Unsupervised Digit Recognition through Biologically Inspired Spike-Timing Dependent Algorithm

Spiking Neural Networks (SNNs) are emerging as a promising energy-efficient, biologically inspired alternative to traditional Artificial Neural Networks (ANNs). Although the theoretical foundations of SNNs date back to several decades, many practical aspects of their implementation and training still remain underdeveloped, making them still not fully competitive with more established, ANN-based systems. SNNs take advantage of neuromorphic hardware to achieve low power consumption and are capable of asynchronous processing and event-driven operations. They are ideal candidates for edge computing, mobile devices, and robotics, where limited battery capacity and computational resources represent the most critical constraints. However, achieving performance comparable to that of standard ANNs is challenging, primarily due to the inability to employ gradient-based optimization methods, due to the non-differentiable nature of the spikes. This work builds upon my own previous research activity carried out for the Data Science Master’s program, with more in depth theoretical discussions regarding the neuron models and methodologies, together with novel simulations, experiments and discussions. The core objective of this Thesis is the implementation of a SNN capable of classifying handwritten digit images using an unsupervised training algorithm based upon the Spike-Timing-Dependent Plasticity (STDP) learning rule. This biologically inspired approach relies on temporal correlations between pre- and postsynaptic spikes to update the synaptic weights. Therefore, differently from backpropagation, no derivatives or backward passes are needed, making this approach suitable for event-driven SNN training. The model incorporates advanced neuron dynamics, including conductance-based synapses, lateral inhibition, and adaptive thresholds, thereby enhancing both classification accuracy and biological realism. Key contributions include a custom synaptic leaky integrate-and-fire (LIF) neuron model that accommodates both excitatory and inhibitory synapses. An exponential-offset STDP rule has been employed to tailor the unsupervised learning process specifically to reduce the computational burden for SNN-based tasks. Additionally, an original confidence function was developed to optimize label assignments after training, ensuring higher reliability and accuracy of output classes. Experimental results underscore competitive classification performance, indicating the potential of unsupervised SNNs for real-world applications.