576654407 by Unknown

Author:Unknown
Language: eng
Format: epub

Deep convolutional neural networks have figured out how to learn directly from image pixels. Since the number of pixels in an image can number in the millions, it looks like the number of training examples would be on the order of one followed by a million zeros. Since one followed by eighty zeros is about the number of atoms in the universe, a million zeros looks to be insurmountable. The genius behind convolutional neural networks is avoidance of this problem. A much smaller set of pixels, called a window, is used to scan the image up and down and right to left, and the neural network trains how strongly each section of the image is connected to the smaller set of roaming pixels in the window. With some additional pixel pushing, this process is repeated again and again in different layers of the convolutional neural network. The curse of dimensionality in this case, it turns out, relates to the number of pixels in the windows and not the number of pixels in the image. The number of training examples needed to train the neural network is still very high but not anywhere near the colossal number required if every image pixel were supplied as an input.

The Problem of Sensitivity

Another major problem is sensitivity. The accuracy of some deep convolutional neural networks can be derailed by changing a handful of pixels on an image. In fact, a 2019 paper titled “One Pixel Attack for Fooling Deep Neural showed that deep neural networks trained on popular databases can be fooled by changing only one pixel in an image.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.