Biomedical AI
Quantum-inspired kernels for scarce ECG labels
A research note on similarity geometry, patient-wise evaluation, and what the QADK study does—and does not—show.
Read noteWriting
Research notes that expose the reasoning.
Short notes on methods, evaluation, reproducibility, and engineering. Dates and references are explicit; claims tied to a paper return to its source record.
Interpretable ML
Interpretability by design with Formal Concept Analysis
How structured constraints can keep a clinical-risk model inspectable, plus the validation still needed before deployment.
Read noteMethods
The kernel trick, revisited for biomedical data
A compact explanation of kernels, inductive bias, and why small-data settings still reward careful similarity design.
Read noteReproducibility
ECG classification: evaluation before architecture
Patient leakage, imbalance, uncertainty, and other decisions that matter before selecting a model.
Read noteRisk-aware learning
Risk-aware reinforcement learning beyond average return
Why sequential decision systems need explicit downside objectives and skeptical backtesting.
Read noteInterpretable ML
Post-hoc explanations and intrinsically interpretable models
A practical distinction between explaining a fitted black box and designing a model whose structure is inspectable.
Read noteRisk-aware learning
CVaR, Value-at-Risk, and the tail that matters
A methods note on threshold risk, expected tail loss, and why evaluation assumptions matter.
Read noteMethods
LSTM or GRU? Start with the protocol
Architecture differences matter less than leakage control, baselines, horizons, and repeated evaluation.
Read noteResearch engineering
Reproducible Python environments for ML experiments
A current, framework-neutral setup checklist built around isolated environments, locked dependencies, smoke tests, and official install selectors.
Read note