02 · Projects and experiments
Projects I took on to understand something better: GPU kernels, evolved neural networks, brain signals and language.
Right now I'm curious whether fine-grained hardware signals — SM occupancy, memory bandwidth, register pressure — can guide kernel optimization better than runtime alone.
Ongoing Self-directed
Softmax, tiled matmul and fused bias+GELU+dropout kernels on an RTX 3060, checked against PyTorch eager and torch.compile and profiled in Nsight Compute to see what really limits each one. A tiled online softmax raised occupancy from 33% to 97% yet ran 73% slower: a useful negative result.
ResultFused bias+GELU+dropout 3.3× over eager; TF32 tensor cores +45% on matmul; softmax 1.9×.
Evolves ProtCNN-style architectures with a genetic algorithm — candidate CNNs built from predefined blocks, trained, ranked and recombined over 50 iterations — for multiclass classification across 17,929 Pfam v32.0 families. Validated on a random and a clustered distant-homolog split, with error reported by sequence identity to avoid inflated results.
Result11.5% overall error vs. ProtENN 12.2%, HMM 18.1%, ProtCNN 27.7% and BLASTp 35.9%.
Genetic programming that evolves neural-network classifiers, with a hybrid crossover sending part of each generation through hill climbing for faster convergence and the rest through standard crossover to avoid local maxima. Evaluated on the Wisconsin Breast Cancer Database (699 samples) across 50-50, 60-40 and 70-30 splits.
Result97% (50-50) and up to 99% (70-30) accuracy, competitive with C4.5 (94.9%) and fuzzy-genetic (97.4%).
Full EEG workflow — Fourier-based frequency decomposition on 20-participant recordings — feeding GP classifiers with tuned fitness functions for four emotional states.
Result89% average accuracy, 6–10 points above SVM, Random Forest and NN baselines.
Neural text classifier (embeddings, CNN and recurrent layers) for a binary label plus three harassment subtypes, on ~25k cleaned social-media posts with threshold tuning and class weighting.
Result8–12% higher multi-label F1 than statistical baselines on noisy Twitter data.
LSTM models trained on 5,000+ lyric sequences and 500+ MIDI files, combining lyric and melody generation into one pipeline with temperature-controlled sampling.
ResultStylistically coherent lyrics and melodies, stable across sampling temperatures.