🧠 Neural Networks -

Deconvolutional

Deconvolutional neural networks (deconvnet) are used to visualize how convnets work.

I more technical terms, deconvnets provide a way to map layer activations back to the input.

• in a sense, they reconstruct images
• main problem was to reverse max pool operation
• max pool: 2x2 square, find max value, take max value and propagate it forward
• to reverse max pooling, keep track of the 2x2 squares. Save which of the squares was chosen.
• "This value to the square, set 0 to the rest."

Deconvnets were originally an unsupervised learning method for finding alternative representation. They work as alternative to wavelets, curvlets, framelets, shapelets, shearlets.