Unfiltered notes from my AI/ML learning journey — the messy middle, not the polished takes.
I'm Aman, self-taught, working through the IITM BS Data Science program while building toward AI systems and research engineering. This repo is where I think out loud: half-formed hypotheses, wrong turns, raw reasoning before I've read the "correct" answer. Some of it will be naive. That's the point — it's a record of how my thinking evolves, not a highlight reel.
For finished projects and code, see my other repos.
- Emotion as Compression: A Wild Guess About How AI Should Remember — Reasoning through continual learning, memory consolidation, and whether emotional intensity could drive what an AI system remembers vs. forgets. Written 3 months into classical ML, flaws included.
- Governor AI: A Supervisory Layer for Machine Learning Systems — An early idea exploring whether ML systems need a supervisory layer to filter spurious patterns from meaningful ones before training, inspired by studying overfitting and the bias-variance tradeoff.
Bi-weekly research sprints — one hypothesis, one small experiment, one honest verdict. Unlike the posts above, these are evidence-backed: real numbers, real plots, a real answer to "was I right." Format: Hypothesis → Method → Result → Verdict, published whether the hypothesis holds or not.
- Does model choice change compute efficiency at similar accuracy? XGBoost v/s GBM — Confirmed: in the range where accuracy converges, XGBoost is ~57x more compute-efficient per second of training time than GBM. Default hyperparameters only.
- Is XGBoost's regularization the reason for the accuracy gap on small datasets, compared to GBM? — Refuted: loosening regularization made XGBoost more volatile at small dataset sizes, not more competitive with GBM — pointing to sampling noise, not regularization strength, as the real driver. Also the memo where a sparse 5-6 point plot taught me not to trust conclusions from too few data points.
New posts and sprints get added roughly monthly and bi-weekly respectively, in step with my learning roadmap.