Hessian-driven unequal protection of DNN parameters for robust inference

Impact of HKMG and FDSOI FeFET drain current variation in processing-in-memory architectures

“…We observe full accuracy recovery to the 85.86 % baseline by using variation aware training in the LeNet-5 emulation for Fashion-MNIST classification. This accuracy recovery becomes much more difficult in the MobileNetV2 classification of CIFAR-10 due to the significantly larger network and the large accuracy reductions caused by noise [24]. Therefore, the drop in classification accuracy caused by the FeFET I DS variation measured in this thesis is shown to not be a limiting factor to our PIM accelerator design in the case of small workloads, but larger workloads present challenges which require more advanced design techniques to remedy.…”

“…However, the PIM architecture still shows promise with accuracy well over 80 % in many cases. Noise aware training can fully recover the accuracy drops observed in PIM emulation of LeNet-5, although this becomes more challenging for the larger and less accurate MobileNetV2 classification of CIFAR-10[24]…”