This card presents a simplified representation of the inner structures of what is known as a Convolutional Neural Network, a prominent expression of what ‘Artificial Intelligence’ means today. In recent years this has become the leading architecture for image classification and object detection. This network processes images in a similar way to our visual cortex, which makes sense of our environment by combining the output of a series of progressively more complex filters. We see here a simplified representation of the filters that would typically appear in such a network, including lines and curves at various angles.
Such networks are often very large and take a long time to train from scratch. To accelerate this process, we can use a technique called transfer learning, which exploits the fact that no matter the domain, the building blocks of an image (lines, curves, etc.) are relatively similar.
The transfer of knowledge across unrelated domains is one of the most powerful creativity hacks. By identifying a set of common fundamentals which exist in any context, it is possible to efficiently move from one discipline to another.
Watch an explanation of the inner layers of a Convolutional Neural Network.
Return home.