Keywords
Spiking Neural Network, Spike Timing Dependent Plasticity, STDP, Machine Learning
Presentation Type
Poster
Research Abstract
Artificial neural networks, that try to mimic the brain, are a very active area of research today. Such networks can potentially solve difficult problems such as image recognition, video analytics, lot more energy efficiently than when implemented in standard von-Neumann computing machines. New algorithms for neural computing with high bio-fidelity are being developed today to solve hard machine learning problems. In this work, we used a spiking network model, and implemented a self-learning technique using a Spike Timing Dependent Plasticity (STDP) algorithm, that closely mimics the neural activity of the brain. The basic STDP algorithm modulates the synaptic weights interconnecting the neurons based on pairs of pre- and post-synaptic spikes. This ignores the timing information embedded in the frequency of the post-synaptic spikes. We calculated the average of the membrane potential of each column of neurons to give an idea of how it behaved and spiked for the particular output neuron for a particular image in the past .The update of the weights or the synapses are done on the basis of the frequency obtained. The resultant synaptic updates are less frequent and made wisely making the learning process better. With the present algorithm, we are able to achieve an accuracy of 79% for classifying images from the MNIST data set for a network of 400 output neurons. So the model was able to identify 79% of the total images correctly which is greater than the original STDP signifying that slow and sensible updates are definitely having a better impact on the learning process.
Session Track
Nanotechnology
Recommended Citation
Sourjya Roy, Gopalakrishnan Srinivasan, and Vijay Raghunathan,
"Brain Inspired Enhanced Learning Mechanism Based on Spike Timing Dependent Plasticity (STDP) for Efficient Pattern Recognition in Spiking Neural Networks"
(August 4, 2016).
The Summer Undergraduate Research Fellowship (SURF) Symposium.
Paper 144.
https://docs.lib.purdue.edu/surf/2016/presentations/144
Included in
Brain Inspired Enhanced Learning Mechanism Based on Spike Timing Dependent Plasticity (STDP) for Efficient Pattern Recognition in Spiking Neural Networks
Artificial neural networks, that try to mimic the brain, are a very active area of research today. Such networks can potentially solve difficult problems such as image recognition, video analytics, lot more energy efficiently than when implemented in standard von-Neumann computing machines. New algorithms for neural computing with high bio-fidelity are being developed today to solve hard machine learning problems. In this work, we used a spiking network model, and implemented a self-learning technique using a Spike Timing Dependent Plasticity (STDP) algorithm, that closely mimics the neural activity of the brain. The basic STDP algorithm modulates the synaptic weights interconnecting the neurons based on pairs of pre- and post-synaptic spikes. This ignores the timing information embedded in the frequency of the post-synaptic spikes. We calculated the average of the membrane potential of each column of neurons to give an idea of how it behaved and spiked for the particular output neuron for a particular image in the past .The update of the weights or the synapses are done on the basis of the frequency obtained. The resultant synaptic updates are less frequent and made wisely making the learning process better. With the present algorithm, we are able to achieve an accuracy of 79% for classifying images from the MNIST data set for a network of 400 output neurons. So the model was able to identify 79% of the total images correctly which is greater than the original STDP signifying that slow and sensible updates are definitely having a better impact on the learning process.