diracdeltafunk / babby_nets

Code for a talk at UCLA, given 11-18-24

Babby Nets

This is the code I used to prepare for & present the talk "Write Your Own Neural Net" in the UCLA GSO Math Seminar, January 18 2024.

The files

• nnet0.ipynb
• nnet1.ipynb
• nnet2.ipynb
• nnet3.ipynb

were written ahead of time, and the file welldoitlive.ipynb was a version of nnet0.ipynb that was live-coded during the talk.

nnet0.ipynb

This implements a very small network (2 inputs, 1 hidden layer with 2 neurons, 1 output) which gets trained via very brutal raw gradient descent. Currently it's set up to approximate the function (x,y) ↦ xy.

nnet1.ipynb

This implements an arbitrary dense neural net, again trained with gradient descent (this time with actual backpropogation).

nnet2.ipynb

This introduces tensorflow, and trains a small network to approximate (x,y) ↦ xy.

nnet3.ipynb

This implements an MNIST classifier using Tensorflow. After training, you can use the make_prediction() function to classify the contents of the image handdrawn.bmp. Currently handdrawn.bmp contains nonsense, but you should replace this with a white digit drawn on a black background.