Questioning Representational Optimism in Deep Learning

Summary

The Fractured Entangled Representation Hypothesis 📝Paper PDF Abstract Much of the excitement in modern AI is driven by the observation that scaling up existing systems leads to better performance. But does better performance necessarily imply better internal representations? While the representational optimist assumes it must, this position paper challenges that view. We compare neural networks evolved through an open-ended search process to networks trained via conventional stochastic gradient descent (SGD) on the simple task of generating a single image. This minimal setup offers a unique advantage: each hidden neuron's full functional behavior can be easily visualized as an image, thus revealing how the network's output behavior is internally constructed neuron by neuron. The result is striking: while both networks produce the same output behavior, their internal representations differ dramatically. The SGD-trained networks exhibit a form of disorganization that we term fractured entangled representation (FER). Interestingly, the evolved networks largely lack FER, even approaching a unified factored representation (UFR). In large models, FER may be degrading core model capacities like generalization, creativity, and (continual) learning. Therefore, understanding and mitigating FER could be critical to the future of representation learning. More Data and Visualizations Here is the all of the supplementary data from the paper. Intermediate Feature Maps: All Weight Sweeps: Select Weight Sweeps From the Paper: Other Assets All other important assets from the paper can be found in ./assets/. Code This repo contains code to: Load the picbreeder genomes from the paper Layerize it into a MLP format Train a SGD network to mimic that output Visualize the internal representation Do weight sweeps and visualize the result Google Colab For a quick start, open src/fer.ipynb in Google Colab: Running Locally To run this project locally, you can start by cloning this repo. git clo...