evt_MNIST: A spike based version of traditional MNIST

📝 Abstract

Benchmarks and datasets have important role in evaluation of machine learning algorithms and neural network implementations. Traditional dataset for images such as MNIST is applied to evaluate efficiency of different training algorithms in neural networks. This demand is different in Spiking Neural Networks (SNN) as they require spiking inputs. It is widely believed, in the biological cortex the timing of spikes is irregular. Poisson distributions provide adequate descriptions of the irregularity in generating appropriate spikes. Here, we introduce a spike-based version of MNSIT (handwritten digits dataset),using Poisson distribution and show the Poissonian property of the generated streams. We introduce a new version of evt_MNIST which can be used for neural network evaluation.

💡 Analysis

Benchmarks and datasets have important role in evaluation of machine learning algorithms and neural network implementations. Traditional dataset for images such as MNIST is applied to evaluate efficiency of different training algorithms in neural networks. This demand is different in Spiking Neural Networks (SNN) as they require spiking inputs. It is widely believed, in the biological cortex the timing of spikes is irregular. Poisson distributions provide adequate descriptions of the irregularity in generating appropriate spikes. Here, we introduce a spike-based version of MNSIT (handwritten digits dataset),using Poisson distribution and show the Poissonian property of the generated streams. We introduce a new version of evt_MNIST which can be used for neural network evaluation.

📄 Content

2016 1st International Conference on New Research Achievements in Electrical and Computer Engineering

evt_MNIST: A spike based version of traditional MNIST an event-based MNIST

Mazdak Fatahi
Computer Engineering Department Razi University Kermanshah, Iran Mazdak.fatahi@gmail.com Tel Number: +98 9183592337 Mahyar Shahsavari CRIStAL laboratory
University of Lille F-59000 Lille, France Mahyar.Shahsavari@ed.univ-lille1.fr

Mahmood Ahmadi Computer Engineering Department Razi University Kermanshah, Iran m.ahmadi@razi.ac.ir Arash Ahmadi Electrical Engineering Departmentine
Razi University Kermanshah, Iran A.ahmadi@razi.ac.ir Philippe Devienne
CRIStAL laboratory
University of Lille F-59000 Lille, France Philippe.Devienne@univ-lille1.fr

Abstract— Benchmarks and datasets have important role in evaluation of machine learning algorithms and neural network implementations. Traditional dataset for images such as MNIST
is applied to evaluate efficiency of different training algorithms in neural networks. This demand is different in Spiking Neural Networks (SNN) as they require spiking inputs. It is widely believed, in the biological cortex the timing of spikes is irregular.
Poisson distributions provide adequate descriptions of the irregularity in generating appropriate spikes. Here, we introduce a spike-based version of MNSIT (handwritten digits dataset), using Poisson distribution and show the Poissonian property of the generated streams. We introduce a new version of evt_MNIST which can be used for neural network evaluation.
Keywords—Neuromorphic; spike train; Spiking Neural Networks; AER; Poisson distribution (key words) I. INTRODUCTION
There are different databases to verify the accuracy and performance of machine learning algorithms. These databases are presented in various fields. For example to evaluate neural networks recognition rate of facial expressions. A lot of training data are required to test and validation. There are some widely used standard databases that become known to compare the results to similar works. However all known databases are applicable for Artificial Neural Networks (ANN).
Consequently, this challenge is remained in usage of traditional frame-based images in existing datasets. Spiking Neural Networks use spikes for communication between the neurons. In this communication, the spikes are the same, a spike by itself will not carry any information and the number and the timing of spikes are important [1]. Therefore, if we want to use these benchmarks to evaluate our works, we need to extract proportional spike sequence from them, considering, the proper number of spikes and interval between them. There are limited public method alternatives to make a suitable dataset to check the validation of SNN. Indeed if it is not possible to use these public methods, we need to transfer the ANN dataset to Spiking one. For instance in [2], each image is presented to the network for 350ms in the form of a spike stream with Poisson-distribution. To encode the input images to the spike trains, firing rates has been transferred between 0Hz and (255/4)Hz. They repeat this process until at least five spikes have been fired during the presentation time. Authors of [3, 4], presented the input as asynchronous voltage spikes using some coding approaches. To convert the static images into events stream (spike train), spikes are generated with probability proportional to the intensity of the pixel of a given image.
According to our investigations, in many perfect SNN architectures [5-8] especially in image processing application, the stimuli is generated directly from a spiking “retina” that naturally presents data as asynchronously. In [7] to response to the visual inputs, the spikes are produced by the Dynamic Vision Sensor (DVS). Furthermore a subset of the emitted spike by the DVS (randomly), are mapped into hidden layer neurons. The Neuromorphic vision sensors are not publicly available. A spike train is a sequence of spikes generated by a single neuron. A spike train can be defined as a sequence of spike times. This chain of events can occur at regular or irregular intervals. Some evidence for neuronal variability and spike-train irregularity were reviewed in [1]. In the biological

2016 1st International Conference on New Research Achievements in Electrical and Computer Engineering

cortex, action potential timing is irregular and it is not periodic. Each irregular spike sequence can be consider as a stochastic sequence generated by a Poisson process. Here, we assume that the generation of each spike is independent of all other spikes. Respecting to this in dependency, the spike train would be completely described through a Poisson process. It is notable that, some aspects of neuronal dynamics can break the independent spike assumptio

This content is AI-processed based on ArXiv data.