Neural Networks Are Blurry, Symbolic Systems Are Fragmented. Sparse Autoencoders Help Us Combine Them.

Neural networks and symbolic systems compress information in opposite ways, with neural models creating smooth, blurry representations and symbolic systems using rigid, high-definition rules. While combining their strengths is desirable, doing so at the scale of modern foundation models has been a major challenge. Sparse autoencoders (SAEs) offer a compelling solution by learning a dictionary of discrete concepts directly from a model's internal activations. These SAEs can then act as a scalable bridge, mapping the model's learned concepts to various external symbolic systems like knowledge graphs to create a powerful neuro-symbolic architecture.