Feature Extraction using Spiking Convolutional Neural Networks

Abstract

Spiking neural networks are biologically plausible counterparts of the artificial neural networks, artificial neural networks are usually trained with stochastic gradient descent and spiking neural networks are trained with spike timing dependant plasticity. Training deep convolutional neural networks is a memory and power intensive job. Spiking networks could potentially help in reducing the power usage. There is a large pool of tools for one to chose to train artificial neural networks of any size, on the other hand all the available tools to simulate spiking neural networks are geared towards computational neuroscience applications and they are not suitable for real life applications. In this work we focus on implementing a spiking CNN using Tensorflow to examine behaviour of the network and study catastrophic forgetting in the spiking CNN and weight initialization problem in R-STDP using MNIST data set. We also report classification accuracies that are achieved using N-MNIST and MNIST data sets. CCS CONCEPTS • Computing methodologies → Machine learning; Machine learning approaches; Bio-inspired approaches;

A Deep Unsupervised Feature Learning Spiking Neural Network with Binarized Classification Layers for EMNIST Classification using SpykeFlow

“…To monitor the weight updates (synapse changes) in the spiking network, the software provides the capability to monitor spike activity, weight evolution (updates), feature extraction (spikes per map per label), and synapse convergence, etc. This software tool was used here and in [34] [35]. Similar to our work, Mozafari et al released the software tool SPYKETORCH in [43] which is based on the PYTORCH [44] deep learning tool.…”

“…Such behavior is expected as the Conv2 layer tries to learn features that are more complex than that of the features in layer L2 (Conv1). For a Conv2 layer trained with spikes collected without lateral inhibition in layer L3 (Pool 1) an early stopping mechanism based on the temporal differences ca be used (see [35]) and an example plot is shown in Figure 24). The net4.spike_statistics() method was used to generate plots shown in Figure 21 and Figure 22.…”

Regression with Deep Learning for Sensor Performance Optimization

“…Deep learning has been used in plethora of applications like autonomous driving, cancer prediction, low power object recognition etc [2] [3] [4]. In particular, neural networks as a regression tool have been used in applications like, time series learning [5], stock prediction [6], pose estimation in computer vision [7], cost predictions [8] etc.…”

Continuous Learning in a Single-Incremental-Task Scenario with Spike Features

“…Typically, Spiking Neural Networks (SNNs) are trained using an unsupervised algorithm called Spike Timing Dependant Plasticity (STDP) [5]. Spike features extracted from latency encoded convolutional variants of SNNs have been used with an SVM [5] and a linear neural network classifier [14] to achieve classification accuracies in excess of 98.5%. However, SNNs tend to achieve lower classification accuracies when compared to Artificial Neural Networks (ANNs) [11].…”

“…In our previous work [14] we used the MNIST dataset split into two disjoint tasks to show that features extracted from a spiking convolutional network (SCN) demonstrated more immunity to catastrophic forgetting compared to their ANN counterparts. In [14], using early stopping, the first five output neurons were trained to classify the digits {0, 1, 2, 3, 4} and then the remaining five output neurons were trained to classify the digits {5, 6, 7, 8, 9}. The network was then tested on the complete test dataset (digits 0-9) and achieved a 93% accuracy on this test data.…”