diffusion model, you know, the typical process is it starts from noise and from noise, it does that sort of iterative process to produce the final image. So instead of starting from the low level image, you could just start from noise and you can get an image that matches the semantic content of the original, but may not necessarily have that spatial information or colors or any of those things correct.

• 54:38Speaker 00

  So yeah, that's it's not a completely necessary part, but it just helps provide sometimes it'll just help improve the image a bit.

• 54:45Nathan Labenz?

  You're definitely standing on the shoulders of, you know, a lot of recent giants here where it's like, this work wouldn't even really have been possible in anything approaching this way. Whatever was the most recently released, you know, pre, uh, you're kind of frozen model that you're that you're building on top of. I guess would be stable diffusion. So that's like, you know, six, eight months ago that that thing first, you know.

• 55:06Speaker 00

  We actually, we, yeah, we of course, we stable diffusion, we actually tested a few, uh, different, uh, pre-trained image generation models. We tested, uh, stable diffusion variations because we're taking in clip image embeddings. So we're looking for the ones that other image generation models to take clip image embeddings. We trained, started with like stable diffusion image variations. I think it was Laffiti. I don't know how to pronounce that exactly, but it was like, I want to get in one of these other papers, Laffiti. And then, uh, versatile diffusion. And that's the one we actually went ahead and, and actually used for the final image generation. It gave the best performance was this versatile diffusion model. And this is actually the one of the models that again, takes in like I talked about the, uh, full clip embedding.

• 55:46Speaker 00

  Instead of the sort of global information, but actually the very like the information for each of the tokens, each of the patches. So that full 257 by 768 tensor. Some of the other models only take in the sort of global vector. That's just like 768 vector, but this one takes in the full information. And again, I think that also really contributes to much improved performance when you get that full information that the model is predicting for the whole image and you're passing all that information into your pre-trained diffusion models. And some of the fusion's models don't take that and those ones are not really necessarily doing well, uh, for this, for this task. But when you have a diffusion model that also takes all that full information, uh, then yeah, it can actually, uh, produce really, uh, really, you know, really nice results.

• 56:31Speaker 00

  And I think, yeah, versatile diffusion, uh, is a paper that I think maybe it went a little bit under the radar. But yeah, their models are actually, actually quite, quite good for, for image generation as well. Uh, and so that's what we use and got the best performance. Uh, yeah, we can use any sort of future pre-trained image generation models that take in these sorts of clip image embeddings. So yeah, we're really excited that. You know, if stability AI releases some other stable diffusion models that taking these clip image embeddings, these sorts of like variation models or whatever, that is much better, that should hopefully also give us a better reconstructions as well. But yeah, again, like you said, it's, you know, we're really building on top of these sorts of existing models, but I think it's also worth highlighting that this, these

• 57:15Speaker 00

  model, our approach can continue to work well for whatever future models also may come in the future. So that's also really exciting about our approach.

• 57:24Nathan Labenz?

  Yeah, that's cool. So, okay, let me try to play this all back in kind of one description. And then I got a few other questions, then we can go on to the other paper, if you have time. So the, I guess, you know, kind of working backwards, it's almost like, all right, we, we have this raw data about the brain, and we know on the other end that there are these image generation models that have recently been created, and they can work in various ways. They could take text in, they could, you know, if we could somehow figure out when they take the text in, they have to embed the text, and then we could figure out how to kind of bypass the text step and like project directly into the text

• 58:07Nathan Labenz?

  space. Then there's all these variations where it's like, well, it could take an image in, it could take a, you know, it could take an image in and noise it, and then kind of take it in some guidance direction. And I think most of our audience will have at least played with these various tools, right, where you start with text and make an image, or we had Su-Hail from Playground AI on, actually, our very first episode, and they have a really nice kind of command-based image editing tool now. And then, of course, you've got your, you know, image to image and all these different variations. People played with all these flavors. So you kind of recognize that, like, okay, that's out there as something we can tap into. And I know that all the information I need is

• 58:52Nathan Labenz?

  contained in these 15,000 voxel numbers. So.

• 58:57Nathan Labenz?

  Then you're like, all right, I'm going to take kind of a pincer movement at it. You identify one kind of blurry representation space where it looks like I've basically just put, you know, a Coke bottle in front of the image and I can just make out kind of color blotches and maybe some vague forms and you figure out how to project the voxels onto that space. So that you have kind of a good starting point for what the image qualitatively should look like. And then you separately say, okay, and I also know that I can guide that toward something semantically meaningful if I have the right representation to guide that, you know,

• 59:42Nathan Labenz?

  the diffusion of that blurry image towards something crisp. So then the other arm of the pincer movement is now we'll project those 15,000 voxel numbers onto this semantic representation, which is used to guide that reconstruction process. You put both of those things into an existing model and now you're in business with reconstructions coming out.

• 1:00:06Speaker 00

  Yep, that sounds about right to me.

• 1:00:08Nathan Labenz?

  All right, I love it. Well, I had to earn that one. There's, and there's some really, again, super interesting things. Yeah, one thing I do really appreciate about the paper too, is you have this just kind of raw code architecture that's like, this is how the model is set up. And they're not that big, right? So can you tell us a little bit about just kind of, I guess, the scale for one thing? I'm also always really interested in how long does this take to train? You know, can you do it on, you know, a machine like overnight? Like, what does your kind of cycle look like in terms of actually training these things? So, you know, your iteration cycle obviously follows from that.

• 1:00:45Speaker 00

  Yeah, um, in terms of training.

• 1:00:49Speaker 00

  I don't know exactly how long it took because I was the one training it. Again, the the main authors, who in this case is Dr. Paul Scotty and Amit the Banerjee. They were the the ones who were running a lot of the experiments. So I don't know how long exactly it takes to train. It's probably on the order of of of several hours. It's not not going to be like days and days of training. That's not not the case here. But yeah, I, you know, I would have to confirm with them the exact timing of how long it takes to train. I think in the in terms of how big the model is, it's it's mentioned in the paper somewhere. I think it was on the order of a billion parameters. Part of it is like also you see that if you get if you use larger models, they do tend to get better performance. So, you know, that's also definitely part part of it.

• 1:01:34Speaker 00

  Yeah, I'm just looking at the in the paper. It's like, I think the model that we finally used had 940 million parameters. So, you know, so, oh, yeah, close to billion parameters that we used in the paper. They have different ablations. We have different ablation studies of the different architectures that are used. So, you know, there are different things that we did try in terms of, yeah, model depth, parameter count, all these different aspects that we we've tried and and we found the best model based based on that. Yeah, so I guess maybe, yeah, it's it's not it's not a too large of a model, but it's certainly certainly sizable, I guess, 1 billion, oh, yeah, 940 million parameters. So so it's still a decent size, yes.

• 1:02:14Nathan Labenz?

  GPT two scale in rough terms. So, yeah, trained on a, just quoting the paper, one A100 machine for 240 epochs with a batch size of 32. So, yeah, just one machine, pretty amazing. Do you know how inference, I assume inference would be like quite fast on this, right? Like similarly, it should take sub second type of deal?

• 1:02:44Speaker 00

  Yeah, it's pretty fast, yes.

• 1:02:45Nathan Labenz?

  And how about on the topic of the four individuals? So you train models specific to the individual. Can we say anything about like how similar or different we are as individuals? Based on if I was doing this, maybe you did this, I would be like, all right, I'm going to put subject A's, you know, data into subject B's model. And if I do that, do I get total garbage, you know, noise out? Do I get something that's like? Decent, but not as good as it was on the actual person's model. How different are we under the hood?

• 1:03:19Speaker 00

  It's a good question. Yeah, I don't know if we actually tried those experiments, to be honest. But, or yeah, maybe they, yeah, I'd have to ask them. But yeah, definitely there are some differences in the sort of visual perception that different people will have. So that would lead to different representations. Yeah, I mean, we are interested in seeing if, for example, there is a way to somehow map them to some sort of shared space that can be used for anybody. But yeah, there are these sorts of minor differences in the visual perception of different people. So that's why we're required to train separate models for the time being. But yeah, again, this is an open research question in terms of how we can. It'd be really nice if you could have one model that just works for everyone.

• 1:04:02Speaker 00

  Part of it is also like, we're trying to think of different approaches. You know, we have different projects that we're working on right now to address this. So again, if they're, for example, people who are interested, who, people listening to this who are interested in working in this further, you know, feel free to check us out and we have different projects like, you know, maybe some sort of foundation model for FMI. Let's just train on all kinds of FMI data sets and, you know, then try to use that as something that we can map to. Or, you know, some form of like, maybe, you know, you have a sort of shared latent space for all these different subjects and for any sort of FMI data.

• 1:04:40Speaker 00

  And then maybe there are subjects that may need to do a few more image, yeah, some more data collection to be able to kind of calibrate the model or something like this. So there are different approaches that we are considering to account for that sort of difference in visual perception of the different, of the different people. But again, this is an open research question and there's lots of interesting avenues of research there. And if people are interested in helping out, that'd be, you know, feel free to join us. And yeah, so I think there's a lot, lot to explore in that direction.

• 1:05:17Nathan Labenz?

  Sounds like if I understand you correctly there, your expectation would be kind of like, we should be able to get to a point where.

• 1:05:26Nathan Labenz?

  Relatively fast calibration would be possible, maybe another, you know, if we sort of extend, we said, okay, originally, we had, you know, five billion images in the, in the data sets that trained like the, you know, foundation models, it took a, you know, a few thousand images to train something that could kind of bridge the, you know, voxel space into the embedding space. And it sounds like you think there's probably enough kind of similarity there that maybe you can bring it down another thousand fold and have like a five image, you know, calibration step that could kind of fit your personal, you know, biology, you

• 1:06:08Nathan Labenz?

  know, shape of your own head, whatever, to the, you know, to the kind of shared FMRI latent space. That's basically where you think this is going?

• 1:06:17Speaker 00

  Yeah, I think that's the, the, the hope that we have, that, that, that something like that may be, may be possible. And this is again, something that, you know, the, the, the main author, Dr. Paul Scotty, he's been really kind of spearhead, spearhead this effort in terms of trying to develop these sorts of approaches. And so he's really excited about, about this and, and really thinks that something like this would be possible. You know, there's various questions that need to be answered, like, what are those sorts of five, for example, if it, if it is just five images, what are those five images going to be? Are there, are those different images per, per, per subject? Maybe there are some images that, you know, maybe there, yeah, you may have to find different optimal images. Uh, and then, yeah, how's that going to map appropriately? So there's lots of open

• 1:06:59Speaker 00

  questions, but the general idea is that's kind of where we're hoping we can, we can get to that kind of, kind of stage in terms of, in terms of image reconstruction. And yeah, that'd be pretty exciting, I think, if we could get something like that working.

• 1:07:12Nathan Labenz?

  Yeah, it's amazing. How much do you think this ultimately feeds into better understanding of the brain, which seems like there's, you know, tremendous potential there versus, you know, the other direction would be like? Wearable devices, sort of more kind of well-personed consumer application. What's your kind of expectation for how this research develops in that respect?

• 1:07:38Speaker 00

  In terms of like, yeah, let's start with the first thing of like wearable devices and things like this. Of course, the fMRI is like, you have to go into this MRI machine. It's this very large machine and it's a very involved and time-consuming process. So to be able to use the fMRI for wearable, I think that would require a lot of development on in terms of MRI technology and stuff like this.

• 1:08:00Nathan Labenz?

  Yeah, it's a strong magnet to be carried around, no doubt.

• 1:08:02Speaker 00

  So it's like, that's more of like a hardware problem that, you know, I don't know if something like that would be solved, and if so, when. There are alternative like wearable approaches for measuring brain activity, things like EEG, I think FNIRS, different approaches. But like, again, they have different sorts of trade-offs in terms of the signal that they provide. You know, the sort of spatial resolution, temporal resolution. So like EEG is good in terms of like, it's got a good temporal resolution, but its spatial resolution is less than fMRI. And we already talked about how fMRI's spatial resolution isn't like actually that great, to be honest. Like, we talked about it's like just this sort of millimeter by milli-, you know,

• 1:08:45Speaker 00

  two millimeters by two millimeters by two millimeters voxel. And EEG is even more coarse-grained than that. So it seems very unlikely that you can get a decent signal from EEG, for example. There may be some other technologies that may be able to get some signal, but I'm not entirely sure what other ones you can get signal from. So I think there's a lot more of like, at this point, it's more like there's a need for better hardware that's able to actually get high-quality signal if you want this to work for wearable applications. So I'm not sure if this is necessarily possible right now. And again, like, even then, it's like, okay, maybe if you don't use fMRI, then if you're using some other technology, then you have to.

• 1:09:27Speaker 00

  Validate if a similar approach would work well for these other technologies. I expect like if you have a high quality data set with high resolution and you get the relevant information, the approach, the general approach should work well. But you know, it still needs to be tested and validated. So like for basically to summarize for variables, you need better, you basically need a better hardware to be able to do this. And if you are using a different technology, then you need to be able to validate that for that other technology. So that's why I'm not really like certain that, you know, this is something that could be used for consumer applications, but it's not going to happen within the next couple of years, I'd say. But you know, maybe if there are some interesting hardware developments, there could be something that that may be possible in the future.

• 1:10:10Nathan Labenz?

  Because what I would imagine would be a fundamental challenge of like a consumer hardware scenario would be what kind of resolution can you get? So I wonder if you could start to anticipate what you might need by taking your current approach and being like, what if we just make it four millimeters cubed instead? And we just, you know, average the, you know, that number across, you know, if you take a, an eight, you know, bit, turn, you know, go from eight voxels to one by, you know, doubling the edge length and then just take the average of those numbers. Now you just have, you know, instead of 15,000, you have whatever about 2,000, a little less than 2,000, right? Numbers?

• 1:10:55Nathan Labenz?

  Have you tried anything like that? Do you think that you could just kind of be like, what if we just fuzz our data and like see how far we can fuzz it before we can't read it anymore?

• 1:11:04Speaker 00

  Yeah, we haven't actually tried anything like that. But one thing that I think we may investigate in the future is just also just some of the other data sets in terms of, yeah, I mean, there are some of these other data sets that use less powerful magnets. I mean, that's what's been used in the field so far until maybe a couple of years ago, when this new data set got released. So, um, you know, it could be worth just also trying those other data sets as well and seeing what kind of performance people get with those data sets.

• 1:11:33Speaker 00

  Yeah. And then again, your idea sounds like an interesting idea too. So there's definitely lots of different.

• 1:11:39Speaker 00

  Approaches to to to test out that, yeah, I mean, yeah, it's definitely an important question to to to think about how much data or how much of the actual signal is needed and the sort of signal to noise ratio needed. But yeah, I think even then, like something like EEG or something like this is still like significantly coarser than than than um, even some of these less powerful FMI machines. Um, yeah. And then I guess the other question was about research sort of applications as well, right? We we mentioned this also a bit in the paper and, you know, there there are some potential and potentially interesting applications. So for example, you could imagine just trying to study the

• 1:12:21Speaker 00

  image reconstructions of of different, you know, patients with different neurological diseases. You could see like how how those sorts of reconstructions change over time as degrees progresses. And one or you could use them potentially for some sort of diagnostic applications as well. So one example that, you know, we were thinking of is like, if you had someone with with depression, maybe their reconstructions, you know, look a little bit different in terms of like, yeah, maybe a little bit more dull, maybe a little bit, yeah, there may be some differences in the reconstructions that we'd be able to pick up on. There are various studies that look at kind of these sorts of responses to images

• 1:13:02Speaker 00

  to or even just like mental imagery, things like this. And a lot of those have been very coarse grained, just be like, imagine some sort of object or like, it's very like, maybe imagine an animal or something. It's very coarse grained. And so being able to also have this more fine grained information, that there could be a lot of really useful neurological studies that could be done if you're able to get that fine grained information from from from a from a subject in terms of what they're imagining, what they're looking at, and being able to reconstruct that. But yeah, I think there's a lot of interesting things in terms of like, yeah, degrees, disease progression and diagnostics. And of course, you know, you have this more kind of basic applications of like, yeah, if you have locked in patients who are in coma or unable to communicate with the outside world,

• 1:13:47Speaker 00

  maybe some of these could be useful. Again, maybe with further development of variable technology and things like this, can use it for interesting medical applications. So yeah, I think there are lots of interesting. Medical applications and research applications out there. And yeah, I think we're starting to reach the point where maybe these sorts of image reconstructions can be used for interest to to to research interesting clinical problems and neurological problems. So yeah, it's I think it's quite exciting. So even if it's not necessarily being used for, you know, the the, you know, mind reading for consumers, I think there are still lots of interesting applications that we'll see in the short term for in clinical research.

• 1:14:30Nathan Labenz?

  Yeah, I think one of my big takeaways from this paper is just how many doors are opening, you know, the so many different doors kind of recently opened that enabled you to do this. And I think, you know, you're obviously going to be showing the way to others. And it seems like as much as we have seen a lot of awesome stuff in the last couple of years, something tells me, you know, we're not anywhere near, you know, kind of the end of the, you know, the fruitful exploration of the current paradigm. And, you know, the fact that you're able to do this kind of work with a single A100, you know, is pretty telling, you know, in terms of how much value is already kind of,

• 1:15:14Nathan Labenz?

  you know, embodied in these foundation models and just waiting for, you know, a super, you know, clever person to come along and figure out how to, you know, piece together the right architecture to kind of bridge different spaces and, you know, make all these different things talk to each other. It's going to be wild, I think.

• 1:15:32Speaker 00

  One of the interesting things about this project is kind of, you know, how it was done, how was how it was conducted in terms of the sort of open research environment and organization that that we had. So this was again done as part of of Medarc, which is the sort of research organization that I've founded. The projects that we work on are done in this sort of collaborative and decentralized manner and done in an open source manner. So of course, the GitHub repository was always open source for people to look at and contribute and and, you know, there would be weekly meetings on Discord and, you know, chat chatting on Discord, sharing research idea, sharing progress. So all that was happening.

• 1:16:14Speaker 00

  And then the compute was provided by by Stability AI and you know, they were able to support our research. And yeah, I think this sort of approach was really great because we were able and a lot of interesting, smart and clever people were able to to contribute to this project. So really was grateful for the contributions, of course, of Dr. Paul Scotty, who's the the lead author and also Ahmed Deep Banerjee, who's also really came up with, you know, really was working on this project a lot and, you know, came up with some of these really interesting ideas that really pushed this forward and and really was able to to help get this working. And then just so many other contributors from around the world that was able to to work on this. I think this is an interesting

• 1:16:56Speaker 00

  project that kind of demonstrates the the value of of this sort of open collaboration as well. And, you know, with this sort of open collaboration, we can do all kinds of incredible things as well. I think, yeah, like you said, there's a lot of other interesting aspects in terms of being able to use foundation models that enables new opportunities. And then, of course, this sort of collaboration that is happening. So I think, you know, this is something that, you know, these sorts of things wouldn't have been possible just just a few years ago. And so it's quite, quite incredible what's what's possible now. And, you know, it can be done by people, you know, that are sitting somewhere halfway across the world and, you know, working from their laptop. And it's it's kind of incredible what's what's what's possible these days. So, yeah.