Building efficient deep Hebbian networks for image classification tasks

Multi-layer models of sparse coding (deep dictionary learning) and dimensionality reduction (PCANet) have shown promise as unsupervised learning models for image classification tasks. However, the pure implementations of these models have limited generalisation capabilities and high computational cost. This work introduces the Deep Hebbian Network (DHN), which combines the advantages of sparse coding, dimensionality reduction, and convolutional neural networks for learning features from images. Unlike in other deep neural networks, in this model, both the learning rules and neural architectures are derived from cost-function minimizations. Moreover, the DHN model can be trained online due to its Hebbian components. Different configurations of the DHN have been tested on scene and image classification tasks. Experiments show that the DHN model can automatically discover highly discriminative features directly from image pixels without using any data augmentation or semi-labeling.