Notes from the work
Papers say what worked. This is where I want to write about the parts that don't fit in eight pages: why an architecture is shaped the way it is, what failed first, and how the reviewing process actually goes.
The first posts are being written
Below is what I plan to publish first. In the meantime, the research page covers what I'm working on, and the publications carry the full technical detail.
What's coming
-
How to design a multimodal architecture that survives missing data
Why the residual path matters more than the attention mechanism, and what happens to each fusion strategy when you start deleting metadata fields at test time.
-
Cross-attention, explained without the tensor soup
One diagram, one worked example, and the specific question cross-attention answers that concatenation does not.
-
Using an LLM as a neural architecture search controller
Notes from an ongoing experiment: what an LLM is good at proposing, where it repeats itself, and how to keep the search honest about the efficiency budget.
-
Why balanced accuracy, and why patient-wise splits
The evaluation choices that decide whether a medical imaging result means anything — and how easy it is to publish a number that doesn't.
-
Answering reviewers
What I've learned from rebuttals: which criticisms to take literally, which to read as a request for a different framing, and how long it really takes.