Hybrid Learners Do Not Forget: A Brain-Inspired Neuro-Symbolic Approach to Continual Learning

TL;DR

The paper tackles catastrophic forgetting in continual learning by introducing NeSyBiCL, a brain-inspired neuro-symbolic framework that marries a fast neural reasoner with a persistent symbolic reasoner guided by a knowledge base of class graphs. A fixed feature extractor powers both pathways, which are coupled through an integration loss to transfer abstract graph information into neural representations. The approach yields state-of-the-art performance on two compositional benchmarks, with the symbolic component providing zero forgetting and the neural component delivering rapid adaptation, and demonstrates meaningful knowledge transfer between systems. This work highlights the value of a dual-system strategy for robust continual learning and offers a viable path toward real-world, compositional AI tasks.

Abstract

Continual learning is crucial for creating AI agents that can learn and improve themselves autonomously. A primary challenge in continual learning is to learn new tasks without losing previously learned knowledge. Current continual learning methods primarily focus on enabling a neural network with mechanisms that mitigate forgetting effects. Inspired by the two distinct systems in the human brain, System 1 and System 2, we propose a Neuro-Symbolic Brain-Inspired Continual Learning (NeSyBiCL) framework that incorporates two subsystems to solve continual learning: A neural network model responsible for quickly adapting to the most recent task, together with a symbolic reasoner responsible for retaining previously acquired knowledge from previous tasks. Moreover, we design an integration mechanism between these components to facilitate knowledge transfer from the symbolic reasoner to the neural network. We also introduce two compositional continual learning benchmarks and demonstrate that NeSyBiCL is effective and leads to superior performance compared to continual learning methods that merely rely on neural architectures to address forgetting.

Figures (3)

• Figure 1: An overview of the proposed framework. (a)-(b) examples from the Clevr and Sketch compositional datasets. In both datasets, the combination of objects (subconcepts) with different shapes, colors, and relational locations form a new class, while other factors, such as rotation, color fullness, or material, are irrelevant. As an example, in the Sketch domain, a red triangle to the right of an orange one forms one class, while the combination of a blue circle, a yellow pentagon on the right, and a blue rectangle on the left creates another class. (c) A schematic of NeSyBiCL which consists of two complementary parallel pathways for data processing. Inspired by the dual thinking system, the symbolic reasoner decomposes the input into its subcomponents/relations and is the best candidate to solve the older tasks with no forgetting. The neural reasoner is responsible for effectively solving the contemporary task. The two components are also coupled through a knowledge integration loss.
• Figure 2: Learning curves for a long continual learning episode consisting of 50 tasks in the Sketch domain: (Left) the average accuracy on all previous tasks at each time step. We observe that NeSyBiC and the symbolic reasoner lead to minimal forgetting effects. (Right) the accuracy on the test samples of the last task at each timestep. We observe that NeSyBiC achieves better performance when solving the contemporary task compared to the symbolic learner.
• Figure 3: Ablative experiments: (Left) performance of the neural reasoner and the symbolic reasoner for learning a single task versus uncertainty in the Sketch domain. We see that the neural reasoner is more robust against the injection of uncertainty compared to the symbolic reasoner. (Right) the average accuracy on the last tasks during the continual episode ($\mathcal{A}_last$), versus the number of training samples in the Sketch dataset. We see the effect of the integration loss and also that the performance of the neural reasoner is highly affected by the number of training samples.