Noise, Scores, and Straight Lines
Diffusion models and flow matching, developed from paths of distributions, scores, velocity fields, SDEs, ODEs, DDPM, guidance, and solver intuition.
Notes
Technical notes, refreshers, and worked derivations on varied ML topics.
Long-form Notes
Functions In, Functions Out
Neural operators as learned solution operators: DeepONets, FNOs, graph operators, low-rank variants, physics-informed losses, and time-dependent maps.
ML 101s
Attention Backward Pass
A full index-notation derivation of self-attention backpropagation, including the row-wise softmax Jacobian and the final matrix-form gradients.
Direct Preference Optimization
A clean DPO derivation from KL-regularised RL and Bradley-Terry preferences, plus implementation details and common variants.
Generalised Advantage Estimation
The GAE recursion from k-step advantage estimators, TD residuals, masking, value targets, and the lambda bias-variance knob.
PyTorch: Storage, Shape, and Axis-Moving
Storage, shape, strides, views, reshapes, transposes, contiguity, and the exact class of tensor bugs that silently scramble data.