Cosmic radiation: a network of quantified neurons repairs its own corrupt weights without a backup copy

Thought for spacecraft and future data centers in orbit, the Projection Healing method doubles the interval between two memory cleanings on commercial equipment.

In space or near particle accelerators, energy particles regularly reverse bits in memory, and the weights of a network of neurons gradually deteriorate. Peter Xiangyuan Ma and Joshua S. Bloom propose a method of repairing execution for quantified networks, on standard commercial equipment, without storing replica weights.

Their approach extends the training "conscious of mistakes" by adding a penalty that concentrates healthy weights and activations in a low-level subspace; when a value is reported as corrupt, it is rebuilt from intact values. On six regression tasks, activating this repair further reduces the cumulative error by about 23%. It prolongs the memory-tolerated cleaning interval by a median factor of 2.04 for a 5% degradation of the error, compared to 1.62 for a classic correction code of type SECEDED.