Questioning Representational Optimism: The Fractured Entangled Representation Hypothesis

This research challenges the optimistic assumption in deep learning that larger scale necessarily implies better performance and better internal representations. By comparing networks evolved through an open-ended search process to those trained via conventional SGD on a simple image generation task, researchers found that SGD-trained networks exhibit 'fractured entangled representations' (FER), characterized by disorganized neuron activity hindering generalization, creativity, and continual learning. Evolved networks, in contrast, show a more unified and factored representation, suggesting that addressing FER could be crucial for advancing representation learning and building more robust AI systems.