Deep Learning Experiments
A collection of applied deep learning notebooks: distribution-comparison metrics, autoencoder-based dimensionality reduction, and CNN architecture benchmarking.
The Problem
Understanding how architectural and statistical choices affect model behavior requires running the comparisons yourself, not just reading about them.
Why I Built It
Self-directed research notebooks built to compare methods head-to-head rather than take published benchmarks at face value.
What I Built
Implemented and compared distribution-divergence metrics (KL divergence, JS divergence, Kolmogorov–Smirnov) for evaluating generative model outputs; built autoencoders for dimensionality reduction and compared classifier performance on the reduced representations; benchmarked VGG-19, ResNet50V2, and Inception v4 for image classification alongside LSTM/GRU/BiRNN variants for time-series prediction.
Technical Challenges
Keeping comparisons fair required controlling for training budget and hyperparameter tuning effort across architectures with very different parameter counts and training dynamics — otherwise the benchmark just measures who got more tuning time.