Hardware friendly probabilistic spiking neural network with long-term and short-term plasticity

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) offer bio-realistic computation.
• Developing efficient and accurate SNNs for complex tasks remains a challenge.
• Plasticity mechanisms are crucial for learning in neural systems.

Purpose of the Study:

• To propose a novel probabilistic spiking neural network (PSNN) with unimodal weight distribution.
• To incorporate both long- and short-term plasticity into the PSNN.
• To evaluate the PSNN's performance, convergence, and hardware-friendliness.

Main Methods:

• The PSNN algorithm was derived using arithmetic gradient descent and bio-inspired methods.
• The PSNN was benchmarked on the Iris and Wisconsin breast cancer (WBC) datasets.
• Performance was further evaluated using odor data from a self-developed electronic nose (e-nose).

Main Results:

• The PSNN achieved high accuracy: 96.7% for Iris and 97.2% for WBC datasets within 40 epochs.
• For e-nose data, the PSNN showed comparable classification accuracy (1.3% less than KNN) with at least 40% less memory.
• The PSNN requires only nine-bit weight resolution and fewer neurons, demonstrating hardware efficiency.

Conclusions:

• The proposed PSNN exhibits fast convergence and high accuracy on diverse datasets.
• The PSNN is hardware-friendly, requiring low weight resolution and reduced neuron count for VLSI implementation.
• The PSNN's robustness to noise and parameter variations makes it suitable for both software and hardware applications.