Summer Session: Kick Off Meeting
Deep Learning Community of Practice
For Next Meeting
We’re starting off from zero and implementing the core engine that powers deep learning: back propagation.
For our next meeting on the 22nd we have an open discussion of backpropagation scheduled. The core video for this session is this one Backpropagation 2h25 in which you’ll build a computational graph and see how values flow forward through this graph to calculate a prediction and backward to calculate gradients. I recommend at minimum watching the video and ideally implementing the code in it as well.
Note that the in the video description there are links to the notebooks that Andrej Karpathy builds during the video. I find this especially helpful for identifying errors in my code.
Please note what you find confusing – someone else probably finds it confusing too!
Materials
Here is a tiny neural network in Excel ( File). Download it and try changing the weights and biases of the layers! In practice we’ll change all the weights and biases at once, but this is nice for building intuition.
Here are the Meeting Slides.