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 local representations in a Spiking Neural Network (SNN) to adapt to novel information while protecting essential information in the remainder of the SNN from catastrophic forgetting. In our Controlled Forgetting Networks (CFNs), novel information triggers stimulated firing and heterogeneously modulated plasticity, inspired by biological dopamine signals, to cause rapid and isolated adaptation in the synapses of neurons associated with outlier information. This targeting controls the forgetting process in a way that reduces the degradation of accuracy for older tasks while learning new tasks. Our experimental results on the MNIST dataset validate the capability of CFNs to learn successfully over time from an unknown, changing environment, achieving 95.24% accuracy, which we believe is the best unsupervised accuracy ever achieved by a fixed-size, single-layer SNN on a completely disjoint MNIST dataset.

Highlights

  • Artificial neural networks have enabled computing systems to successfully perform tasks previously out of reach for traditional computing, such as image and audio classification

  • When traditional artificial neural networks are presented with changing data distributions, more rigid parameters interfere with adaption, while more flexibility causes the system to fail to retain important older information, a problem called catastrophic interference or catastrophic forgetting

  • Using the hyper-parameter selection for each network size discussed above, we simulate online dopaminergic learning on the completely disjoint MNIST dataset, compared with an identical offline setup allowed to access the MNIST dataset in randomized order allowing for class interleaving

Read more

Summary

Introduction

Artificial neural networks have enabled computing systems to successfully perform tasks previously out of reach for traditional computing, such as image and audio classification. These networks, are typically trained offline and do not update during deployed inference. Lifelong learning is the process of successfully learning from new data while retaining useful knowledge from previously encountered data that is statistically different, often with the goal of sequentially learning differing tasks while retaining the capability to perform previously learned tasks without requiring retraining on data for older tasks (see Figure 1). We take inspiration from the brain to help overcome this obstacle

Methods
Results
Discussion
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call