What we study

Adaptive resonance models

Fuzzy ARTMAP learns new categories on the fly and keeps the old ones. It is simple, fast to train and easy to inspect. The open questions are practical ones: how its vigilance setting and decision readout should be tuned, and how it compares with other methods when compute is held equal.

Explainable classification

In health, security and other regulated work, someone is accountable for each decision. We favour models whose reasoning can be read and checked by that person.

Honest evaluation

Equal compute budgets for every compared method. An analysis plan fixed before the main run. Deviations logged. Null results reported. Code and full results published when a study is finished.

Work in progress

• Study 1design stage
  A budget-matched factorial comparison of heuristic search, a genetic algorithm and Q-learned vigilance control in Fuzzy ARTMAP, on public UCI datasets.
• Edge studyplanned
  Where a small language model's parameters can be stored on a Raspberry Pi (memory, SSD or SD card), and what each choice costs per generated token.

Status labels are literal: nothing here is described as finished until it is published.