Uncertainty, phase and oscillatory hippocampal recall

Máté Lengyel, Peter DayanNeurIPS 2006 · OpenReview

Introduction

Many neural areas, notably the hippocampus, exhibit structured, dynamical population behavior such as coordinated oscillations. Traditional theories suggest that the timing of a spike relative to this rhythm (its phase) encodes analog information, such as an animal's spatial position. However, these models usually assume neurons fire exactly once per cycle and ignore a fundamental requirement of neural computation: handling uncertainty.

When the brain retrieves a memory from a noisy or partial cue, it shouldn't just provide a "best guess"; it needs to know how reliable that guess is. This paper proposes that the "burstiness" or concentration of spikes within a hippocampal oscillation cycle isn't just noise—it’s a sophisticated code for uncertainty. By treating the CA3 region as a probabilistic retrieval system, the authors demonstrate how an oscillatory neural network can represent and manipulate this uncertainty without needing a separate, complex architecture.

The Coding Scheme

The authors develop a variational framework to interpret neural activity as a probability distribution. They identify three key parameters in a single oscillatory cycle that define this code:

  1. $r$ (Number of spikes): Represents the probability that a neuron is "active" in a memory pattern, capturing binary uncertainty.
  2. $\phi$ (Mean phase): Represents the analog value associated with that neuron (e.g., a specific location).
  3. $c$ (Concentration): Represents the certainty about that phase. Highly concentrated spikes (bursts) signal high certainty, while scattered spikes signal high uncertainty.
The proposed neural coding scheme illustrating how spikes within an oscillation cycle represent binary state, analog phase, and uncertainty.

By minimizing the Kullback-Leibler (KL) divergence between the network's internal state and the true Bayesian posterior distribution of the memory (given a noisy cue and stored synaptic weights), the authors derive recurrent neural network dynamics. This allows the network to naturally "settle" into a state that represents the most likely memory while explicitly signaling its confidence through spike concentration.

Key Insights and Experimental Validation

By using Bayes' rule to define the target distribution, the authors transform the standard Hopfield-style associative memory into a probabilistic inference machine. This yields several critical insights:

  • The "Whisper vs. Shout" Principle: Neurons that are "erring" or uncertain should "whisper" (fire with low concentration/rate) rather than "shout." This prevents a single uncertain neuron from corrupting the collective retrieval process of the entire population.
  • Burstiness as a Signal: Instead of viewing bursts as a metabolic quirk or a simple gain mechanism, the paper views the degree of burstiness as a fundamental channel for communicating confidence to downstream partners.
  • Dimensionality Advantage: Unlike population codes for single variables (like orientation), this model allows each neuron to represent its own independent analog value and its own uncertainty. This scales the representational power linearly with the number of neurons.

To validate these theoretical claims, the authors tested the model on both synthetic and real-world datasets:

  • Predictive Uncertainty in Simulations: In a simulated network of 100 neurons with 10 stored patterns, "burst strength" (a combination of rate and concentration) was inversely correlated with phase error. Essentially, when the model was "sure," it was usually right. The "full model" performed as well as specialized models that only handled rate or phase uncertainty, proving that adding the dimension of phase uncertainty doesn't compromise the network's ability to recall basic binary patterns.
  • Biological Validation: The authors analyzed extracellular recordings from rat hippocampal area CA1 during a linear track task. They found that the error in a neuron's firing phase (relative to its average "true" phase for a location) was indeed predicted by its burst strength, strongly suggesting the brain actually utilizes this coding scheme.

Limitations and Future Work

While the framework provides a compelling link between oscillations and probabilistic inference, the authors note a few limitations:

  • Simplified Firing Rates: The simulations treated firing rates as continuous real-valued variables for simplicity, whereas real neurons fire discrete, integer numbers of spikes.
  • Factorized Approximation: The model assumes neurons contribute independently to the probability distribution (a "mean-field" approach), which may ignore complex correlations between neurons.
  • Multi-Frequency Oscillations: The paper focuses on a single rhythm (like the theta rhythm). Future work is needed to understand how uncertainty is handled when multiple oscillations (e.g., theta and gamma) interact, or when different brain regions with different frequencies communicate.
  • CA1 vs. CA3: The experimental data used for validation came from CA1, but the model is specifically designed for the recurrently connected CA3 region. Direct testing in CA3 remains a necessary next step.

Summary updated

Abstract

Many neural areas, notably, the hippocampus, show structured, dynamical, population behavior such as coordinated oscillations. It has long been observed that such oscillations provide a substrate for representing analog information in the firing phases of neurons relative to the underlying population rhythm. However, it has become increasingly clear that it is essential for neural populations to represent uncertainty about the information they capture, and the substantial recent work on neural codes for uncertainty has omitted any analysis of oscillatory systems. Here, we observe that, since neurons in an oscillatory network need not only fire once in each cycle (or even at all), uncertainty about the analog quantities each neuron represents by its firing phase might naturally be reported through the degree of concentration of the spikes that it fires. We apply this theory to memory in a model of oscillatory associative recall in hippocampal area CA3. Although it is not well treated in the literature, representing and manipulating uncertainty is fundamental to competent memory; our theory enables us to view CA3 as an effective uncertainty-aware, retrieval system.
BibTeX
@inproceedings{lengyel2006uncertainty,
  author = {Máté Lengyel and Peter Dayan},
  title = {{Uncertainty, phase and oscillatory hippocampal recall}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2006},
  url = {https://openreview.net/forum?id=HyWHmYW_WH},
}
Figure