Convolve4D: A Novelty Approach to Improve Convolutional Process

Abstract

Convolutional neural networks (ConvNet or CNN) are deep learning algorithms that can process input images, assign meaning to various aspects or objects in the image (biases and learnable weight) and recognize one image from another. The bigger kernel size will take more time to process the input.We present a novelty way to use a 4D rank tensor to improve a convolutional process. At the early stage of the Convolve4D development, the edge detection with 3×3 kernel and The Laplacian of Gaussian (LoG) with 5×5 kernel size was used to demonstrate the convolutional process improvement. The Convolve4D needs more elaboration to be used into a CNN algorithm. The advantage of convolve4D is only need 9 loops to calculate 81 outputs, whereas convolve2D need 9 × 9 × 3 × 1 × 7 × 7 = 11.907 loops. The result is 18.5% shorter when using a 5×5 kernel; it reduces from 0.54 seconds to 0.44 seconds for the edge detection convolution process.