Thanks Prof. Shen's wonderful talk! In your talk, you mentioned the Floor-ReLU network. May I ask that how do you replace ReLU by Floor, is it randomly chosen? Or you use some other methods to do the replacement?

Thank Prof Shen for your good representation, it is quite interesting! And I have some problems.1.whether your theorem is related to the architecture of DNN?2. if I use some other activation function, can i get similar result?3. and can i use this in U-net?

Thanks!

Thank you for your talk.Have you ever try smoother activation funcitons, such as RePU(max{0,x^s})

That's okay! Thank you!

I see.Thank you!

Thank Prof. Shen for the nice talk. Do you think it more efficient to enforce some properties for the dictionary/generators by introducing a loss in Deep leaning approximation? For example orthogonality.

Do you consider the complexity of the NN?

I see. Thanks a lot.

Prof. Shen, thank you for the wonderful talk. Do you think it is possible to extend the above analysis to semi-supervised or unsupervised learning? Thanks.

Can you explain more on the equation on Page38

how to construct the function g，is there any special requirement for vector a