Research Notes
One note per piece of published work
Status of This Document
An index, revised whenever a note is added. Each entry below is a document in its own right, numbered in the order it was written rather than by subject. Papers are cited inside the notes; where a note and its paper disagree, the paper is authoritative.
Both notes so far come from the same line of work: an instrument records a signal, looks away for a while, and the interesting part is missing. Different methods were built to fill those gaps and, more importantly, to say honestly how much the filling should cost you in confidence.
Reconstruction Under Missing Data
An attention-gated U-Net, built for a nine-method comparison, which produced the largest reduction in parameter uncertainty of anything tested. Also an account of the measure on which it lost to a much simpler model.
A Model That Did Not Work
Time-aware neural ODEs for irregularly sampled signals. The idea was right, the reconstructions were good, and it still ended up in an appendix. Why, and what that was worth learning.
Elsewhere in the series:
Typeset in Source Serif 4 and IBM Plex Mono, self-hosted. Built without a framework. The only JavaScript on this site switches the theme. No tracker, no analytics. Just documents. Index · rev. 1 · 2026-08-08 · nikunjindoriya.com/research/