A Novel Verification Method of Minimizing Verification Error in AND-Type Flash Arrays

A novel verification method for accurate neuromorphic systems is proposed. First, the origin of verification errors is investigated, pinpointing the off-current and the discrepancy between the verify and inference phases as the primary contributors. Subsequently, we analyze the verification error relative to the cell states in a fabricated AND-type flash array with 28-word lines. The verification error was measured to be in the range of 150–200 nA, accounting for 2.3 % of the total current. Experimentally, the proposed verification method demonstrates 3x lower verification error than conventional methods on average. Consequently, the proposed method effectively mitigates an accuracy drop, defending against drops of up to 11.57 %p in classification tasks.