Adaptive Kernel Selection for Convolutional Neural Network

SAHAN AHMAD, Aminul Islam

Keywords: Adaptive, Kernel, CNN, Computer Vision

Abstract: Convolutional Neural Networks (CNN) are used for various applications ranging from computer vision to natural language processing. A kernel, known as the matrix of weights, performs a convolution operation on input data. In general, the optimizer updates the weights of the kernel. Recent research suggests that applying a deterministic kernel after convolution operation can help a CNN to gain better generalization. However, how to compute the weights of the deterministic kernel is still a field of active research. In this work, we derived a lemma that shows of the representativeness of an adaptive deterministic kernel. We construct an adaptive deterministic kernel based on the Gaussian distribution of convoluted data. We generate many set of kernels by shifting weights to different positions of the initially created Gaussian kernel. We notice a pattern of weight distribution in deterministic kernels constructed from the Gaussian distribution of convoluted data. Using the derived lemma, it is possible to sort out a set of kernels (from many set of kernels) that tends to gain better performance in CNN for image classification task. The main object of this research work\footnote{https://github.com/PaperUnderReviewDeepLearning/KernelSet} is to identify these patterns and recommend the better set of kernels to gain performance.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)