A Robust Model of Gated Working Memory

Anthony Strock, Xavier Hinaut, Raphaël MonatbioRxiv 2019 · DOI

TL;DR

This paper proposes a novel approach to modeling gated working memory using a simple, random recurrent neural network (a reservoir model) instead of explicit, dedicated mechanisms. The key insight is that robust working memory capabilities can emerge implicitly from the dynamics of such a network, even with minimal hypotheses about its architecture.

What Problem Does This Paper Address?

Working memory, the ability to hold and manipulate information over short periods, is crucial for higher cognitive functions. While neurophysiological evidence points to the prefrontal cortex (PFC) as a key region, the exact mechanisms are debated. Some models propose explicit "gates" and sustained neural activity to maintain information. However, recent findings suggest that information might be encoded in more distributed, high-dimensional ways, challenging these explicit models. This paper tackles the question of whether working memory can be an implicit property of a neural system, emerging from the interaction of its components rather than being hardwired. Specifically, they aim to model "gated working memory," where information can be stored and retrieved at arbitrary times, and explore how a random recurrent neural network can achieve this.

Methods

The authors propose a reservoir computing (RC) approach, specifically using an Echo State Network (ESN). In RC, a fixed, randomly connected recurrent neural network (the reservoir) processes input signals, and only a linear readout layer is trained to produce the desired output. This bypasses the computationally expensive training of all recurrent weights.

Key aspects of their method include:

  • Implicit Gating: Instead of explicit gating mechanisms, the model learns to "gate" information through the learned weights of the readout layer, exploiting the non-linear dynamics of the reservoir.
  • Minimal Model: They first introduce a simplified, three-neuron "minimal model" to illustrate the core principles of gating using non-linearities and asymptotic behaviors of tanh units. This model serves as a conceptual stepping stone.
  • Reservoir Model: The main model is an ESN with leaky neurons and feedback from readout units to the reservoir. The reservoir's connectivity is random but structured with specific sparsity and spectral radius.
  • Tasks: They define several working memory tasks, including scalar value storage and retrieval (1-value 1-gate, 1-value 3-gate, 3-value 1-gate) and a digit recognition task. These tasks involve receiving an input value, a trigger signal to store it, and then later retrieving it.

Datasets

The paper does not use pre-existing datasets in the traditional sense. Instead, they generate synthetic data for their defined working memory tasks. This includes:

  • Scalar values: Randomly sampled values from a uniform distribution between -1 and +1.
  • Trigger signals: Binary signals generated with a specific probability (e.g., 0.01) to indicate when to store information.
  • Digit inputs: Grayscale bitmap representations of digits (0-9) generated using a specific font.

Key Insights and Ideas

  • Working Memory as an Emergent Property: The most significant insight is that robust gated working memory can emerge implicitly from a randomly connected recurrent neural network. This challenges the necessity of explicit, hardwired gating mechanisms.
  • Robustness to Hyperparameters: The model demonstrates remarkable robustness across a wide range of hyperparameters (number of neurons, sparsity, spectral radius, etc.). This suggests that the learned working memory capability is not overly sensitive to precise network configurations.
  • Segment Attractor Dynamics: In the absence of input, the reservoir dynamics settle into an "approximate segment attractor," which is essentially a line attractor with bounded values. This provides a theoretical framework for how information is maintained.
  • Sustained Activity Not Mandatory: Contrary to some traditional models, the paper shows that clear, sustained neural activity is not strictly necessary for working memory. The information can be decoded even when neural activity is highly variable or transient.
  • Implicit Gating via Readout: The gating mechanism is implicitly learned through the readout layer, which learns to selectively access information from the reservoir based on the trigger signal.

Key Results

  • High Precision: The reservoir model achieves high precision (RMSE on the order of 10⁻³ to 10⁻⁴) in storing and retrieving scalar values across various tasks.
  • Generalization: The model shows good generalization performance, even when trained on a limited number of discrete values (as few as four).
  • Robustness: Performance remains largely stable across a broad range of hyperparameter values, indicating a robust mechanism.
  • Digit Recognition: The model successfully performs a non-linear task (digit recognition) in addition to memory tasks, demonstrating its versatility.
  • Minimal Model Equivalence: The minimal three-neuron model successfully captures the essential gating behavior, and equivalent sub-populations can be identified within the larger reservoir model.

Limitations and Future Work

  • Offline Training: While the primary analysis uses offline training (training the readout layer after the reservoir is fixed), the authors briefly mention that online learning (e.g., FORCE learning) also yields similar results, suggesting a path for more biologically plausible learning.
  • Minimal Model Proof: The authors acknowledge that while their minimal model demonstrates the principles, they haven't formally proven that a model with only two neurons in the reservoir cannot solve the task.
  • Biological Plausibility: While the model offers an alternative to explicit mechanisms, further work is needed to directly map its emergent properties to specific neural circuits and their dynamics in the brain.
  • Complex Tasks: Future work could explore more complex working memory tasks involving multiple items or more intricate manipulations of information.

Summary updated

Abstract

Abstract Gated working memory is defined as the capacity of holding arbitrary information at any time in order to be used at a later time. Based on electrophysiological recordings, several computational models have tackled the problem using dedicated and explicit mechanisms. We propose instead to consider an implicit mechanism based on a random recurrent neural network. We introduce a robust yet simple reservoir model of gated working memory with instantaneous updates. The model is able to store an arbitrary real value at random time over an extended period of time. The dynamics of the model is a line attractor that learns to exploit reentry and a non-linearity during the training phase using only a few representative values. A deeper study of the model shows that there is actually a large range of hyper parameters for which the results hold (number of neurons, sparsity, global weight scaling, etc.) such that any large enough population, mixing excitatory and inhibitory neurons can quickly learn to realize such gated working memory. In a nutshell, with a minimal set of hypotheses, we show that we can have a robust model of working memory. This suggests this property could be an implicit property of any random population, that can be acquired through learning. Furthermore, considering working memory to be a physically open but functionally closed system, we give account on some counter-intuitive electrophysiological recordings.
BibTeX
@misc{strock2019robust,
  author = {Anthony Strock and Xavier Hinaut and Raphaël Monat},
  title = {{A Robust Model of Gated Working Memory}},
  year = {2019},
  doi = {10.1101/589564},
  howpublished = {bioRxiv},
}
Figure