Comparison of Hebbian Learning and Backpropagation for Image Classification in Convolutional Neural Networks

Abstract: Current commonly used image recognition convolutional neural networks share some similarities with the human brain. However, the differences are many and the well established backpropagation learning algorithm is not biologically plausible. Hebbian learning is an algorithm that could minimize these differences and potentially provide image recognition networks with brain-like advantageous features. Here we explore the differences between Hebbian learning and backpropagation, both regarding accuracy and representations of data in hidden layers. Overall Hebbian networks performed considerably worse than conventional backpropagation-trained networks. Experiments with incomplete training data and distorted test data resulted in smaller but still visible performance differences. However, the convolutional filter structure of Hebbian networks proved to be simpler and more interpretable than the backpropagation equivalent. We hypothesize that improvements to increase scaling capabilities of Hebbian networks could render them a powerful alternative for image classification networks with more brain-like behavior.