From statistical inference to a differential learning rule for stochastic neural networks

Luca Saglietti, Federica Gerace, Alessandro Ingrosso, Carlo Baldassi, Riccardo ZecchinaInterface 2018 · DOI

TL;DR

This paper introduces a novel "delayed-correlations matching" (DCM) learning rule for stochastic neural networks. DCM enables networks to learn and store patterns robustly, even with biologically plausible constraints like noisy signals and asymmetric connections, while avoiding spurious memories. This rule offers a unified framework for both attractor network learning and generative modeling.

What Problem Does This Paper Address?

Understanding how biological neural networks learn and store information in the face of inherent noise and biological constraints is a fundamental challenge in neuroscience and computational modeling. Traditional learning rules, like Hebbian plasticity, often struggle with these complexities, leading to issues like catastrophic forgetting, proliferation of spurious memories, and incompatibility with biological realism (e.g., Dale's principle, asymmetric synapses). This paper aims to bridge this gap by developing a learning rule that is both biologically plausible and capable of robustly learning complex statistical structures from data.

Methods

The core contribution is the derivation of the Delayed-Correlations Matching (DCM) learning rule. This rule is derived from a statistical inference perspective, specifically by minimizing the Kullback-Leibler (KL) divergence between conditional probability distributions of network states under different levels of external field intensity. The key innovation is that DCM relies only on delayed activity correlations observed during the network's dynamics. This allows for:

  • Biological Plausibility: DCM is local, uses finite and noisy signals, and is compatible with asymmetric synaptic connections and Dale's principle.
  • Robustness: It avoids catastrophic forgetting and the proliferation of spurious attractors, even with correlated patterns.
  • Generality: It applies to networks with or without hidden units, and supports one-shot learning.

The paper also explores connections to:

  • Maximum Pseudo-Likelihood: In the simplified case of fully visible networks and infinite initial fields, DCM reduces to optimizing the log-pseudo-likelihood.
  • Perceptron Rule: In the noise-free limit, DCM recovers the perceptron learning rule.
  • Boltzmann Machines: When hidden units are introduced, DCM can be understood within a KL minimization framework similar to learning for Boltzmann Machines.

The implementation involves a specific learning protocol where the external field is gradually reduced, and the network's dynamics are observed at different field levels to estimate the necessary correlations.

Datasets

  • MNIST dataset: Used for evaluating the network's performance in feature extraction, generation, and classification tasks, particularly in the context of visible-to-hidden directed synapses.

Key Insights and Ideas

  • Delayed Correlations as the Learning Signal: The paper demonstrates that learning can be effectively driven by matching time-delayed correlations in neural activity, rather than just instantaneous ones. This is a significant departure from many traditional plasticity rules.
  • Unified Framework for Inference and Learning: DCM provides a single learning rule that can address both unsupervised pattern storage (attractor networks) and generative modeling (networks with hidden units), bridging statistical inference and neural learning.
  • Biological Plausibility without Sacrificing Performance: The rule's ability to incorporate constraints like Dale's principle and asymmetric synapses while maintaining high storage capacity is a crucial insight, suggesting that biological realism doesn't necessarily mean reduced computational power.
  • Avoiding Spurious Attractors: A key advantage of DCM is its remarkable ability to avoid the generation of spurious memories, a common problem in attractor networks, leading to more coherent and stable memory representations.
  • One-Shot Learning Capability: The finding that DCM can achieve extensive palimpsest capacity (learning new memories without catastrophically overwriting old ones) in a one-shot setting is particularly noteworthy.

Key Results

  • High Storage Capacity: DCM can store an extensive number of patterns as attractors in stochastic recurrent neural networks, outperforming the standard Hopfield model in terms of storage load and robustness to noise (Figure 2).
  • Robustness to Correlated Patterns: The rule effectively handles correlated patterns without requiring explicit knowledge of their statistics, unlike modified Hebbian rules (Figure 5).
  • Reduced Spurious Attractors: DCM significantly reduces the number of spurious attractors compared to the naive Hebb rule, leading to more stable memory retrieval (Figure 7).
  • Effective One-Shot Learning: DCM achieves an extensive palimpsest capacity (around 0.05N) in a one-shot learning scenario, a significant achievement for this type of learning.
  • Generative and Classification Performance: When applied to the MNIST dataset with hidden units, DCM demonstrates reasonable performance in feature extraction, generation, and classification, even under biologically constrained conditions (Figure A1, A2, and performance metrics on page 29).

Limitations and Future Work

  • Computational Cost: While the learning rule is local, the heuristic implementation relies on Monte Carlo simulations to estimate time-delayed correlations, which can be computationally intensive. The paper mentions the TAP approximation as a potential analytical shortcut.
  • Continuous Time and Spiking Neurons: The current work focuses on discrete-time synchronous dynamics. Future work could generalize the framework to continuous-time dynamics and more realistic spiking neuron models.
  • Learning Dynamical Patterns: The current model learns static patterns. Extending it to learn dynamical activation patterns is a promising direction.
  • Generalization of Hidden Unit Architectures: While the paper explores visible-to-hidden directed synapses for RBM-like models, further investigation into more general architectures with hidden units is suggested.
  • Comparison with Other State-of-the-Art Methods: While comparisons are made, a more extensive comparison with other advanced learning algorithms, especially for complex tasks, would be beneficial.

Summary updated

Abstract

Stochastic neural networks are a prototypical computational device able to build a probabilistic representation of an ensemble of external stimuli. Building on the relationship between inference and learning, we derive a synaptic plasticity rule that relies only on delayed activity correlations, and that shows a number of remarkable features. Our <i>delayed-correlations matching</i> (DCM) rule satisfies some basic requirements for biological feasibility: finite and noisy afferent signals, Dale's principle and asymmetry of synaptic connections, locality of the weight update computations. Nevertheless, the DCM rule is capable of storing a large, extensive number of patterns as attractors in a stochastic recurrent neural network, under general scenarios without requiring any modification: it can deal with correlated patterns, a broad range of architectures (with or without hidden neuronal states), one-shot learning with the palimpsest property, all the while avoiding the proliferation of spurious attractors. When hidden units are present, our learning rule can be employed to construct Boltzmann machine-like generative models, exploiting the addition of hidden neurons in feature extraction and classification tasks.
BibTeX
@inproceedings{saglietti2018from,
  author = {Luca Saglietti and Federica Gerace and Alessandro Ingrosso and Carlo Baldassi and Riccardo Zecchina},
  title = {{From statistical inference to a differential learning rule for stochastic neural networks}},
  booktitle = {Interface},
  year = {2018},
  doi = {10.1098/rsfs.2018.0033},
}
Figure