Continual Learning of Multiple Cognitive Functions with Brain-inspired Temporal Development Mechanism

TL;DR

The paper presents TD-MCL, a brain-inspired continual-learning framework that leverages temporal development to grow and prune Spiking Neural Network modules across perception, motor, and interaction tasks. By evolving long-range inter-module connections and applying feedback-guided local inhibition, the method achieves continual learning across a cross-domain PMI dataset without regularization, replay, or freezing, while progressively reducing network size. The approach demonstrates positive knowledge transfer between tasks, maintains or improves performance on prior tasks during pruning, and aligns with biological development patterns, offering scalable, energy-efficient cognition. This work has potential practical impact for low-power AI systems needing robust cross-domain continual learning and interpretable, brain-like organization of cognitive functions.

Abstract

Cognitive functions in current artificial intelligence networks are tied to the exponential increase in network scale, whereas the human brain can continuously learn hundreds of cognitive functions with remarkably low energy consumption. This advantage is in part due to the brain cross-regional temporal development mechanisms, where the progressive formation, reorganization, and pruning of connections from basic to advanced regions, facilitate knowledge transfer and prevent network redundancy. Inspired by these, we propose the Continual Learning of Multiple Cognitive Functions with Brain-inspired Temporal Development Mechanism(TD-MCL), enabling cognitive enhancement from simple to complex in Perception-Motor-Interaction(PMI) multiple cognitive task scenarios. The TD-MCL model proposes the sequential evolution of long-range connections between different cognitive modules to promote positive knowledge transfer, while using feedback-guided local connection inhibition and pruning to effectively eliminate redundancies in previous tasks, reducing energy consumption while preserving acquired knowledge. Experiments show that the proposed method can achieve continual learning capabilities while reducing network scale, without introducing regularization, replay, or freezing strategies, and achieving superior accuracy on new tasks compared to direct learning. The proposed method shows that the brain's developmental mechanisms offer a valuable reference for exploring biologically plausible, low-energy enhancements of general cognitive abilities.

Figures (3)

• Figure 1: The procedure of TD-MCL model.A) Correspondence between cognitive function development and brain structure development in children. B) Multiple Cognitive Functions Continual Learning Dataset Paradigm. C) General overview of the procedure of efficient continual multiple cognitive function learning algorithm. D) Local connection suppression and pruning integrating biological synaptic plasticity and knowledge generalization. E) Long-range connection growth based on evolutionary algorithms.
• Figure 2: Progressive continual learning performance.A-B) Examples of successful processes for motion control and environmental interaction tasks. C) Effect of pruning on continual learning performance. D) Fine-tuning performance of sparse networks after pruning. E) Comparison with other continual learning methods. F-G) Performance of ablation experiments. H) Module reuse determined by long-range connectivity.
• Figure 3: Local connectivity and long-range connectivity dynamics.A) Local connectivity changes in the overall and cognitive-functional networks.B) Local connectivity changes in single-task networks. D-E) Number of retained parameters from previous tasks. C, F) Correlation of biological synaptic plasticity with local connection pruning. G) Long-range connection parameter counts and sparsity. H-I) Long-range connection mode selection.