Locally connected spiking neural networks for unsupervised feature learning.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) show promise for machine learning tasks.
• Learning image features efficiently is a key challenge in AI.
• Biological neural systems offer insights into efficient information processing.

Purpose of the Study:

• To introduce a novel method for learning image features using locally connected SNNs (LC-SNNs).
• To leverage spike-timing-dependent plasticity (STDP) for unsupervised feature learning.
• To explore biologically inspired parallel processing for image classification.

Main Methods:

• Developed LC-SNNs with competing sub-networks and inhibitory interactions.
• Employed a biologically inspired n-gram classification approach for parallel patch processing.
• Utilized the BindsNET library for efficient SNN implementation.

Main Results:

• Achieved state-of-the-art classification accuracy on two image datasets.
• Demonstrated fast convergence and required fewer parameters than other unsupervised SNNs.
• Showcased robust performance degradation under synapse and neuron deletion.

Conclusions:

• LC-SNNs offer an efficient and robust approach for image feature learning.
• The proposed method mimics biological neural interactions for improved performance.
• This work highlights the potential of SNNs for advanced machine learning applications.