Teaching a gesture radar a new user without backpropagation

A radar can read hand gestures, but a model trained on some people does worse on someone new. I tested whether it can adapt to a new user with forward passes only, which is cheaper to run on small devices than normal training.

On a new user the model scored 68.9 percent with no adaptation. Normal gradient training on 5 examples per gesture raised that to 79.3 percent. Forward-only adaptation reached 81.2 percent within 4,000 forward passes.

The cost depends on an effective dimension of about 8, far below the model's 14,843 parameters. My formula predicted it within 15 percent in 93 percent of the last-layer tests.

I wrote it with Ashley Mendez as a two-page Late News short paper for IEEE-EPIC 2026, and we submitted it on October 20, 2026.