Difference between revisions of "10-601 Deep Learning 1"

Revision as of 14:17, 11 April 2016

This area is moving very fast and the textbooks are not up-to-date. Some recommended readings:

Neural Networks and Deep Learning An online book by Michael Nielsen, pitched at an appropriate level for 10-601, which has a bunch of exercises and on-line sample programs in Python.

For more detail, look at the MIT Press book (in preparation) from Bengio - it's very complete but also fairly technical.

The underlying reasons deep networks are hard to train
- Exploding/vanishing gradients
- Saturation
The importance of key recent advances in neural networks:
- Matrix operations and GPU training
- ReLU, cross-entropy, softmax
Convolutional networks
- 2-d convolution
- How to construct a convolution layer
- Architecture of CNN: convolution/downsampling pairs

@@ Line 10: / Line 10: @@
 This area is moving very fast and the textbooks are not up-to-date.  Some recommended readings:
 * [http://neuralnetworksanddeeplearning.com/index.html  Neural Networks and Deep Learning] An online book by Michael Nielsen, pitched at an appropriate level for 10-601, which has a bunch of exercises and on-line sample programs in Python.
-* [http://cs231n.github.io/  Stanford CS class CS231n: Convolutional Neural Networks for Visual Recognition] has nice on-line notes.
-I also used some on-line visualizations in the materials for the lecture, especially the part on ConvNets.
-* [https://en.wikipedia.org/wiki/Convolution the Wikipedia page for convolutions] has nice animations of 1-D convolutions.
-* [http://matlabtricks.com/post-5/3x3-convolution-kernels-with-online-demo  On-line demo of 2-D convolutions] for image processing.
-* [https://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist.html  There's an on-line demo of CNNs] which are trained in your browser (!)
-* [http://scs.ryerson.ca/~aharley/vis/conv/   3D visualization of a trained net.]
 For more detail, look at the [http://www.deeplearningbook.org/ MIT Press book] (in preparation) from Bengio - it's very complete but also fairly technical.