Vision that keeps learning after it ships.
Teach it new objects a few at a time. No retraining, no backprop, no stored photos.
Taught in lessons. Scored on everything.
87.17 ± 0.19% on CIFAR-10 with no backprop: one look, three seeds, settings chosen on validation, nothing previewed, 5 eyes each taught in a different lesson order. One strict eye: 84.93%. CIFAR-10.1, a fresh photo set: 74.38%.
See every number and its labelSee it on your own categories.
A sealed evaluation: you keep the answer key, we send back predictions, you score them. Report in two weeks.
Run it on your photosProtocols differ; we’d welcome a matched head-to-head. 87.17% takes 5 passes, one per eye, not one stream. These are our own runs, not yet independently replicated.
New things show up. It learns them.
New things on site.
New parts, products and objects on site, without a retraining cycle.
New variants on the line.
A new product variant on the line next week.
New targets, offline.
Learn a new target offline, in the field.
On-chip and wearable learning are where we’re headed, not a claim today.
Honest specs.
- Forgets on command5,000 photos removed in 0.149 s; its sums match a model that never saw them, bit for bit
- Two copies mergeIn opposite orders, bit-identical
- Readout memory7,835 MB down to 1.887 MB live, eye and on-disk codes counted separately
- Small whole model97.40% MNIST and 89.39% Fashion-MNIST in 150.46 KB (timing-based eye, one-layer local error rule, no matrix solve)
- Timing neurons16 neurons name 3,963 synthetic sequences at 99%; graded-release threshold reaches 99.57%
- Hardware todayRuns on a laptop with 64 GB of memory; not yet run on a phone, robot or chip
- RulesNo backprop, no stored photos
- IP5 US provisional patents filed