4 ms·
You can't train fully connected deep models with backprop, or at least not easily or well. An alternative solution to this problem is spatial weight pooling (Ya
by iskander 14y ago
You can't train fully connected deep models with backprop, or at least not easily or well. An alternative solution to this problem is spatial weight pooling (Yann's convolutional networks) which play well with SGD.
- mark_l_watson 14y agoThat is correct. The problem is that the gradients get smaller and smaller as you back propagate back towards the input layer. So learning on the front part of the net is slow. Hinton has a lot of good material about htis in his Coursera lectures.
- wookietrader 14y agoYes you can. Check out the publications by Ciresan on MNIST, have a look at Hinton's dropout paper or at the Kaggle competition that used deep nets. Or try it yourself and spend a descent amount of time on hyper parameter tuning. :)
- iskander 14y agoWhich of Ciresan's projects are you referring to? Everything I've seen by him uses convolutional layers of some sort.