An Online Unsupervised Structural Plasticity Algorithm for Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Traditional neural networks often require complex hardware and extensive training data.
• Developing efficient and hardware-friendly neural network architectures is crucial for real-world applications.

Purpose of the Study:

• To propose a novel winner-take-all (WTA) architecture with nonlinear dendrites and binary synapses.
• To introduce an online unsupervised structural plasticity rule for training the WTA network.
• To evaluate the network's performance in multi-class classification tasks using spike time inputs.

Main Methods:

• Implemented a novel WTA architecture utilizing neurons with nonlinear dendrites.
• Employed an unsupervised structural plasticity learning rule inspired by spike-timing-dependent plasticity.
• Trained the network using binary synapses for hardware implementation feasibility.
• Tested the network on two-, four-, and six-class classification of random Poisson spike time inputs.

Main Results:

• The inhibitory time constant of the WTA network allows tuning of specificity and sensitivity.
• Classifying patterns subdivided into 5 or 10 subpatterns achieved 100% successful trials.
• Classifying patterns without subdivision resulted in 92%, 88%, and 82% success for two-, four-, and six-class problems, respectively.
• Unsubdivided patterns demonstrated higher resilience to timing jitter compared to subdivided patterns.

Conclusions:

• The proposed WTA architecture and learning rule offer an efficient approach for spike-based pattern classification.
• Pattern subdivision enhances classification accuracy at the cost of jitter resilience.
• The inhibitory time constant is a key parameter for balancing network sensitivity and specificity.