loading…
A4.3.9 HL · deep learning
A convolutional network sees images through small filters that slide across them, each one lighting up where it finds its pattern, like an edge. Watch a filter build a feature map, then pool it down. Stack this idea many times and you have a CNN. After convolution and pooling have distilled the image into features, a fully connected layer weighs them all together to make the final call: which class the image belongs to.
input image
feature map
The filter (below) slides over every patch of the image; each patch becomes one cell of the feature map, bright where the filter's pattern is found. Pooling then keeps the strongest response in each block, shrinking the map but keeping what matters.