# Development of digital hardware for a spiking image recognition network employing a novel burst-based reinforcement learning approach

## TL;DR

This paper presents a new digital hardware design for spiking neural networks that improves image recognition accuracy and speed using a novel reinforcement learning approach.

## Contribution

The paper introduces a novel RBTDP learning algorithm and efficient digital hardware design for spiking networks.

## Key findings

- The proposed spiking network achieved 98.2% accuracy on MNIST with only 6 training iterations.
- The model demonstrated 94% accuracy on CIFAR10 and 75.6% on CIFAR100.
- The design offers faster convergence and higher accuracy compared to previous methods.

## Abstract

The primary focus of accurate and cost-effective computation in machines endowed with advanced cognitive abilities is to enhance the accuracy and speed of learning in the bio-inspired spiking machine vision networks. This paper introduces a novel reinforcement burst time dependent plasticity (RBTDP) learning algorithm, implemented as a digital circuit within a spiking network that utilizes low-cost neuron circuits. This paper introduces an efficient hardware solution that employs linear substitution technique, motivated by the need for precise and fast calculations that minimize costly resource consumption in machine vision platforms, particularly those utilizing neural networks. The suggested digital designs, emphasizing the linear substitution method within digital learning and neuron blocks, are meticulously detailed to ensure maximum speed enhancement, minimal resource utilization, and high accuracy. The suggested digital learning mechanism and neuron modules were employed to build a bio-inspired spiking vision network consisting of three layers and Actor and Critic neural population, which supports unsupervised and reinforcement training, utilizing excitatory AMPA and inhibitory GABA neural interactions. Consequently, the suggested bio-inspired spiking network, utilizing the proposed RBTDP learning method, demonstrated exceptional performance in spiking vision networks. Upon training on the MNIST, CIFAR10, and CIFAR100 datasets for 6, 6, and 30 training iterations respectively, the model achieved remarkable accuracies of 98.2%, 94%, and 75.6%. These results reflect both enhanced accuracy and faster convergence compared to earlier studies.