loading…
A4.3.9 HL · deep learning
A convolutional neural network (CNN) uses filters to detect features such as edges in an image. Watch a filter build a feature map, then use max pooling to reduce its size. The filters here are fixed. CNN filters are normally learned during training. Successive layers can detect more complex features. For image classification, fully connected layers can use these features to predict a class.
input image
feature map
The filter slides across the image. Each patch gives one value in the feature map. Blue shows positive values and orange shows negative values, with stronger colour for greater magnitude. Max pooling keeps the highest numerical value in each block. ReLU is an activation function that sets negative values to zero. This activity keeps the filter's signed outputs, so pooling can discard a strong negative response.