Abstract
"You never forget how to ride a bike" — but how is that possible? The brain is able to learn complex skills, stop the practice for years, learn other skills in between, and still retrieve the original knowledge when necessary. The mechanisms of this capability, referred to as lifelong learning (or continual learning, CL), are unknown. We suggest a bio-plausible meta-plasticity rule building on classical work in CL which we summarize in two principles: (i) neurons are context selective, and (ii) a local availability variable partially freezes the plasticity if the neuron was relevant for previous tasks. In a new neuro-centric formalization of these principles, we suggest that neuron selectivity and neuron-wide consolidation is a simple and viable meta-plasticity hypothesis to enable CL in the brain. In simulation, this simple model balances forgetting and consolidation leading to better transfer learning than contemporary CL algorithms on image recognition and natural language processing CL benchmarks.
Introduction
Artificial Neural Networks (ANNs) excel at learning specific tasks but suffer from catastrophic forgetting: when trained on a new task, they tend to overwrite the weights necessary for previous ones. This is a manifestation of the stability-plasticity dilemma—the need for a system to be stable enough to retain old knowledge but plastic enough to acquire new information.
While the human brain manages this effortlessly, current AI approaches often rely on "replay" (storing and re-training on old data), which is memory-intensive and often biologically implausible. Furthermore, many "regularization" methods that freeze important weights eventually "saturate"—the network runs out of free parameters and becomes unable to learn anything new. GateON (Gated Obstruction of Neural updates) is a biologically inspired framework designed to be computationally efficient and capable of handling a large, open-ended stream of tasks without saturation.
Methods: The GateON Framework
The authors propose two core principles to address the stability-plasticity dilemma:
- Gated Context Selectivity: Each neuron is equipped with a context-dependent gate. For any given task, only a subset of neurons is "active." This induces sparsity and ensures that different tasks utilize different (though potentially overlapping) neural pathways.
- Gradient Obstruction via Dynamic Availability: This is the "meta-plasticity" rule. Each parameter (or neuron) has an availability variable ($A$) between 0 and 1. If a neuron is highly relevant to a task, its availability drops toward 0, "freezing" its weights to protect that knowledge. Crucially, if a neuron becomes irrelevant for a long time, its availability slowly drifts back toward 1, "unfreezing" it for future use.
The paper introduces two versions of this mechanism:
- p-GateON: A parameter-centric version where every weight has its own availability variable.
- n-GateON: A more biologically plausible neuro-centric version where the entire neuron (all its incoming weights) freezes or unfreezes as a single unit. They show that a neuron's own activity ($x^2$) can serve as a local proxy for relevance, eliminating the need for complex global error signals.
Experimental Results
The researchers evaluated GateON across several challenging benchmarks, including MNIST variants (Permuted, Rotated, Shuffled, and Split), Split CIFAR-100, and NLP benchmarks using pre-trained BERT models.
Key Findings:
- Superior Scaling: On a 100-task MNIST challenge, GateON maintained ~95% accuracy, while standard regularization methods like EWC saw performance plummet as they saturated. As shown in the figures above, accuracy remains high even as the total network availability decreases.
- State-of-the-Art on CIFAR: n-GateON achieved 84.6% accuracy on Split CIFAR-100, significantly outperforming EWC (75.3%).
- NLP Excellence: When applied to BERT, GateON ranked in the top two across all metrics, often beating algorithms that use "perfect replay" of old data. In some cases, "negative forgetting" was observed, where the network improved on old tasks while learning new ones.
Discussion and Insights
One of the most surprising findings is that freezing at the neuron level (n-GateON) is nearly as effective as freezing individual synapses. This aligns with biological observations of neuron-wide consolidation and is computationally efficient.
Furthermore, because GateON doesn't strictly partition the network, new tasks can "recruit" neurons that were useful for similar previous tasks. This leads to forward transfer, where the network learns later tasks faster because of earlier knowledge. Most importantly, the ability to "unfreeze" parameters (controlled by a threshold hyperparameter $\epsilon$) allows GateON to avoid the saturation trap that plagues other continual learning methods.
Conclusion and Future Work
While GateON demonstrates robust performance, some challenges remain. The primary version assumes the network is told which "context" it is currently performing, though the authors' unsupervised "Change-point detection" algorithm shows promise. Future work will explore the exact molecular mechanisms that might implement the "availability" integrator in biological neurons and investigate how these principles scale to massive generative models like LLMs.
Summary updated