Supervised Learning without Backpropagation using Spike-Timing-Dependent Plasticity for Image Recognition

Abstract

This study introduces a novel supervised learning approach for spiking neural networks that does not rely on traditional backpropagation. Instead, it employs spike-timing-dependent plasticity (STDP) within a supervised framework for image recognition tasks. The effectiveness of this method is demonstrated using the MNIST dataset. The model achieves approximately 40\% learning accuracy with just 10 training stimuli, where each category is exposed to the model only once during training (one-shot learning). With larger training samples, the accuracy increases up to 87\%, maintaining negligible ambiguity. Notably, with only 10 hidden neurons, the model reaches 89\% accuracy with around 10\% ambiguity. This proposed method offers a robust and efficient alternative to traditional backpropagation-based supervised learning techniques.

Figures (6)

• Figure 1: (a) A single hidden layer unit comprising ten neurons, each dedicated to a specific digit between 0 and 9. Neurons in a unit share identical spatial coordinates and are fully connected with inhibitory synapses during validation and testing. (b) An example of a network with two hidden layers. The input layer encodes image information through Poisson-distributed spike trains, with firing rates proportional to pixel intensities. Full connectivity exists between the input layer and the first hidden layer, as well as between the first and second hidden layers. During training, the image label is provided as supervisory information to each excitatory neuron.
• Figure 2: The influence of training sample size per worker on (left) overall test accuracy and (right) test accuracy with ambiguity for networks with different hidden layer sizes. Each color represents a specific number of excitatory neurons in the hidden layer. The size of test sample is $10^4$ and the bars associated with data points represent statistical uncertainties.
• Figure 3: Spike count distribution of each neuron group for correctly identified test samples across different input labels (highlighted in blue). The network has a single hidden layer with 250 neurons (equivalent to 25 workers) employing parallel training with diversity. Each worker processed 10 training stimuli. The size of test sample is $10^4$.
• Figure 4: Spike count distribution of each neuron group for correctly identified test samples across different input labels (highlighted in blue). The network has a single hidden layer with 250 neurons (equivalent to 25 workers) employing parallel training with diversity. Each worker processed $10^4$ training samples. The size of test sample is $10^4$.
• Figure 5: The influence of input labels on (left) overall test accuracy and (right) test accuracy with ambiguity for networks with different hidden layer sizes. Each color represents a specific number of excitatory neurons in the hidden layer. The size of test sample is $3\times10^4$ and the bars associated with data points represent statistical uncertainties.