Learnable Gabor Filters in CNNs: Avoiding Filter Degeneration via Early Stopping Based on Similarity Metrics

The advent of Deep Learning introduced a paradigm shift in the design and implementation of machine learning models, from the feature engineering paradigm towards the feature learning one; nowadays much less effort is dedicated to the manufacture of feature extraction methods, albeit at the expense of requiring larger volumes of training data and extended training times for deep models to learn the meaningful features. Nonetheless, it has been observed that the initial layers of many image models tend to converge to some of the earlier engineered feature extractors, mainly in the form of Gabor filters and other spatial filters, thus generating a growing interest in replacing the first layers in CNNs with learnable spatial filters, and in particular, Gabor filters. In this work we investigate the problem of parameter convergence in learnable Gabor filters, discover that the filters can exhibit degradation after a few epochs of training, and propose a method based on similarity metrics between the Gabor filters to address this issue. This research can contribute to the design of more efficient training strategies of networks employing learnable spatial filters, to leverage their intrinsic advantages over the more popular non-engineered convolutional filters.