Predictive Coding Talk (2025)

Brain oscillations do the math of prediction error — and this paper shows exactly how.

The brain is constantly making predictions about what it will sense next, then comparing those predictions against reality. The difference — the prediction error — is what drives learning and perception. Everyone agrees this happens. Nobody has nailed down the circuit-level mechanism. This paper proposes a specific answer: the comparison is done by cross-frequency coupling, the phenomenon where slow brain waves (alpha, ~10 Hz) modulate the amplitude of fast ones (gamma, ~40 Hz).

The key intuition borrows from AM radio. In AM transmission, a low-frequency audio signal is encoded in the envelope — the slowly varying amplitude — of a high-frequency carrier wave. The paper argues the brain does something analogous: sensory inputs ride as envelopes on fast gamma oscillations, while top-down predictions arrive as slow alpha rhythms. Two coupling modes then implement the comparator. Signal-Envelope Coupling (SEC) lets the slow prediction signal directly suppress the gamma envelope when it matches the input — "explaining away" the sensory signal. Envelope-Envelope Coupling (EEC) operates one level up, letting slow envelope fluctuations gate how much weight (precision) the error signal carries. SEC gives you fast error detection; EEC gives you confidence-weighted error gating.

The simulations use the LaNMM, a biophysical model of a cortical column with distinct superficial and deep layers, built to reproduce the laminar structure where feedforward and feedback signals actually travel. When a matching prediction is injected, gamma-band error output collapses. Mismatched or absent predictions leave it elevated. This is the "explaining away" principle made concrete in a circuit.

The paper then stress-tests the framework with two clinical perturbations. Simulating serotonergic psychedelics (by boosting excitatory drive onto deep pyramidal cells) reduces the effective weight of predictions and inflates error signals — consistent with the REBUS model, which frames psychedelic experience as a collapse of top-down prior beliefs. Simulating Alzheimer's disease by progressively weakening parvalbumin interneurons produces a two-stage trajectory: early dysfunction amplifies error signals (too much surprise), while severe late-stage dysfunction causes gamma to collapse entirely and error propagation to fail.

This is a slide deck from a May 2025 conference presentation, so the source is sparser than a full paper — figures are referenced but not reproduced here, and some methodological details are gestured at rather than fully specified. The core mechanistic claim is clear and internally consistent, but readers wanting quantitative depth should track down the companion bioRxiv preprints cited in the references.