Asynchronous Hebbian/anti-Hebbian networks (2501.02402v1)

Abstract: Lateral inhibition models coupled with Hebbian plasticity have been shown to learn factorised causal representations of input stimuli, for instance, oriented edges are learned from natural images. Currently, these models require the recurrent dynamics to settle into a stable state before weight changes can be applied, which is not only biologically implausible, but also impractical for real-time learning systems. Here, we propose a new Hebbian learning rule which is implemented using plausible biological mechanisms that have been observed experimentally. We find that this rule allows for efficient, time-continuous learning of factorised representations, very similar to the classic noncontinuous Hebbian/anti-Hebbian learning. Furthermore, we show that this rule naturally prevents catastrophic forgetting when stimuli from different distributions are shown sequentially.

• The paper introduces a new asynchronous Hebbian plasticity rule that allows neural networks to learn continuously without requiring system stabilization, a key limitation of traditional methods.
• This asynchronous approach successfully learns sparse, factorized representations while effectively preventing catastrophic forgetting, where new learning erases previous knowledge.
• The model offers significant potential for applications requiring real-time adaptation, energy efficiency, and long-term continual learning in dynamic artificial intelligence systems.

The paper "Asynchronous Hebbian/anti-Hebbian networks" explores a new approach to how neural networks can learn and remember important features from sensory inputs, inspired by biological processes. The goal of this research is to develop a system that learns efficiently and continuously, similar to how our brains function.

Background and Importance

Neural networks try to learn patterns and features from data inputs. One classic method for training these networks utilizes Hebbian learning—often summarized by the phrase "cells that fire together, wire together"—which strengthens connections between neurons that are activated together. The opposite, anti-Hebbian learning, weakens connections. Such models can extract meaningful features from complex data, like identifying edges in an image. However, traditional Hebbian models require the network to stabilize before making updates, which isn't biologically realistic and is inefficient for real-time applications.

Proposed Model

This paper introduces a new Hebbian plasticity rule that allows neurons to update their connections asynchronously without needing the recurrent neural network to stabilize first. This is more in tune with actual biological processes where neuron activities and learning happen continuously and simultaneously. Here are the main features of the proposed model:

• Asynchronous Updates: Neurons only update their connections when a certain activity level is reached, unlike traditional models that require a stable system.
• Refractory Period: After a neuron updates, there is a waiting period before it can update again. This concept is similar to a natural biological refractory period observed during Long-Term Potentiation (LTP).

Key Findings

• Continuous Learning: The asynchronous model successfully learns factorized representations similar to traditional networks that require equilibrium, but does so more continuously and efficiently.
• Preventing Catastrophic Forgetting: An exciting result of this approach is its ability to resist forgetting past information when learning new stimuli. Traditional networks often forget old information when new tasks are learned unless extra measures are taken.
• Sparsity and Efficiency: The representations learned by the model are sparse, meaning only a few neurons are active at a given time. This is efficient both computationally and biologically, as it can reduce energy use and increase memory capacity.

Technical Explanation

In Hebbian learning models, specific neurons learn to represent particular features in the input data. The asynchronous model achieves this using several mathematical techniques:

• Feed-forward Weights Updates: The update rule uses the neuron's activity and input signals to adjust the weights, applying an asynchronous gating mechanism to determine when updates can happen.

$\Delta w_{i,j} = \beta_{i}(y_{i}x_{j} - w_{i,j})$

Where: - $\Delta w_{i,j}$ is the change in weight for neuron $i$ from input $j$. - $y_{i}$ is the activity of the post-synaptic neuron $i$. - $x_{j}$ is the input from the pre-synaptic neuron $j$. - $\beta_{i}$ is a gating function that governs whether an update occurs based on activity.

• Recurrent Weights: Similarly, this method works for recurrent connections where the neuron receives feedback from other neurons including itself, helping the network quickly adapt to the information.

Practical Implications and Recommendations

For researchers developing neural networks in artificial intelligence, this asynchronous approach can be significant:

• Real-Time Learning: Models that need to adapt quickly to changing data or operate in dynamic environments could benefit from these methods.
• Energy Efficiency: In environments with energy constraints, such as on portable devices, asynchronous methods may be preferable due to their sparsity and reduced computation requirements.
• Continual Learning Applications: Systems that engage in long-term learning tasks, such as robots or online learning algorithms, could apply these methods to avoid forgetting previous experiences.

Conclusion

The paper emphasizes the potential of asynchronous Hebbian learning as a biologically inspired, efficient method for teaching artificial neural networks. This model addresses critical limitations of traditional Hebbian models, allows for continuous updates without requiring network stability, and inherently prevents catastrophic forgetting, making it a promising approach to designing more effective and durable neural networks.