[Paper] Leakage and Second-Order Dynamics Improve Hippocampal RNN Replay

Source: arXiv - 2602.18401v1

Overview

Recent work shows that the hippocampus can “replay” sequences of neural activity even when no external stimulus is present—a process thought to support memory consolidation and planning. This paper revisits noisy recurrent neural networks (RNNs) that have been used to model such replay, treating the dynamics as a form of stochastic sampling. By digging into the underlying mathematics, the authors uncover three mechanisms—state leakage, adaptation (negative feedback), and momentum—that together make replay faster, more exploratory, and more biologically plausible.

Key Contributions

• Theoretical proof of time‑varying gradients: Demonstrates that the optimal replay gradients change over time, making naïve Langevin sampling inefficient.
• Hidden‑state leakage justification: Shows that allowing a fraction of the hidden state to “leak” (i.e., decay toward zero) stabilizes replay and aligns the network’s dynamics with the derived gradients.
• Adaptation as exploration driver: Provides empirical evidence that negative‑feedback adaptation encourages the network to wander through state space, but also introduces non‑Markovian effects that can slow down sampling.
• Momentum‑based temporally compressed replay: Introduces a hidden‑state momentum term that yields under‑damped Langevin dynamics, compressing replay time while preserving exploratory behavior.
• Comprehensive validation: Tests the combined mechanisms on 2‑D triangular and T‑maze trajectories as well as high‑dimensional synthetic place‑cell data, demonstrating robust improvements over baseline noisy RNNs.

Methodology

1. Path‑integrating RNN model: The authors start with a recurrent network trained to integrate velocity inputs, mimicking how hippocampal place cells encode an animal’s position.
2. Noisy dynamics as sampling: By adding Gaussian noise to the hidden state updates, the network’s trajectory can be interpreted as a stochastic sampler of possible paths.
3. Analytical gradient analysis: Assuming a simple quadratic loss that measures deviation from the true path, they derive the optimal gradient field that the sampler should follow. This field turns out to be time‑dependent and hard to estimate online.
4. Introducing modifiers:
   • Leakage adds a decay term (‑λ h) to the hidden state update, effectively regularizing the dynamics.
   • Adaptation implements a negative‑feedback term that reduces the hidden state’s magnitude based on recent activity, encouraging the network to explore new regions.
   • Momentum adds a velocity‑like hidden variable that carries inertia across updates, turning the dynamics into an under‑damped Langevin process.
5. Simulation suite: The modified RNNs are evaluated on synthetic navigation tasks (triangular loop, T‑maze) and on high‑dimensional place‑cell representations generated from a simulated rat moving through a maze.

Results & Findings

When leakage, adaptation, and momentum are combined, the network reproduces biologically realistic replay: compressed in time (≈10–20 ms per replay segment) yet still explores alternative routes, matching experimental observations of hippocampal replay bursts.

Practical Implications

• Memory‑augmented AI: Incorporating leakage and momentum into RNN‑based memory modules could enable faster, more diverse retrieval of past experiences, improving planning in reinforcement‑learning agents.
• Robust sequence generation: The momentum‑driven under‑damped dynamics offer a principled way to generate temporally compressed yet plausible sequences—useful for video prediction, speech synthesis, or any task requiring “fast‑forward” replay of learned patterns.
• Neuro‑inspired hardware: The simple linear leakage and momentum terms map well onto analog neuromorphic circuits, suggesting low‑power implementations of stochastic replay for edge devices.
• Exploration strategies: Adaptation provides a biologically grounded mechanism for intrinsic exploration that could replace hand‑crafted curiosity bonuses in RL, especially in environments where external rewards are sparse.

Limitations & Future Work

• Assumption of quadratic loss: The analytical gradient derivation relies on a simple loss landscape; real‑world tasks may involve more complex, non‑convex objectives.
• Non‑Markovian side effects: Adaptation introduces history dependence that can complicate theoretical guarantees about convergence to the target distribution.
• Scalability to very high dimensions: While the paper shows improvements up to a few hundred dimensions, scaling to millions of units (e.g., large language models) remains untested.
• Biological validation: The proposed momentum term aligns with under‑damped Langevin theory, but direct physiological evidence for such inertia in hippocampal circuits is still lacking.

Future research could explore adaptive schedules for leakage and momentum, integrate these ideas into transformer‑style memory architectures, and seek experimental collaborations to verify the presence of momentum‑like dynamics in vivo.

Authors

• Josue Casco‑Rodriguez
• Nanda H. Krishna
• Richard G. Baraniuk

Paper Information

• arXiv ID: 2602.18401v1
• Categories: cs.LG, cs.AI, q-bio.NC, stat.ML
• Published: February 20, 2026
• PDF: Download PDF