project #1
a simple 10 class mnist digit classifier pre-trained to classify handwritten digits from 0 to 9 implemented in desmos
try it yourself!
training dataset courtesy of oliver gibson on kaggle
the process
i downloaded the 10x10 handwritten digits dataset from kaggle (linked over yonder) before cleaning it and creating a training and testing split which i fed to a stochastic gradient descent model (also built from scratch with just numpy and pandas). i used cross-entropic loss and trained over 10 or so epochs and then extracted the weight and bias vectors from each layer into desmos as a column of a table by formatting them in desmos markup format which i just pasted in using desmodder. from there i just had to implement regular forward propagation using lists. i also took the time to visualize individual weights, coloring them red when positive and blue when negative and also adjusting thickness and opacity according to magnitude.
