Figure 2.

Hyperdimensional memory mechanism in HAP (Mitrokhin et al., 2019). A memory unit m consists of velocity bins as fields that are bound to another record of summed vector representations of time image slices from training. An input image is converted to its hyperdimensional vector representation and XOR'd with m. If it matches approximately with one of the image slice sums, the result contains the matching velocity representation with some noise.