[Paper Review] Continual Competitive Memory: A Neural System for Online Task-Free Lifelong Learning

Alexander G. Ororbia|arXiv (Cornell University)|Jun 24, 2021

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper proposes Continual Competitive Memory (CCM), a novel unsupervised neural system for online, task-free lifelong learning that mitigates catastrophic forgetting through competitive Hebbian learning. CCM outperforms traditional competitive learning models and achieves performance competitive with state-of-the-art lifelong learning methods on Split MNIST and Split NotMNIST without requiring task descriptors or data replay.

ABSTRACT

In this article, we propose a novel form of unsupervised learning, continual competitive memory (CCM), as well as a computational framework to unify related neural models that operate under the principles of competition. The resulting neural system is shown to offer an effective approach for combating catastrophic forgetting in online continual classification problems. We demonstrate that the proposed CCM system not only outperforms other competitive learning neural models but also yields performance that is competitive with several modern, state-of-the-art lifelong learning approaches on benchmarks such as Split MNIST and Split NotMNIST. CCM yields a promising path forward for acquiring representations that are robust to interference from data streams, especially when the task is unknown to the model and must be inferred without external guidance.

Motivation & Objective

• To address catastrophic forgetting in online continual learning without relying on task descriptors or external supervision.
• To unify competitive neural models under a generalized Neural Competitive Learning (NCL) framework for streaming data.
• To develop a scalable, biologically plausible system that learns from data streams while retaining knowledge across tasks.
• To evaluate CCM's robustness and performance against established baselines in task-free continual learning settings.
• To enable integration of CCM with downstream predictors for improved generalization in continual learning.

Proposed method

• The NCL framework unifies competitive learning models by formalizing key components like competitive activation, weight updates, and memory unit dynamics.
• CCM uses a competitive mechanism where the most responsive neuron (winner) updates its weights based on input via a Hebbian-like rule, promoting memory consolidation.
• A dynamic thresholding mechanism determines whether a new input triggers the creation of a new memory unit or is stored in an existing one.
• The system operates in an online, streaming fashion, processing one sample at a time without storing raw inputs or relying on replay buffers.
• The framework supports incremental growth of memory units and incorporates mechanisms inspired by Adaptive Resonance Theory (ART) for stability.
• CCM is integrated with a feedforward predictor to evaluate performance on classification benchmarks like Split MNIST and Split NotMNIST.

Experimental results

Research questions

• RQ1 Can a competitive learning-based system achieve robust lifelong learning without task descriptors or external supervision?
• RQ2 How does CCM compare to classical competitive learning models in terms of knowledge retention and generalization under continual learning?
• RQ3 To what extent does competitive learning provide inherent resistance to catastrophic forgetting in streaming data scenarios?
• RQ4 Can CCM match or exceed the performance of modern, backpropagation-based lifelong learning methods on standard benchmarks?
• RQ5 How does the NCL framework generalize and unify diverse competitive learning models for continual learning?

Key findings

• CCM outperforms classical competitive learning models such as ART-C 2A and SOINN in terms of accuracy and knowledge retention on Split MNIST and Split NotMNIST.
• CCM achieves performance competitive with state-of-the-art lifelong learning methods, including those using regularization or experience replay, despite being unsupervised and task-free.
• The system demonstrates robust resistance to catastrophic forgetting, maintaining high accuracy across sequentially presented tasks without task boundary signals.
• The NCL framework successfully unifies diverse competitive models, including ART, SOM, and neural gas, under a common computational abstraction.
• CCM enables effective integration with feedforward predictors, showing strong generalization in continual classification tasks.
• The model exhibits stable memory growth without explicit upper bounds on unit count, though dynamic thresholding and unit merging remain open challenges.