[Paper Review] Neuromimetic metaplasticity for adaptive continual learning

Suhee Cho, Hyeonsu Lee+2 more|arXiv (Cornell University)|Jul 9, 2024

Neuroblastoma Research and TreatmentsMedicine3 citations

TL;DR

This paper proposes a neuromimetic metaplasticity framework that enables deep neural networks to perform continual learning without catastrophic forgetting, by dynamically varying synaptic flexibility through biologically inspired mechanisms. The model achieves stable retention of old and new knowledge across streaming data and demonstrates robustness against data poisoning by leveraging the Hebbian repetition effect to reinforce salient memories.

ABSTRACT

Conventional intelligent systems based on deep neural network (DNN) models encounter challenges in achieving human-like continual learning due to catastrophic forgetting. Here, we propose a metaplasticity model inspired by human working memory, enabling DNNs to perform catastrophic forgetting-free continual learning without any pre- or post-processing. A key aspect of our approach involves implementing distinct types of synapses from stable to flexible, and randomly intermixing them to train synaptic connections with different degrees of flexibility. This strategy allowed the network to successfully learn a continuous stream of information, even under unexpected changes in input length. The model achieved a balanced tradeoff between memory capacity and performance without requiring additional training or structural modifications, dynamically allocating memory resources to retain both old and new information. Furthermore, the model demonstrated robustness against data poisoning attacks by selectively filtering out erroneous memories, leveraging the Hebb repetition effect to reinforce the retention of significant data.

Motivation & Objective

• To address catastrophic forgetting in continual learning by emulating human working memory mechanisms in artificial neural networks.
• To enable deep learning models to retain both old and new knowledge without architectural or training modifications.
• To enhance robustness against data poisoning by selectively reinforcing significant memories through Hebbian repetition.
• To dynamically allocate memory resources based on input stream characteristics, including unexpected changes in input length.
• To achieve a balanced tradeoff between memory capacity and model performance in continual learning scenarios.

Proposed method

• The model implements a heterogeneous synaptic architecture with distinct types of synapses ranging from stable to flexible, enabling variable learning rates across connections.
• Synaptic flexibility is randomly intermixed during training, allowing the network to adaptively allocate memory resources to new and old information.
• The framework leverages the Hebb repetition effect to strengthen significant memories, enhancing resilience to data poisoning attacks.
• No pre- or post-processing or structural modifications are required, maintaining the original network architecture throughout continual learning.
• The approach dynamically adjusts memory allocation in response to changes in input length, ensuring stability under unexpected data variations.
• Metaplasticity is implemented through a biologically inspired mechanism that modulates synaptic efficacy based on activity history.

Experimental results

Research questions

• RQ1 Can a biologically inspired metaplasticity mechanism enable continual learning without catastrophic forgetting in DNNs?
• RQ2 How does variable synaptic flexibility impact memory retention and performance in streaming learning scenarios?
• RQ3 To what extent can the model maintain performance under unexpected input length variations?
• RQ4 Can the Hebbian repetition effect enhance robustness against data poisoning in continual learning?
• RQ5 What is the tradeoff between memory capacity and model accuracy in the proposed framework?

Key findings

• The model successfully learns a continuous stream of data without catastrophic forgetting, even when input lengths change unexpectedly.
• The framework achieves a balanced tradeoff between memory capacity and performance without requiring additional training or architectural changes.
• The model demonstrates robustness against data poisoning by selectively filtering out erroneous memories through Hebbian repetition.
• The dynamic allocation of synaptic flexibility enables effective retention of both old and new information across diverse learning tasks.
• The approach maintains high performance across multiple continual learning benchmarks, outperforming standard DNNs in stability and retention.
• The model's resilience to input variation and adversarial data is attributed to its neuromimetic metaplasticity mechanism.