Train a network to tolerate undesired weight shift during inference

1

Hi,

Is there anything could be done in the training to help the network be more resilient to undesired weight shift during inference?

For instance, suppose we train a simple MLP 784-512-10 and load the trained weights to some hardware, but the weight values are drifting down due to the reliability issue of memory components, thus degrading the accuracy. Assuming a known decreasing rate (v) of weight values, Is there any way to enforce a safety margin of weights that can retain a good accuracy to a certain time (e.g., 1 year) ?

Any inputs are appreciated! Thanks!

2

Could you explain this (interesting) use case a bit more?
Why would weight values drift down, if your memory is unreliable?
Wouldn’t this introduce random errors in the memory and thus corrupting the values in an unpredictable way?