Store-n-Learn: Classification and Clustering with Hyperdimensional Computing across Flash Hierarchy

Processing large amounts of data, especially in learning algorithms, poses a challenge for current embedded computing systems. Hyperdimensional (HD) (HDC) is brain-inspired paradigm that works with high-dimensional vectors called hypervectors . HDC replaces several complex computations bitwise and simpler arithmetic operations at the expense an increased amount data due to mapping into space. These hypervectors, more often than not, cannot be stored memory, resulting long transfers from storage. In this article, we propose Store-n-Learn, in-storage solution performs classification clustering by implementing encoding, training, retraining, inference across flash hierarchy. To hide latency training enable efficient computation, introduce concept batching HDC. We also present on-chip acceleration encoding planes. This enables us exploit high parallelism provided hierarchy encode multiple points parallel both batched non-batched fashion. Store-n-Learn implements single top-level FPGA accelerator novel implementations inference, on encoded data. Our evaluation over 10 popular datasets shows average 222× (543×) faster CPU 10.6× (7.3×) state-of-the-art solution, INSIDER (clustering).