Implementation of an unsupervised fully-connected spiking neural network on SpiNNaker for pattern classification

Machine Learning (ML) is an application of Artificial Intelligence based on the concept that machines should be given access to data and learn specific tasks by themselves, without being explicitly programmed. Artificial Neural Networks (ANNs) are a special type of ML algorithms inspired by the human brain and its functioning is similar to the way neurons work in our system. That is, just like how the neurons can learn from the past data, similarly, the ANN, in order to mimic the behavior of the biological neural networks, learn from the data and provide responses in the form of predictions or classifications. The third generation on neural networks, called Spiking Neural Networks, represents a new computation theory that mimics the brain working principles better than traditional artificial neural networks. The human brain offers an efficient signal-encoding paradigm for processing temporal real-time data, in which the information is encoded in the time of binary spike events, for instance, spikes per time unit. Spiking Neural Networks are inspired by the efficient signal-encoding paradigm for processing temporal real-time data offered by the brain and are being studied for building a new class of information processing engines that incorporate more complex models of neurons, synapses and networks. In this study we implement a fully-connected SNN model, able to perform a small pattern recognition task, incorporating unsupervised learning rules in a neuromorphic hardware architecture. The choice to use unsupervised learning is based on the interest of solving the problem in the scenario where labels may not be presented. Just the input without annotation, as for example, the picture of a dog without the label "dog". The SNN architecture is a simplified model of the one discussed by Diehl and Cook in the article "Unsupervised learning of digit recognition using spike-timing-dependent plasticity", it consists of 3 layers (input, excitatory and inhibitory) with a small amount of neurons (4 input neurons, 12 excitatory and 12 inhibitory neurons). Working with SNNs is computationally expensive as it requires solving many differential equations. To reduce the latency, to overcome the problem of Von Neumann bootleneck, the network implementation and simulation are performed on the neuromorphic computer platform SpiNNaker, using sPyNNaker, a software package for running PyNN simulations on SpiNNaker. To perform the unsupervised learning, the networks must have two fundamental characteristics: the lateral inhibition and the adaptive threshold. There were no already existing neuron models on sPyNNaker that include the characteristic of the adaptive threshold. For this reason, the most important contribution of this work is the creation of a new neuron model on SpiNNaker that includes this characteristic. This goal was also achieved thanks to the support from the SpiNNaker team at University of Manchester. This model was tested on the SNN explained before and shows the correct behaviour, the threshold value of each neuron adapts to the correspondent firing rate, fundamental for those who want to implement the unsupervised learning technique in this neuromorphic hardware.