WritingML
Models built by hand to understand what the libraries hide: backpropagation from first principles and Bayesian model selection.
2 posts · newest first
Fitting five polynomial regression architectures to four EEG input channels using nothing but normal equations, then comparing them with RSS, log-likelihood, AIC and BIC, validating on a held-out split, and running rejection ABC — plus the two places where the numbers quietly disagreed with the method.
15 minR
Working out the forward pass, softmax with cross-entropy, the chain rule and mini-batch gradient descent by building a multi-layer perceptron in plain Java — no ML libraries — and learning why data layout decides what the algorithm can express.
9 minJava