SAPIENS Chip: Human-Like Learning
SAPIENS is the first integrated chip demonstrating on-chip, one-shot learning, enabled by exposing the unique properties of RRAM with the cross-layer design and integration for memory-augmented neural networks.
Aiming to continuously learn from one or few examples on the fly, SAPIENS is built to work with memory-augmented neural networks (MANNs). The name for this chip was inspired by how human can learn new concepts with very few or even single example, dating back to the early stage of intelligence about 150 thousand years ago.
For MANN models, the associative memory is the backbone that augments the feature learning capabilities of shallow neural nets with scarce data. The ability to memorize the associations of the features with the similarity information is the key mechanism to eliminate the need of abundant data.
SAPIENS is the first integrated chip for one-shot learning, having 65 thousand resistive memories monolithically integrated with mixed-signal silicon CMOS in TSMC 40-nm process, while achieving software-comparable accuracies on 32-way, one-shot learning workloads (i.e., learn and infer on 32 unseen classes with one image per class) from real-time chip measurements. Unlike incumbent memory technologies that are volatile requiring constant power or refresh, the non-volatile nature of the SAPIENS chip means that the features are embedded on chip with zero standby power, showing a viable path towards life-long learning and inference at the edge.
Related publications:
VLSI 2021, TED 2021