Unsupervised Feature Learning With Winner-Takes-All Based STDP.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Computer Vision

Background:

• Unsupervised feature learning is crucial for image analysis.
• Biological learning rules offer novel approaches for artificial intelligence.
• Spike-Timing-Dependent-Plasticity (STDP) is a key mechanism in neural plasticity.

Purpose of the Study:

• To develop a novel unsupervised feature learning strategy for image applications.
• To leverage the biological STDP learning rule for efficient image representation.
• To demonstrate the effectiveness of the proposed method on various image datasets.

Main Methods:

• Equivalence between rank order coding Leaky-Integrate-and-Fire neurons and ReLU artificial neurons for non-temporal data.
• Application of rank-order coding for single feed-forward pass network simulation on GPU hardware.
• Introduction of a binary STDP learning rule compatible with batch image training.
• Implementation of Winner-Takes-All (WTA) framework and feature-wise normalization for training stabilization.
• Training of multi-layer architectures of convolutional sparse features.

Main Results:

• Demonstrated efficient feature extraction from MNIST, ETH80, CIFAR-10, and STL-10 datasets.
• Showcased the relevance of learned features for image classification tasks.
• Achieved competitive performance compared to state-of-the-art unsupervised learning methods.

Conclusions:

• The proposed STDP-inspired method offers an effective strategy for unsupervised feature learning in images.
• The method enables efficient training of deep convolutional sparse feature architectures.
• The learned features are highly relevant for downstream classification tasks.