You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
Every AI system in use is frozen the day it ships, and retraining it makes it forget. I am measuring, in public and under predeclared criteria, whether a hyperdimensional computing (HDC) learning layer can keep learning without forgetting, and releasing the benchmark harness and the layer as open public goods.
I built FlyBoxLab, a whole fruit fly brain simulator (166,700 neurons) that runs live in a browser and draws about 50 visitors a day, as a public testbed. On it, and on standard digit data, I ran head-to-head tests between an HDC learning layer and a neural network baseline, with win conditions written down before each run and identical data and splits. The layer learned new classes from 5 examples at 100% (digits) and 94.5% (fly behaviour) accuracy while losing under 2 points on known classes; the baseline lost 12 to 28 points learning the same classes, and 97.8 points (digits) to 100% of first-half accuracy (fly data) in a sequential training test. All write-ups, including the losses (the layer currently costs 32x to 84x more compute per correct answer; its robustness edge held only on digits), are public.
The work will have worked if the benchmark harness can be rerun by any researcher on their own tasks, the fast compiled runtime is released openly, and a first external adopter is running the layer inside a real product that must adapt in the field, with measurements published.
(1) The open benchmark harness, packaged so any researcher can rerun the forgetting comparisons on their own tasks; (2) a fast compiled runtime for the layer, released openly; (3) a first external adopter running the layer inside a real product that must adapt in the field, with measurements published. Total budget $25,000, essentially the researcher's time; the work is CPU-only, so there is no significant compute cost.
Gaurav Narlawar, 23, independent builder and software engineer in New Delhi, India. Primary engineer on a 15,000+ user product by day; this research is self-funded and done in the open (36 public GitHub repositories). Links: https://gvrv.cc ; https://flyboxlab.vercel.app ; https://the-learning-layer.vercel.app
The main risks are already visible in the published results: the layer currently costs 32x to 84x more compute per correct answer, and its robustness edge held only on digits. The benchmark harness and the layer would still be released openly.
None. This research is self-funded.