[Paper Review] DualNet: Continual Learning, Fast and Slow

Quang Pham, Liu Chenghao+1 more|arXiv (Cornell University)|Oct 1, 2021

Domain Adaptation and Few-Shot LearningComputer Science52 references60 citations

TL;DR

DualNet proposes a two-system continual learning framework with a slow self-supervised learner for general representations and a fast learner for rapid adaptation to new labeled data, synchronized online. It outperforms state-of-the-art baselines on CORE50 and miniImageNet benchmarks.

ABSTRACT

According to Complementary Learning Systems (CLS) theory~\citep{mcclelland1995there} in neuroscience, humans do effective \emph{continual learning} through two complementary systems: a fast learning system centered on the hippocampus for rapid learning of the specifics and individual experiences, and a slow learning system located in the neocortex for the gradual acquisition of structured knowledge about the environment. Motivated by this theory, we propose a novel continual learning framework named "DualNet", which comprises a fast learning system for supervised learning of pattern-separated representation from specific tasks and a slow learning system for unsupervised representation learning of task-agnostic general representation via a Self-Supervised Learning (SSL) technique. The two fast and slow learning systems are complementary and work seamlessly in a holistic continual learning framework. Our extensive experiments on two challenging continual learning benchmarks of CORE50 and miniImageNet show that DualNet outperforms state-of-the-art continual learning methods by a large margin. We further conduct ablation studies of different SSL objectives to validate DualNet's efficacy, robustness, and scalability. Code will be made available upon acceptance.

Motivation & Objective

• Motivate continual learning via Complementary Learning Systems (CLS) theory by separating fast task-specific learning from slow general representation learning.
• Decouple representation learning from supervised learning using a slow SSL-based module and a fast, per-sample adaptation module.
• Demonstrate that synchronously trained slow and fast learners improve forgetting, transfer, and overall accuracy on challenging benchmarks.
• Show robustness of DualNet to different SSL objectives and optimizers and analyze ablations and semi-supervised settings.

Proposed method

• Two-model architecture: a slow learner φ trained with a self-supervised loss on memory samples to learn general representations, and a fast learner θ that adapts these representations for labeled data via a per-sample transformation mechanism.
• Slow learner objective: adopt a self-supervised loss; the paper implements BarLow Twins-style loss to encourage invariances and reduce redundancy.
• Optimization of the slow learner uses Look-ahead SGD to enhance convergence given the evolving memory distribution in continual learning.
• Fast learner objective: per-sample adaptation that conditions on the input to produce adapted features, enabling rapid learning from new labeled samples without task identifiers.
• Experience replay: combine incoming labeled data with memory samples using a soft-label loss and KL-divergence to stabilize learning.
• Training is synchronized: the slow SSL objective runs in the background while the fast learner updates on new labeled data, with backpropagation through both components.

Experimental results

Research questions

• RQ1 Can decoupled slow SSL-based representation learning paired with a fast per-sample adaptation mechanism improve continual learning performance over state-of-the-art methods?
• RQ2 Does synchronous training of slow and fast learners enable robust knowledge retention and transfer across tasks in online continual learning settings?
• RQ3 How does DualNet's performance vary with different SSL objectives and optimization strategies?
• RQ4 Is DualNet effective under semi-supervised continual learning where labeled data are scarce?
• RQ5 How does varying the amount of SSL iterations between supervised updates affect forgetting and transfer?

Key findings

• DualNet outperforms baselines (including ER, DER++, CTN) on CORE50 and Split miniImageNet benchmarks in both task-aware and task-free settings.
• The slow learner trained with SSL provides a robust, task-agnostic representation that supports better generalization across tasks.
• Per-sample adaptation in the fast learner enables rapid acquisition of new knowledge without requiring task identifiers.
• Using Look-ahead optimization for the slow learner and BarLow Twins SSL yields strong convergence and performance gains, with ablations confirming the efficacy of the chosen SSL objective.
• DualNet remains effective with varying SSL iterations, showing improved forgetting and transfer as more SSL updates are performed between supervised updates.
• In semi-supervised settings with partial labeling, DualNet leverages unlabeled data via the slow SSL branch to improve performance beyond baselines.