Research Areas for 2025 - Building

permalink

I'm focusing on starting 2025 by getting more actively connected to research in a variety of fields for the year.

Topics

Compilers

  • SSA
  • SIMD

I'd also like to see if I can make progress on tooling to improve code generation that is only intended to implicitly leverage SIMD instead of having to generate it explicitly. For example, if I can generate 4 independent add operations and write them in sequence, it should be feasible for Superword-Level Parallelism tooling to match to SIMD if the cost model lines up.

Drones

  • Visual Odometry
  • Visual Inertial Odometry
  • VI SLAM w/ Factor Graphs

Celestial Navigation

One topic that I rediscovered because of drones, but originally connected with because of sailing and history is celestial navigation. Specifically, celestial navigation for drones made it to Hacker News (discussion link), and I quite enjoyed reading the paper (DOI, open access via the MDPI journal). It sparked some new ideas for me and generally renewed my interest in learning celestial navigation as a skill

Underwater Autonomy

An additional topic area that I see as logically adjacent to drones is underwater autonomous vehicles. Once I've taken some time to explore through the visual odometry for flying drones, I'd also like to spend some time exploring similar challenges for underwater vehicles.

Learning Practices

The research is also taking something of an applied turn, in that in order to maximize my understanding, I'm aiming to implement more of the papers as I go instead of just passively consuming them. I see this as a proxy for understanding, in that if I can't "explain" the paper to a computer, I probably don't understand the algorithm well enough.

Discovery

I'm also trying to refine how I discover papers to maximize my reading for what I'm most likely to find interesting, enjoyable and tractable. In a previous attempt, I'd used an interactive rating system to recommend arXiv papers; however, it didn't allow me to easily revisit papers that I had read to upgrade or downgrade them. I'm actually starting to think that an interface more like a spreadsheet could help me side-step the interactivity and storage portion, and then I can build the recommender system as a batch processing.

As a short addendum, the previous approach that I had taken focused on TF-IDF as a measure of interesting terms and then doing a search that way. In the going-forward mode, I'd like to explore with using sentence or paragraph embeddings (e.g. BERT, sentence-transformers ), and some combination. For example, given papers that I've read and found interesting enough to implement, find me papers that match as well as invert the embeddings to find a three word (or similar short description) of a topic area. I'd also like to be able to do topic search.