BioLCNet: Reward-Modulated Locally Connected Spiking Neural Networks

Reinforcement Learning in Deep Spiking Neural Networks with Eligibility Traces and Modifying the Threshold Parameter

“…The highest test accuracy of our proposed model, obtained by optimizing the hyperparameters including the adaptive LIF neurons threshold voltage increase, membrane time constant, and threshold voltage of simple and adaptive LIF neurons relative to their resting and reset voltages, was 87.84% on the MNIST dataset. This accuracy represents an improvement compared to the previous similar work in [21]. Our proposed model, indicated by the red box in the table, shows a comparison with its spike-based counterpart.…”

“…The performance of the proposed approach was evaluated on MNIST, by working on the threshold voltage increase of adaptive LIF neurons, as well as the membrane time constant and voltage threshold of simple and adaptive LIF neurons in comparison to their resting voltage and reset values in the network, and a recognition accuracy of 87.84% was achieved on the test set. Simulation results demonstrate the improved accuracy of this network compared to a similar SNN in [21]. Also, the efficient mapping of the model to the neuromorphic hardware is feasible because adjusting the voltage reduces the convergence time.…”

“…These two proposed learning mechanisms can solve the XOR problem with both rate and temporal encoding schemes and learn the target pattern efficiently. In the article [21], a two-layer locally connected SNN called BioLCNet is proposed, with architecture and learning methods that are biologically plausible. The input's visual features are extracted using local filters and then, they are learned using STDP.…”

“…Both models, due to their use of rate coding, locally connected architecture, and reinforcement learning, are biologically plausible and exhibit high convergence speed. In [21], Ghaemi et al proposed a two-layer locally connected SNN called BioLCNet, with a biologically plausible architecture and learning method similar to our utilized model. A locally connected layer is trained using STDP learning to learn the visual features of the input images.…”

“…The effectiveness and credibility of the introduced approach are evaluated on the MNIST dataset [37,38]. Simulation results confirm that the recognition accuracy of this model has improved significantly compared to the previous work in [21]. This article is structured as follows: Previous related works are presented in section 2.…”