New Theory Explains Slow Thinking and Active Perception in AI

A new theoretical framework aims to provide a mathematical foundation for cognitive functions like thinking and perception, specifically focusing on "slow thinking" and active perception in AI. This theory starts by considering how probability distributions on observable and latent spaces can be lifted and projected to represent complex data using simpler function families, such as neural networks. The core concept, "active lifting," is based on sampling latent sequences and an inherent drive to reduce uncertainty at the maximum possible rate. This framework defines a broad design space that includes slow-thinking models as a specific subspace. It also outlines how these models can be improved by ascending representation and sampler hierarchies. Furthermore, the theory derives an inference process with an internal time axis and a training objective that resembles minimum-length coding, suggesting how languages might emerge. Practical implications include a three-stage pathway for enhancing slow-thinking models, a unified approach for building encoders and generative models across data modalities, and a potential explanation for human-like visual representations and the resolution of policy collapse.