Machine learning and signal processing — one-page summaries, full papers linked.
CNN Music Genre Classification
79% accuracy from a from-scratch neural network
Problem. Classify music genres from audio without off-the-shelf models — derive the mathematical foundations of convolutional networks and implement them from first principles.
Approach.
Derived the convolution, pooling and backpropagation mathematics by hand before writing code.
Implemented the network and training loop in Python, no ML frameworks.
Evaluated against a held-out set with per-class error analysis.
Results. 79% classification accuracy. Recognized on "Ivanov's Top Papers of the Decade", and the work led directly to a programming assistant job offer.
Confusion matrix — most confusion between acoustically adjacent genres.
