AI that keeps working on damaged chips
I tested an AI that keeps working when the chip it runs on starts to break.
Military AI is a glass cannon. It is very accurate on perfect computer chips, and it breaks down when those chips wear out from heat, radiation, or manufacturing defects. Drones and aircraft cannot stop in the middle of a mission to swap a worn-out chip.
I tested a different kind of AI, called hyperdimensional computing, against a standard neural network. Both had to recognize radio signals while I simulated chips with more and more damage.
With 20 percent of the chip damaged, the standard network lost 28.7 points of accuracy. Mine lost 1.4. That is about 20 times more resilient.
My first approach did not work. The write-up shows the failure and how I changed course.
I wrote the paper with Dr. Jeong Yang. It was published in Military Cyber Affairs on September 21, 2026. Three days after it came out, I presented it on the Stars track at HammerCon, at Johns Hopkins APL.
20 percent of the chip is damaged. The same cells on both.
Standard neural network
My approach (HDC)
About 20 times more resilient.
- PaperCognitive Resilience at the Edge: Hyperdimensional Computing versus Deep Learning for Hardware-Degraded RF Classification
- PublishedMilitary Cyber Affairs, Vol. 9, Iss. 1, Article 13 (2026)
- DateSeptember 21, 2026
- Co-authorDr. Jeong Yang
- Accuracy lost at 20 percent chip damageStandard neural network 28.7 points. Hyperdimensional computing 1.4 points.
- PresentedHammerCon 2026, Stars track, Johns Hopkins APL
