• Lecture slides
• Lecture on deep learning (russian) - https://www.youtube.com/watch?v=8008XQzoUEs
• Seminar on theano (russian) - https://yadi.sk/i/54STsEBVpubkn
• Intro to neural nets and backprop (english) - https://www.youtube.com/watch?v=uXt8qF2Zzfo
• Intro to convnets (english) - https://www.youtube.com/watch?v=FmpDIaiMIeA
• Theano tutorial from Lamblin (english) - https://www.youtube.com/watch?v=OU8I1oJ9HhI

• Karpathy's course on deep learning (english) - http://cs231n.github.io/
• Nuts and Bolts of deep learning by Andrew Ng (english) - https://www.youtube.com/watch?v=F1ka6a13S9I
• Deep learning demystified - https://www.youtube.com/watch?v=Q9Z20HCPnww
• Karpathy's lecture on deep learning for computer vision - https://www.youtube.com/watch?v=u6aEYuemt0M
• Our humble DL course: HSE'autumn16, Skoltech/YSDA'spring16 courses on deep learning (english).
• Srsly, just google "deep learning %s"%s for s in what_you_want_to_know.

If you are already familiar with lasagne or you are super-good with tensorflow/pytorch/similar, pick one of the alternative options. Otherwise we highly recommend the first one as we'll need convolutional networks soon enough.

• recommended go to Seminar3.5-*-mnist.ipynb and follow the instructions (ends with lasagne MNIST classifier)

• alternative task go to Seminar3.5-approx-q-learning.ipynb and follow the instructions (ends with simple NN for q-learning)

• [alternative frameworks] The equivalent of recommended track would be

• [tensorflow] learning through this google course from start till "Convolutional neural networks" (inclusive).

• [manual/other] surviving past assignment2 of cs231

• [alternative task and frameworks] Implement the simple q-learning network that solves CartPole-v0. You're not required to implement experience replay / any advanced stuff, just set sgd learning rate to a small enough number (10^-4) and pray that trains smoothly.

Here's a convenient translation to tensorflow: notebook

Agent can maintain low reward for long enough, but it should at least show some progress by the end of the default loop.