A vision transformer system that can dial down its own precision to keep working through hardware glitches, instead of just crashing.
Researchers describe WARD, a framework for running Vision Transformer AI models on edge devices - the kind used in drones, satellites, and factory sensors - that have to keep working as power budgets shift and hardware bits start flipping from radiation or wear. WARD splits the network into two physically separate subnetworks and switches between four operating modes (full precision, low power, high reliability, and adaptive) depending on conditions. The team built it on a small FPGA-based accelerator with hardware support for switching modes in a few clock cycles. In tests under high bit-error rates, the split design held the network's failure rate to 1.79 percent while adding less than 5 percent extra chip area.
Most fault-tolerant AI designs assume conditions stay fixed, and most continual-learning systems ignore the possibility that the chip underneath is failing at the same time. WARD's contribution is folding both problems into one adaptive control loop, aimed at places where AI can't just reboot and try again - a satellite in orbit, a robot on a factory floor. That combination, more than the raw numbers, is what sets it apart from prior work.
These are lab results on one FPGA prototype, not production silicon, so treat the sub-5-percent overhead figure as a promising data point to watch rather than a guarantee it survives contact with a real product line.