gradCSCG: Differentiable cognitive maps learned end-to-end from raw images

A new paper from Simula Research Lab and international collaborators presents gradCSCG, a fully differentiable reformulation of the Clone-Structured Causal Graph (CSCG) algorithm. Originally a hippocampus-inspired model for learning interpretable maps from aliased observations, CSCG required a predefined discrete alphabet and used expectation-maximization, making it incompatible with modern deep learning pipelines. gradCSCG removes this barrier by enabling end-to-end gradient-based training directly from raw image sequences.

Coupled with a learned vector-quantized variational autoencoder (VQ-VAE), gradCSCG uses a soft emission forward pass to propagate the map-learning objective back into perception. Loss-balancing mechanisms prevent module collapse during joint training. Experiments show gradCSCG replicates CSCG's room topology recovery in symbolic grid worlds, and—more importantly—recovers underlying adjacency graphs from heavily aliased MNIST image sequences with high edge precision and recall. This work demonstrates that structured causal graph learning can now be a composable building block in deep architectures, opening the door to neuroscience-inspired world models that learn directly from pixels.