Benchmarking Out-of-Distribution Generalization Capabilities of DNN-based Encoding Models for the Ventral Visual Cortex [video]
- All Captioned Videos
- Publication Releases
[RHYTHMIC MUSIC]
SPANDAN MADAN: I'm Spandan Madan. I'm a PhD student in computer science at Harvard University, working with Professor Hans Peter Fischer and Gabriel Kreiman, and I'm also a research affiliate at the Boston Children's Hospital. So my research focuses on machine learning and artificial intelligence. Specifically, I look into why machines can't generalize out of the training data distribution.
What that means is that humans very easily, effortlessly adapt to new unseen situations they haven't seen before, but machines often struggle to. And that's what I look into. So one of the great things about Harvard and Boston Children's Hospital is that they foster a lot of interdisciplinary research. And my co-author, Will Xiao, is someone who has worked a lot with brains. I have worked a lot with artificial intelligence. And we've started talking about how the behavior of these models and how the behavior of brains might be similar or dissimilar. And that's how we started working on this project.
So one of the big challenges in neuroscience is collecting a large-scale data set. In AI, it's easy. You just go to the internet, you download a data set, and you're good to go. But in neuroscience, you have to collect data from brains, which can be pretty challenging. And that's where my co-authors came in. So my co-author Will Xiao and Professor Marge Livingstone collected this really large-scale, amazing data set, which allowed us to do all the analysis that we did.
So most modern models of the brain rely on deep neural networks. However, from AI research, we know that these models don't really work very well with respect to out-of-distribution data. That is to say, they work very well on the data they've been trained with, but they don't generalize to data that's very different from it. So what we did is we took a bunch of brain data, and we split it into different chunks.
So for the first set, what we did is we took low contrast images at the training set and the high contrast images as the test set. We trained on the low-contrast images, and then retested on the out-of-distribution high-contrast images. And what we found is that these models of the brain do not generalize. The same was true for high hue, contrast, temperature, intensity, and a bunch of different image computable metrics that we found.
Well, one of the big challenges in neuroscience is to understand how the brain processes images and how it makes sense of the world that we see. So think about it like this. If we have a model of the brain, it should work no matter what image you test it on. The same is saying, if you had Newton's laws, it should work on every single object, from an apple to the planets. And it would not be very satisfying if Newton's laws worked only for planets, but not for the apple that fell on Newton's head. Very similarly, a model of the brain is not a satisfying model of the brain if it only works on the training data and not on the testing data.
So the large-scale data set I just talked about, which is like an IT branch, it's one of the largest data sets of neuronal recordings from the IT cortex. This data set is freely available to download. You can find more in the paper. And we've already made it available for researchers to easily download it and use it by splitting the data into in-distribution and out-of-distribution data sets, which people can use to study generalization.
The problem of generalization has been well known in AI for a very long while, and a lot of people, including me, have been working on it. Going forward, as AI and neuroscience become increasingly intertwined, we hope that this problem also becomes of importance to neuroscience researchers. And with the data sets and the tools that we provide in this paper, we hope that we can bring the two fields together and work on this problem together.
Authors: Spandan Madan, Will Xiao, Mingran Cao, Hanspeter Pfister, Margaret Livingstone, Gabriel Kreiman
Link to Paper: https://arxiv.org/abs/2406.16935
Abstract: We characterized the generalization capabilities of DNN-based encoding models when predicting neuronal responses from the visual cortex. We collected \textit{MacaqueITBench}, a large-scale dataset of neural population responses from the macaque inferior temporal (IT) cortex to over 300,000 images, comprising 8,233 unique natural images presented to seven monkeys over 109 sessions. Using \textit{MacaqueITBench}, we investigated the impact of distribution shifts on models predicting neural activity by dividing the images into Out-Of-Distribution (OOD) train and test splits. The OOD splits included several different image-computable types including image contrast, hue, intensity, temperature, and saturation. Compared to the performance on in-distribution test images -- the conventional way these models have been evaluated -- models performed worse at predicting neuronal responses to out-of-distribution images, retaining as little as 20% of the performance on in-distribution test images. The generalization performance under OOD shifts can be well accounted by a simple image similarity metric -- the cosine distance between image representations extracted from a pre-trained object recognition model is a strong predictor of neural predictivity under different distribution shifts. The dataset of images, neuronal firing rate recordings, and computational benchmarks are hosted publicly at: https://bit.ly/3zeutVd