A Convolutional Network, sometimes called “ConvNet”, “Convolutional Neural Network”, or “CNN”, is a class of trainable models for processing and classifying images, video, audio and other signals that can be organized as 1D, 2D or 3D arrays.
ConvNets are composed of multiple stages processing, each of which is composed of three layers: a (convolutional) filter bank, a non-linearity, and a pooling operation. The filters are local. The architecture is somewhat reminiscent of the feed-forward portion of the ventral pathway of the mammalian visual cortex. The filter outputs are akin to Hubel and Wiesel's “simple cells” and the pooling units are similar to their “complex cells”.
The distinct advantage of ConvNets over other recognition architectures is that all the filters in all the layers are trained, often in purely supervised mode, sometimes in unsupervised mode. This allows ConvNets to automatically learn low-level and mid-level features that are tuned to the task at hand.
ConvNets have been commercially deployed in many applications, including document recognition (reading checks), handwriting recognition for pen-based computers, medical image analysis, video surveillance, recognizing the the gender and estimating the age of vending machine customers, tracking customer behavior in supermarkets, controlling pan-tilt cameras for video conference, detecting faces and license plates in web images for privacy protection,….
- Publications and research papers on ConvNets
- Tutorials, introductory material on ConvNets here and around the Web
- Talks about ConvNets, with slides and videos
- Demos of ConvNets applications
- Software and library to train and run ConvNets on CPUs, GPUs, etc.
- Applications of ConvNets, including commercially deployed ones
- Hardware designed specifically to run ConvNets and similar models
- History of ConvNets and related models
- Press articles about ConvNets and their applications
- Links to various on-line material about ConvNets
- People and Labs that work on ConvNets
An unusually large number of papers at ICML 2011 talk about ConvNets.