Unsupervised Learning to Overcome Catastrophic Forgetting in Neural Networks

Abstract

Continual learning is the ability to acquire a new task or knowledge without losing any previously collected information. Achieving continual learning in artificial intelligence (AI) is currently prevented by catastrophic forgetting, where training of a new task deletes all previously learned tasks. Here, we present a new concept of a neural network capable of combining supervised convolutional learning with bio-inspired unsupervised learning. Brain-inspired concepts such as spike-timing-dependent plasticity (STDP) and neural redundancy are shown to enable continual learning and prevent catastrophic forgetting without compromising standard accuracy achievable with state-of-the-art neural networks. Unsupervised learning by STDP is demonstrated by hardware experiments with a one-layer perceptron adopting phasechange memory (PCM) synapses. Finally, we demonstrate full testing classification of Modified National Institute of Standards and Technology (MNIST) database with an accuracy of 98% and continual learning of up to 30% non-trained classes with 83% average accuracy. INDEX TERMS Catastrophic forgetting, continual learning, convolutional neural network (CNN), neuromorphic engineering, phase-change memory (PCM), spike-timing-dependent plasticity (STDP), supervised learning, unsupervised learning.

Bio-Inspired Techniques in a Fully Digital Approach for Lifelong Learning

“…with a memristive-based network with PCM synapses, in terms of area, energy consumption and testing efficiency (Munoz-Martin et al, 2019). Figure 10 shows the classification results of the continual learning accuracy for every combination of two non-trained classes of the MNIST (a) and the Fashion-MNIST (c) datasets.…”

“…However, the FPGA-based fully digital approach is not the only feasible way to perform continual learning. In particular, other works have described the possibility of implementing a hybrid supervisedunsupervised neural network using a PCM-based approach (Bianchi et al, 2019;Munoz-Martin et al, 2019). PCM devices are among the best candidates for building efficient synaptic elements, especially for their 3D stacking integration and multilevel programming capability (Kuzum et al, 2013).…”

“…Class Filters are designed to recognize only one specific class of the dataset. To determine these class-selected filters we used a fully convolutional approach as described in Munoz-Martin et al (2019). Thanks to the training procedure, the system yields a positive response on the output neuron ("1" or "VDD") when only that specific class is detected.…”

“…However, all these attempts have only partially enabled continual learning mainly because they lack an intimate link with bio-inspired techniques. In fact, bio-inspired learning algorithms like spike-timing-dependent plasticity (STDP), neuronal redundancy and spike-frequency adaptation appear as key elements for achieving continual incremental learning in various neural networks (Takiyama and Okada, 2012;Chicca et al, 2014;Bianchi et al, 2019Bianchi et al, , 2020Munoz-Martin et al, 2019).…”

Memristive and CMOS Devices for Neuromorphic Computing

“…Importantly, note that the weight change in the 1T1R synapse can be induced only via spike overlap, hence only for delays in the range −10 ms < Δt < 10 ms in this experiment [152]. Although the STDP characteristics achieved in the 1T1R RRAM synapse [151,152] display a squared shape due to binary operation of the RRAM cell instead of the exponentially decaying behavior observed in biological experiments, the plasticity of the 1T1R synapse was exploited in Although the STDP characteristics achieved in the 1T1R RRAM synapse [151,152] display a squared shape due to binary operation of the RRAM cell instead of the exponentially decaying behavior observed in biological experiments, the plasticity of the 1T1R synapse was exploited in many SNN implementations enabling neuromorphic tasks, such as unsupervised learning of space/spatiotemporal patterns [151,152,154,155], the extraction of auditory/visual patterns [156,157], pattern classification [158][159][160], and associative memory [161][162][163], in both simulation and hardware. Figure 16a shows the schematic representation of the RRAM-based SNN used in ref.…”