Scientists from Bar-Ilan University have discovered how machine learning successfully classifies images, revealing that each filter in a deep learning architecture fine-tunes the recognition and acknowledges of image clusters through layers.Current AI architectures can successfully carry out image classification jobs, taking on human capabilities. What is the mechanism that makes device knowing so successful?Image category is a complicated task that deep knowing architectures perform successfully. Those deep architectures are normally consisted of many layers, with each layer consisting of numerous filters. The typical understanding is that as the image progresses through the layers more enhanced functions, and features of features, of the image are revealed. Those functions and functions of features are not quantifiable, and therefore how machine learning works remains a puzzle.In a post recently published in Scientific Reports, scientists from Bar-Ilan University reveal the system underlying successful machine learning, which enables it to carry out category tasks with definite success. “Each filter basically recognizes a small cluster of images and as the layers advance the recognition is honed. We discovered a way to quantitatively measure the efficiency of a single filter,” said Prof. Ido Kanter, of Bar-Ilans Department of Physics and Gonda (Goldschmied) Multidisciplinary Brain Research Center, who led the research.A video explaining the research study. Credit: Prof. Ido Kanter, Bar-Ilan University” This discovery can pave the path to much better understanding how AI works,” said PhD trainee Yuval Meir, one of the key factors to the work, adding, “This can improve the latency, memory usage, and complexity of the architecture without lowering general accuracy.” While AI has been at the forefront of recent technological progress, understanding how such machines actually work can break the ice for much more advanced AI.Reference: “Towards a universal system for successful deep knowing” by Yuval Meir, Yarden Tzach, Shiri Hodassman, Ofek Tevet and Ido Kanter, 11 March 2024, Scientific Reports.DOI: 10.1038/ s41598-024-56609-x.
