Acetylcholine: a candidate substrate for hippocampal predictive learning?

Abstract

Acetylcholine release in the hippocampus has been associated with diverse neural functions in learning and memory, including novelty, uncertainty detection, error correction, arousal and hidden state inference, while also modulating theta oscillations. Confronted with this plurality of roles, a unifying framework for interpreting cholinergic function is lacking. Recently, predictive models have emerged as a normative lens for understanding neural function, viewing the brain as learning to predict environmental dynamics. Within this framework, we propose that hippocampal acetylcholine serves as a fundamental learning signal, encoding state transition prediction errors — the magnitude of mismatches between predicted and actual transitions in environmental states. We suggest that, analogously to how dopamine signals reward prediction errors to guide value learning, acetylcholine guides structural learning, signalling state transition prediction errors. By condensing behaviour onto timescales amenable to spike-timing-dependent plasticity, theta sequences provide a potential substrate for calculating state transition prediction errors and driving the synaptic updates that revise the internal world model. Within the septo-hippocampal circuit, we propose that theta sequences relay predictions to the septum for comparison against observed state transitions, with the resulting cholinergic feedback gating plasticity in proportion to the mismatch. Plausibly, this mechanistic account unifies the diverse roles of acetylcholine, from novelty detection to hidden state inference, as aspects of a single computational principle — learning predictive internal world models from precisely-timed neural sequences. Predictive models have emerged as a normative lens for understanding neural function. In this Perspective, de Cothi, Shipley and Barry plausibly unify the diverse roles of hippocampal acetylcholine in learning and memory by proposing it serves as a fundamental learning signal, encoding the mismatch between predicted and actual environmental state transitions.