Introduction
The hippocampus has long been characterized as a "Swiss Army knife" of the mammalian brain, playing a critical role in both long-term associative memory and spatial navigation. Historically, these two functions have been modeled using distinct algorithmic frameworks. The memory literature often employs models like Hopfield Networks or vector-based similarity lookups, viewing retrieval as a one-step associative process. Conversely, the navigation literature utilizes Reinforcement Learning (RL) constructs, specifically the Successor Representation (SR), to explain how the brain builds predictive maps for planning and trajectory evaluation.
This theoretical note bridges this gap by demonstrating that Personalized PageRank (PPR)—a staple of graph theory and web indexing—is mathematically isomorphic to the Successor Representation. By proving this equivalence, we propose a unified computational role for the hippocampus: it serves as a generalized graph processor that computes the stationary distribution of random walks across arbitrary input structures, whether they are physical maps or semantic networks.
Algorithmic Alignment
Using Marr’s levels of analysis, we can compare the fixed-point iteration of PPR with the analytical solution of the SR. The alignment of these two frameworks reveals a deep structural identity between memory retrieval and spatial planning.
1. Personalized PageRank (PPR)
PPR ranks nodes in a graph based on the probability of a random walker visiting them. Crucially, it includes a "teleportation" or restart factor, which periodically resets the walker to a specific starting node (the "query"). The stationary distribution $\pi^*$ is defined by: $$\pi^* = (1 - \alpha) (I - \alpha P)^{-1} p^*$$ Where:
- $\alpha$ is the teleportation coefficient.
- $P$ is the transition matrix of the graph.
- $p^*$ is the personalization or restart vector (the query).
2. Successor Representation (SR)
In RL, the SR summarizes the expected future occupancy of states under a given policy. It allows an agent to compute the value $V$ of a state by multiplying the SR matrix by a reward vector $r$: $$V = (I - \gamma T)^{-1} r$$ Where:
- $\gamma$ is the discount factor.
- $T$ is the policy transition matrix.
- $r$ is the reward function.
The Formal Proof of Equivalence
By identifying the variables across domains, the two equations become identical. Specifically, if we set the discount factor $\gamma$ equal to the teleportation coefficient $\alpha$, the transition matrix $T$ equal to $P$, and the reward function $r$ equal to the normalized query $(1 - \alpha) p^*$, the value function of the SR is exactly the stationary distribution of PPR.
Key Insights
Memory as Multi-Step Search
Traditional memory models often treat retrieval as a simple similarity lookup (akin to modern Retrieval-Augmented Generation or RAG). However, complex reasoning requires "walking" a graph of associations. This paper argues that advanced memory retrieval is essentially a multi-step search process, where the hippocampus navigates semantic space in the same way an agent navigates physical space.
The Reward-Query Equivalence
A striking implication of this isomorphism is that a "user query" in a memory system is mathematically equivalent to a "reward" in a navigation system. Both serve as a "pull" or a bias that directs the random walk toward specific nodes of interest. This suggests that the hippocampal circuitry does not distinguish between seeking a physical goal and seeking a specific piece of information.
The Hippocampus as a Graph Processor
We hypothesize that the hippocampus is agnostic to the nature of the input. Whether the entorhinal cortex provides spatial coordinates or abstract conceptual relationships, the hippocampus performs the same fundamental operation: computing the stationary distribution of the resulting graph. This explains why the same neural architecture is so effective at two seemingly disparate tasks.
Limitations and Future Directions
While the mathematical equivalence is robust, several questions remain regarding biological implementation. The current analysis does not provide a low-level neural circuit diagram explaining how CA1 or CA3 neurons perform matrix inversions or temporal-difference (TD) updates. Furthermore, it remains to be seen if the brain truly performs multi-step PageRank-style retrieval or relies on simpler approximations.
Future empirical work should focus on identifying "place cell" and "grid cell" equivalents in non-spatial, semantic tasks. If this theory holds, we should observe firing patterns that reflect the topological structure of abstract graphs. Additionally, integrating this framework with Bayesian "Active Inference" models could provide a more probabilistic understanding of how these representations are formed and updated.
Summary updated