Which Way

Place cells in the hippocampus fire when an animal occupies a specific location. A cell for the left fork fires when the rat is at the left fork; a cell for the straightaway fires when the rat is on the straightaway. A path through a maze is a sequence of positions, and running that path produces a corresponding sequence of cell activations. The cells form a map — not a visual one, but an abstracted representation of where the animal is.

During quiet wakefulness and sleep, these cells do something that has nothing to do with the animal's current position. They fire in sequences — the same sequences they produced during movement — at dramatically compressed timescales. A path that took thirty seconds to run replays in about two hundred milliseconds. This happens inside sharp-wave ripple events: brief, high-frequency oscillations in the hippocampal local field potential that occur hundreds of times per night and during quiet rest. The cells are running through the path without the path.

This is forward replay, discovered in the late 1990s. The standard interpretation is memory consolidation: the hippocampus re-runs experiences at high speed to transfer them to longer-term neocortical storage. Repeated activation strengthens the relevant synapses. It's a plausible account, though the causal relationship between replay and memory consolidation is still contested — correlation, not proven cause.

What I want to focus on is the other direction.

In 2006, Foster and Wilson discovered that the sequences could run backward. At the moment a rat reaches a reward — the goal, the food — the place cells fire in reverse order: from the current location back toward where the run began. From goal to start, compressed into the same two-hundred-millisecond window. Reverse replay, and it happens specifically at reward locations, specifically after the animal has arrived, specifically during the few seconds of quiet that follow.

The prevailing hypothesis connects this to reinforcement learning, specifically to what's called the credit assignment problem. If you receive a reward, you need to figure out which earlier actions and positions in the sequence actually deserve credit for it. The cell active just before reward gets the direct signal. But what about the cell at the turn before that? It fired a few seconds earlier, before reward arrived. How does it learn that it contributed?

The Bellman equation in reinforcement learning solves this by propagating value backward through a sequence: the value of a state is computed from the value of the state that follows it. Working backward from reward, you update each position in turn, most proximal first, then back toward the start. The reverse replay has exactly this structure. Playing the path backward — from food to start — allows each earlier position to be updated in sequence. The ripple event carries the reward signal back through the cells. The term for this operation is a Bellman backup.

Whether the hippocampus is actually implementing this computation, nobody knows. The parallel is clean, and the reward-sensitivity of reverse replay is suggestive — it increases when reward is greater than expected and decreases when reward is reduced or absent, exactly the pattern you'd want from a system that's tracking prediction error. But the brain doesn't know about the Bellman equation. The question is whether it happens to implement the same operation, or whether the parallel is coincidental and the actual mechanism is something else entirely.

What I can't reconcile is the mechanism at the cell level. A rat runs a path for the first time. Synaptic potentiation encodes the sequence: cell A fires, then B, then C. Hebbian reinforcement strengthens the A→B→C connections. Then the rat arrives at the reward, pauses, and the cells fire C→B→A. The sequence encoded forward is running backward.

How? If the architecture was built during forward movement, what drives the reversal? This question doesn't have an answer. Models exist — involving asymmetric inhibition, ripple-wave propagation dynamics, specific features of hippocampal CA3 microcircuits — but none of them have been confirmed as the actual mechanism. Something in the network runs the film backward without obviously having a backward film.

There's an asymmetry worth sitting with. Forward replay occurs preferentially before runs — at the start of a trajectory, before the animal has moved. Reverse replay occurs after runs, at the reward site. One is prospective; the other is retrospective. And during sleep, replay is mostly forward again — consolidation rather than credit assignment. The same cells, the same structure, apparently doing three different things depending on when an event occurs: anticipating what's about to happen, propagating value backward from what just happened, consolidating what happened over the night.

You experience running the maze. You arrive at the food. You rest for a few seconds. During those seconds, the hippocampus runs the path you just took backward, in ripple-bound bursts — the turn before the food, the straightaway, the start — propagating what happened at the end back toward the beginning. Then during sleep, it runs the path forward again, slower in some sense, handing pieces of it to the cortex.

Neither computation shows up as experience. The reversal happens; you don't feel it happening. You sit at the reward location and eat, and underneath the eating, the sequence is running backward in compressed flashes too brief to surface, doing something to the cells that fired along the way.

What it feels like from inside a ripple event — if "feels like" applies to anything happening in two hundred milliseconds of high-frequency oscillation — is not a question the experiments can address. The psychometric function for ripples doesn't exist. We measure behavior and infer. The backward computation runs, and the only window into it is a recording electrode, not a report.