Fig. 4. The MANN architecture with the binary key memory using analog in-memory computations.

The architecture is simplified for efficient few-shot inference: (1) The transformed HD support vectors are stored in a memristive crossbar array as the binary key memory; the query vectors are binarized too. (2) The cosine similarity (α) between the input query vectors and the support vectors is computed through in-memory dot products in the crossbar using Eq. (10). (3) To further simplify the inference pipeline, the normalization of the attention vectors and the regular absolute sharpening function are bypassed. The accumulation of similarity responses belonging to the same support label in the value memory and finding the class with maximum accumulated response are implemented in software. The binary query/support vectors have 512 dimensions. m and n stand for “way” and “shot” of the illustrated problem respectively. A similar architecture with the bipolar key memory is shown in Supplementary Fig. 1.