CALLIOPE · BCOM

This note summarizes the basic idea of diffusion models and uses it to formulate a speculative hypothesis about brain-like planning. A diffusion model deliberately corrupts real data with noise, then learns to reverse that corruption. This transforms generation into a sequence of small denoising operations rather than a single decoding step. The same principle may be useful for brains: endogenous noise can provide candidate variation, while learned denoising dynamics constrain that variation into plausible percepts, futures, or model-unpackings. In this view, a planner can exploit noise to sample possible futures, denoise them into coherent trajectories under the current world model, and then evaluate them according to the agent's objective function. We connect this proposal to a recent neuroanatomical hypothesis that casts the hippocampus as a latent diffusion engine, and to our own Kolmogorov-Theory treatment of computational neuropsychiatry, in which the hippocampus is identified as a natural host for the agent's latent space.

Noise is not the enemy of thought — it might be the engine of imagination.

Diffusion models, the technology behind modern AI image generators, work by learning to undo damage. You take a real image, corrupt it with noise step by step until it's unrecognizable, then train a neural network to reverse each small corruption step. At generation time, you start from pure noise and run the learned cleanup process in reverse. The result: coherent images emerge from randomness, not in one leap but through many small refinements. Each step is a tractable local problem rather than one impossible global one.

This note asks whether the brain might exploit the same principle. Not literally — the brain isn't running a U-Net with a Gaussian noise schedule. But the computational logic maps surprisingly well. The brain is constantly inferring stable causes from noisy, partial, ambiguous sensory input. That's already denoising. Recurrent cortical processing looks more like iterative refinement than one-shot decoding. The proposal here is that endogenous neural noise — the brain's own internal fluctuations — could serve as the variation source, while learned recurrent dynamics act as the cleanup process, carving plausible percepts, memories, and imagined futures out of that noise.

The most interesting application is planning. To plan, you need candidate futures — not one deterministic prediction but a diverse set of plausible continuations. A diffusion-like mechanism gives you exactly that: inject noise into the current world-model state, denoise it into coherent candidate trajectories, then score those candidates against your goals. Noise becomes the branching substrate from which possible futures are carved. The paper connects this to BCOM's own Kolmogorov Theory framework, where a Modeling Engine, Objective Function, and Planning Engine collaborate in an agent loop — diffusion-like dynamics would slot in as the candidate-generation step.

The hippocampus gets a starring role. Jones and Breakspear (2026) independently proposed that the hippocampus operates as a "latent diffusion engine," compressing sensory and cognitive inputs into a low-dimensional latent space and regenerating percepts and episodic memories via stochastic oscillatory dynamics. This converges with BCOM's earlier identification of the hippocampus as the natural host for the agent's latent space. The note argues these aren't coincidentally similar ideas — hippocampal-cortical oscillations could supply the time-indexed denoising schedule, and hippocampal pathology in dementia could be reread as degradation of this generative machinery.

The paper is honest about what this isn't. Modern diffusion training requires backpropagation and engineered noise schedules — neither is biologically realistic. The brain probably doesn't generate full scenes from pure noise; it starts from partial states, goals, and memories. And denoising alone produces plausible candidates, not good decisions — valuation and selection still need to do their work. The contribution here is conceptual: noise as a resource, iterative denoising as a mechanism, and the hippocampus as a plausible neural locus for all of it.