Differentiable Causal Graphs Learn Cognitive Maps from Images

Agents typically struggle to build structured world maps from raw, non-repeating sensory input and movements, especially when exact patterns are rare. The original Clone-Structured Causal Graph (CSCG) algorithm offered a solution for learning interpretable maps but required predefined discrete inputs and wasn't easily integrated with neural networks for end-to-end processing of raw data like images.Researchers have now addressed this by reformulating CSCG into a fully differentiable module called gradCSCG. This new module is coupled with a learned vector-quantized variational autoencoder (VQ-VAE) for perceptual input. This allows the map-learning objective to flow back into the perception module, enabling joint training and robust map recovery from visual input.The gradCSCG pipeline successfully reproduces original CSCG results on symbolic grid worlds and demonstrates robust map recovery on MNIST image sequences, even in heavily aliased environments. This work provides a foundational proof that CSCG can serve as a composable building block within deep learning architectures for cognitive map learning.