Exploring how
machines learn.
Architecture, data, optimization, and the questions in between. A collection of academic studies with interactive explanations and room for the experiments to come.
Sound Event Recognition — MLPC Project
University MLPC project: multi-label acoustic event recognition with MFCC features, a PyTorch AudioMLP model and recording-aware validation.
Transformer Attention Lab — PyTorch
Implemented multi-head self-attention and transformer blocks, then tested causal masking, padding masks and gradient propagation on synthetic tensors.
Graph Neural Networks — GCN vs MLP
Built a two-layer Graph Convolutional Network and an MLP baseline to study how graph structure affects node classification.
FNN vs CNN
Exploring how architectural inductive biases shape learning on a sequence-based latch task.
Seeing data differently
From high-dimensional features to interpretable projections with principal component analysis.
Finding structure without labels
Investigating hierarchical clustering, distance metrics, and the structure hidden in unlabeled data.
Inside the training loop
Connecting neurons, gradients, and parameter updates to the mechanics of learning.
Tiny Transformer Language Model & KV Cache
Planned advanced implementation: decoder-only language model, causal training and token-by-token inference with a KV cache.