Visualizing and Understanding Convolutional Networks ( What are deep ConvNets learning?)


In the video “What are deep ConvNets learning?” Andrew presented a way to visualize what activations are learning from the image from the paper of Matthew D. Zeiler “Visualizing and Understanding Convolutional Networks”. I read the paper which overall stated that the visualization is done using deconvolutional network, where the inputs are the activations and the output is the reconstructed image. The layers are comprised of unpooling and transpose convolution layers. I am not getting how can transpose convolution reconstruct the image. In case anyone understood the paper and have an idea how this works can you please clarify the process or suggest relevant resources.

Thank you in advance.

1 Like

Interesting! I never read the paper linked in that section and just listened to Prof Ng’s description. Based on that, I assumed it was implemented in a very different way: by adding instrumentation to the neurons in a conv layer to see which input gives the highest activation value for a given neuron.

If you want to know more about transpose convolutions, Prof Ng covered them in Week 3 in the lead up to the U-Net architecture, which uses them in the “upsampling” phase. You can also google “understanding transpose convolutions” and you find lots of hits. Here’s one from Jason Brownlee’s website that seems worth a look.