Summary: In AI research, larger models are often assumed to be better, but that approach drives up energy use and computational cost. Taking inspiration from human brain development, a research team has introduced a brain-inspired “selective pruning” framework for spiking neural networks (SNNs) that reduces network size while improving continual learning across perception, motor control, and interaction tasks.
The study shows that AI systems do not need ever-denser connectivity to master complex tasks; they need the right connections. By imitating how an infant brain strengthens useful long-range links while pruning redundant local connections, this method preserves and reuses knowledge, mitigates forgetting, and becomes more energy-efficient over time.
Key Facts
- The “Infant” Approach: Instead of only adding connections, the model matures core modules first—such as perception—then progresses to higher-level capabilities, following a simple-to-complex developmental order.
- Selective Pruning: Rather than freezing weights to avoid forgetting, this framework uses feedback-guided inhibition to remove redundant local connections formed during earlier learning phases.
- Knowledge Reuse: While pruning local clutter, the system strengthens cross-regional long-range connections so that high-level structures and concepts can be reused for new tasks without increasing overall network size.
- Reduced Forgetting: The developmental mechanism substantially reduces catastrophic forgetting without relying on memory-intensive approaches like experience replay, regularization schedules, or weight freezing.
- Sustainable Growth: The network becomes progressively more compact as it acquires skills, offering a low-energy route toward broader, general cognitive abilities.
Source: Science China Press
How can artificial intelligence keep improving efficiently?
For years, increasing model size has been the main tactic for boosting neural network performance, but that strategy drives up electricity consumption and hardware demands. By contrast, biological brains expand and then refine connectivity: early growth is followed by selective pruning that preserves critical pathways while removing redundant ones.
Motivated by these biological principles, the team developed a temporally developmental continual learning framework for spiking neural networks. The method stages the formation and reorganization of connections across modules and time, allowing the system to learn in a perception→motor→interaction sequence while reducing network scale and energy consumption.
Temporally Development–Inspired Continual Learning Mechanism
Brain development follows a temporal pattern: connectivity initially increases and then becomes refined. During this process, long-range cross-regional links are progressively strengthened while local connections that add little value are pruned. Primary regions mature earlier to support higher cognition, and feedback from higher-level areas optimizes lower-level structures. Inspired by this timeline, the researchers implemented a development-inspired continual learning method for SNNs.
The approach lets cognitive modules in the network grow in sequence—perception, then motor control, then interaction—while evolving inter-module long-range connections to promote positive transfer of knowledge. Concurrently, feedback-driven inhibition prunes redundant local connections tied to earlier tasks, enabling the network to shrink in size yet retain functional capabilities.
Energy-Efficient Cross-Domain Continual Learning
Experiments demonstrate that this method delivers stable continual learning across diverse cognitive domains, including perception, motor control, and interactive tasks, and achieves strong performance on standard continual learning benchmarks. Compared to direct training or blind pruning, the temporally staged approach learns complex tasks more reliably along a simple-to-complex trajectory.
Even as the network progressively reduces in scale, it preserves memory of prior tasks and significantly mitigates catastrophic forgetting. Analysis indicates this improvement stems from brain-like network dynamics: local connections expand early and are later inhibited and pruned to remove outdated or interfering information, while long-range connections progressively strengthen to support reuse of shared structures and high-level concepts.
Crucially, these gains are achieved without conventional continual learning tools such as replay buffers, explicit regularization, or freezing of parameters, pointing to a biologically plausible path for low-energy, scalable cognitive learning in AI.
The authors argue that developmental brain mechanisms offer a promising design principle for creating AI systems that grow smarter while consuming less energy—an important direction for sustainable, general-purpose intelligence.
Key Questions Answered:
A: Not necessarily. The model prunes redundant local patterns while preserving high-level concepts via strengthened long-range links. This removes noisy, task-specific details but keeps the abstract knowledge needed for later tasks.
A: SNNs process information in discrete spikes, which is closer to biological neural signaling than continuous activations. That spiking behavior, combined with selective pruning, enables highly energy-efficient implementations.
A: Instead of increasing parameters and power consumption as capabilities grow, this approach follows a biological growth-and-refinement curve: the network becomes more compact and efficient as it matures, lowering energy and computational requirements.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The referenced journal paper was reviewed in full.
- Additional context was provided by the editorial staff.
About this AI and neuroscience research news
Author: Bei Yan
Source: Science China Press
Contact: Bei Yan – Science China Press
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Continual Learning of Multiple Cognitive Functions with Brain-inspired Temporal Development Mechanism” by Bing Han, Feifei Zhao, Yinqian Sun, Wenxuan Pan, and Yi Zeng.
DOI: 10.1093/nsr/nwag066
Abstract
Continual Learning of Multiple Cognitive Functions with Brain-inspired Temporal Development Mechanism
Current artificial intelligence systems often tie new capabilities to exponential growth in network size, while the human brain learns many functions with remarkably low energy use. This efficiency partly stems from cross-regional temporal development: the staged formation, reorganization, and pruning of connections from basic to advanced regions facilitates knowledge transfer and prevents redundancy.
Building on these principles, the proposed TD-MCL (Temporal Development–inspired Continual Learning) framework enables cognitive enhancement from simple to complex across perception, motor control, and interaction tasks. It promotes sequential strengthening of long-range inter-module links to encourage positive knowledge transfer, while feedback-guided local inhibition and pruning remove redundancies from prior tasks.
Experiments on cross-domain perception–motor–interaction datasets and general benchmarks (including CIFAR-100 and ImageNet subsets) show that TD-MCL can achieve continual learning while reducing network scale, without relying on replay, regularization, or freezing strategies, and often delivers higher accuracy on new tasks than direct training. These results suggest that developmental brain mechanisms provide a useful template for biologically plausible, low-energy routes to more general cognitive abilities in AI.