A novel technique for optimizing the filter size of CNNs without backpropagation

Image filters play a crucial role in the performance of convolutional neural networks (CNNs). Yet, the optimisation of those filters tends to focus solely on optimising their weights. As a result, the machine learning practitioner either uses the standard 3×3 filter size or has to select the filter size empirically, as the optimal filter size depends on the application at hand. There has been a lack of serious attempts to address this issue in CNNs. To this end, we propose a novel technique to learn the filter size without depending on either gradient descent or backpropagation. We compare our technique with the only serious attempt existing and show that not only our method performs better, but also converges to an optimal solution at a much smaller number of training iterations.