Provable local learning rule by expert aggregation for a Hawkes network

Sophie Jaffard, Samuel Vaiter, Alexandre Muzy, Patricia Reynaud-Bouret

Abstract

We propose a simple network of Hawkes processes as a cognitive model capable of learning to classify objects. Our learning algorithm, named HAN for Hawkes Aggregation of Neurons, is based on a local synaptic learning rule based on spiking probabilities at each output node. We were able to use local regret bounds to prove mathematically that the network is able to learn on average and even asymptotically under more restrictive assumptions.

Introduction: Bridging the Gap Between Neurons and Algorithms

In neuroscience, learning is understood as a local process—synaptic changes occur based on the timing of spikes between adjacent neurons, such as in Spike-Timing-Dependent Plasticity (STDP). Conversely, modern machine learning predominantly relies on global rules like backpropagation, which requires information to flow backward through the entire network—a mechanism widely considered biologically implausible.

The central challenge addressed in this work is whether a network using only local, biologically inspired rules can be mathematically proven to learn a global task, such as classification. While Spiking Neural Networks (SNNs) are common, they often lack rigorous "oracle inequalities"—proofs that the network will perform as well as the best possible theoretical configuration. We fill this gap using the framework of Hawkes processes.

The HAN Model and Architecture

The proposed architecture, HAN (Hawkes Aggregation of Neurons), consists of a two-layer network of Hawkes processes. Hawkes processes are a type of point process where past events increase the probability of future events, mimicking the excitatory nature of neural firing.

Architecture of the Hawkes Aggregation of Neurons (HAN) network.

In this model, input neurons represent specific features (e.g., "blue," "square"), while output neurons represent categories (e.g., "Class A"). The core innovation lies in treating presynaptic (input) neurons as experts and the output neuron as a forecaster. Synaptic weights are interpreted as a probability distribution over these experts.

Instead of standard gradient descent, HAN utilizes Expert Aggregation algorithms, such as the Exponentially Weighted Average (EWA). When an object is presented, if an output neuron is intended to fire (correct class), it increases the weights for active "helpful" inputs. Conversely, if it should not fire, those weights are penalized. By applying the Kalikow decomposition of Hawkes processes, we justify treating weights as probabilities, allowing the application of regret bounds from online learning theory to prove convergence.

Key Results and Performance

Our analysis demonstrates that minimizing "regret" locally at each neuron directly translates to maximizing the "network discrepancy"—the system's ability to distinguish between classes. We prove an oracle inequality with a constant of 1, meaning that over time, the network performs as well as an "oracle" with perfect weight knowledge, with an error term that vanishes at a rate of $O(M^{-1/2})$.

Learning performance and weight convergence of the HAN model.

As shown in the results, HAN achieves perfect classification significantly faster than the traditional "Component-Cue" model from cognitive science. Furthermore, the full HAN model, which incorporates inhibitory connections and non-linear activation, outperforms the simplified linear "HAN Solo" version. This suggests that biological complexity, such as inhibition, provides significant computational advantages.

The weight dynamics illustrate how the network correctly identifies diagnostic features. For instance, weights for features associated with a specific class increase over time, while irrelevant or misleading features are suppressed, eventually converging to a stable, optimal configuration.

Conclusion and Future Directions

This work provides a rigorous mathematical foundation for local learning rules in spiking networks. By framing synaptic updates as expert aggregation, we bridge the gap between biological plausibility and algorithmic efficiency.

Future research will focus on extending these proofs to multi-layer architectures, as the current proof is limited to two-layer networks. Additionally, exploring how the brain learns to encode raw sensory data into features and investigating more nuanced temporal coding strategies remain important open questions.

Summary updated

Abstract

We propose a simple network of Hawkes processes as a cognitive model capable of learning to classify objects. Our learning algorithm, named HAN for Hawkes Aggregation of Neurons, is based on a local synaptic learning rule based on spiking probabilities at each output node. We were able to use local regret bounds to prove mathematically that the network is able to learn on average and even asymptotically under more restrictive assumptions.
BibTeX
@inproceedings{jaffard2024provable,
  author = {Sophie Jaffard and Samuel Vaiter and Alexandre Muzy and Patricia Reynaud-Bouret},
  title = {{Provable local learning rule by expert aggregation for a Hawkes network}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2024},
  url = {https://openreview.net/forum?id=cODYuXmAh4},
}
Figure