TaOx ReRAM Cell-level Unidirectional Neural Network Weight Control for Non-linearity & Variation Robust Transfer Learning of Low Cost Digital eCiM

96.4% high inference accuracy in transfer learning of edge all-digital AI is achieved by proposed Gradual Reset and by proposed Unidirectional Dual-cell training (UDCT) for 40nm TaO
x
ReRAM embedded Computation-in-Memory (eCiM). Proposed Gradual Reset writes Neural Network (NN) weights with large dynamic range and linearity. Optimizing Initialization before Gradual Reset controls slope of read current in Gradual Reset cycles. Proposed UDCT introduces adaptive conductance threshold to update NN weights by utilizing unidirectional current shift at Cell-level and to remove non-linearity and variation at Chip-level.