Hebbian semi-supervised learning in a sample efficiency setting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep Convolutional Neural Networks (DCNNs) often require large labeled datasets for effective training.
• Sample efficiency remains a significant challenge in DCNN training, particularly in real-world applications with limited labeled data.

Purpose of the Study:

• To develop and evaluate a novel semi-supervised training strategy for DCNNs to address sample efficiency issues.
• To leverage unsupervised Hebbian learning for pre-training internal network layers, reducing reliance on labeled data.

Main Methods:

• A hybrid training approach combining unsupervised Hebbian learning for internal layers and supervised Stochastic Gradient Descent (SGD) for the classification layer.
• Experiments conducted on various object recognition datasets under different sample efficiency conditions.
• Comparison against end-to-end supervised backpropagation and Variational Auto-Encoder (VAE)-based semi-supervised methods.

Main Results:

• The proposed semi-supervised approach demonstrated superior performance compared to other methods in low-sample regimes.
• Significant improvements in object recognition accuracy were observed when labeled data was scarce.
• Hebbian learning effectively pre-trained DCNN internal layers without requiring labeled examples.

Conclusions:

• The novel semi-supervised strategy significantly enhances sample efficiency in DCNNs.
• This method offers a promising solution for training DCNNs in data-limited scenarios.
• Combining unsupervised Hebbian learning with supervised SGD is effective for improving DCNN performance with limited labeled data.