A fundamental feature of learning in animals is the "ability to forget" thatallows an organism to perceive, model and make decisions from disparate streamsof information and adapt to changing environments. Against this backdrop, wepresent a novel unsupervised learning mechanism ASP (Adaptive SynapticPlasticity) for improved recognition with Spiking Neural Networks (SNNs) forreal time on-line learning in a dynamic environment. We incorporate an adaptiveweight decay mechanism with the traditional Spike Timing Dependent Plasticity(STDP) learning to model adaptivity in SNNs. The leak rate of the synapticweights is modulated based on the temporal correlation between the spikingpatterns of the pre- and post-synaptic neurons. This mechanism helps in gradualforgetting of insignificant data while retaining significant, yet old,information. ASP, thus, maintains a balance between forgetting and immediatelearning to construct a stable-plastic self-adaptive SNN for continuouslychanging inputs. We demonstrate that the proposed learning methodologyaddresses catastrophic forgetting while yielding significantly improvedaccuracy over the conventional STDP learning method for digit recognitionapplications. Additionally, we observe that the proposed learning modelautomatically encodes selective attention towards relevant features in theinput data while eliminating the influence of background noise (or denoising)further improving the robustness of the ASP learning.
展开▼
