[Paper Review] Stimulating STDP to Exploit Locality for Lifelong Learning without Catastrophic Forgetting.

Jason M. Allred, Kaushik Roy|arXiv (Cornell University)|Feb 8, 2019

Domain Adaptation and Few-Shot Learning10 references5 citations

TL;DR

This paper proposes a spiking neural network (SNN) framework that leverages Spike Timing Dependent Plasticity (STDP) with dopamine-inspired stimulation to selectively update only the neurons processing novel, outlier information, thereby preventing catastrophic forgetting during lifelong learning. On MNIST, the system achieves up to 93.88% accuracy when learning disjoint datasets sequentially, demonstrating robustness in uncontrolled, dynamic environments.

ABSTRACT

Stochastic gradient descent requires that training samples be drawn from a uniformly random distribution of the data. For a deployed system that must learn online from an uncontrolled and unknown environment, the ordering of input samples often fails to meet this criterion, making lifelong learning a difficult challenge. We exploit the locality of the unsupervised Spike Timing Dependent Plasticity (STDP) learning rule to target subsets of a segmented Spiking Neural Network (SNN) to adapt to novel information while protecting the information in the remainder of the SNN from catastrophic forgetting. In our system, novel information triggers stimulated firing, inspired by biological dopamine signals, to boost STDP in the synapses of neurons associated with outlier information. This targeting controls the forgetting process in a way that reduces accuracy degradation while learning new information. Our preliminary results on the MNIST dataset validate the capability of such a system to learn successfully over time from an unknown, changing environment, achieving up to 93.88% accuracy for a completely disjoint dataset.

Motivation & Objective

• Address the challenge of catastrophic forgetting in lifelong learning systems operating in uncontrolled, non-stationary environments.
• Overcome the limitation of stochastic gradient descent requiring uniformly random data ordering, which is often violated in real-world online learning.
• Leverage the locality of STDP to selectively update only the neural subnetworks processing novel or outlier inputs.
• Minimize accuracy degradation on previously learned tasks while adapting to new information.
• Enable stable, incremental learning in deployed systems without requiring retraining or prior knowledge of data distribution.

Proposed method

• Segment the spiking neural network into subnetworks to localize learning and isolate memory regions.
• Use Spike Timing Dependent Plasticity (STDP) as the core unsupervised learning rule to strengthen synapses based on pre- and post-synaptic spike timing.
• Introduce dopamine-like stimulation signals that trigger enhanced STDP activity only in neurons processing novel or outlier inputs.
• Apply the stimulation selectively to target synapses associated with new information, thereby focusing plasticity and minimizing interference with existing knowledge.
• Dynamically identify outlier inputs through spiking activity patterns and activate stimulation only when such inputs are detected.
• Maintain the integrity of previously learned representations by restricting plasticity to specific, contextually relevant subnetworks.

Experimental results

Research questions

• RQ1 Can STDP-based plasticity be selectively enhanced in response to novel or outlier inputs to reduce catastrophic forgetting?
• RQ2 How does dopamine-inspired stimulation improve the stability of learned representations during incremental learning?
• RQ3 To what extent can a spiking neural network learn from a sequence of disjoint datasets without performance degradation?
• RQ4 Does targeting plasticity to specific subnetworks based on input novelty improve learning efficiency and accuracy retention?
• RQ5 Can this approach function effectively in an online, non-stationary environment without requiring uniformly random data ordering?

Key findings

• The system successfully learns from a sequence of disjoint MNIST datasets without catastrophic forgetting.
• The proposed method achieves a test accuracy of up to 93.88% on the final dataset, indicating strong retention of prior knowledge.
• Selective stimulation of STDP in response to novel inputs effectively reduces interference with previously learned representations.
• The use of dopamine-inspired signals enables targeted plasticity, minimizing uncontrolled weight updates across the network.
• The approach maintains high performance across multiple learning phases, even when data distributions are completely disjoint.
• The system demonstrates robustness in uncontrolled, dynamic environments where data ordering violates the i.i.d. assumption of standard SGD.