Deep face recognition training on a mobile device

Question

I want to create a mobile face recognition application, and I want to do this with regular Deep Learning (convolutional network). I will train him using my computer and use the prepared data in a mobile application.

My question is: can I get very fast calculation in regulat phone like iPhone? I need it to be very fast and in less than 1 second you can detect a face in the video. Is this possible on a mobile device? or does this kind of task require more powerful equipment?

I know that the training phase must be on a powerful computer, but I mean the production phase on the mobile device.

for example, if I put my phone outside, can He find that all people encounter the same deep network during the training phase?

Answer 1

Yes, it is possible, but not with standard CNN architectures, some changes are needed:

• One approach is CNN with binary weights, so CNN can only be evaluated using bitwise operations. There are many publications about this, for example this , this or this . I saw a YOLO implementation with binary scales working in real time on the iPhone, so this is definitely possible.
• , , 5000 , , , . .
• - . , L-CNN , cuDNN.

- 1- .