The problem
Analog crossbar arrays offer an efficient way to execute generative adversarial networks, but process and conductance variations can degrade output quality. Exhaustive testing is costly.
What we did
We developed an adaptive testing framework for compute-in-memory GANs that uses backpropagation to guide test compaction. The framework reduces the test workload while targeting the effects of variation in analog crossbar implementations.
What we found
The resume reports approximately 7× testing speedup relative to exhaustive test sets. This work makes test efficiency part of the design of reliable analog AI hardware.
Paper
Anurup Saha, Ashiqur Rasul, Thomas Walton, Amirali Aghazadeh, Abhijit Chatterjee. “Adaptive Testing of Compute-in-Memory GANs Using Backpropagation-Guided Test Compaction.” DATE 2026.