From the course: Neural Networks and Convolutional Neural Networks Essential Training
Unlock the full course today
Join today to access over 22,600 courses taught by industry experts or purchase this course individually.
Convolutions
From the course: Neural Networks and Convolutional Neural Networks Essential Training
Convolutions
- [Instructor] Fully connected neural networks typically don't work well on images. This is because if each pixel is an input, then as we add more layers the amount of parameters increases exponentially. Let's say you had a 32 by 32 image. That's 32 wide, 32 high, and with three color channels. A single, fully connected neuron in the first hidden layer of a regular neural network would have 32 multiplied by 32 multiplied by three, and that's 3,072 weights. A color image, which isn't significantly larger, so something that is 200 wide by 200 high, with three color channels, so a fully connected neuron in the first hidden layer of a regular neural network would have 200 multiplied by 200 multiplied by three, which is 120,000 weights. The other challenge is that the number of parameters this large can quickly lead to over-fitting. One work around is that we can use smaller images, but clearly we will lose information. What we have not taken into account is that what makes one image…
Practice while you learn with exercise files
Download the files the instructor uses to teach the course. Follow along and learn by watching, listening and practicing.