Fig. 1. Proposed methodology to merge deep network representations with vector-symbolic representations in high-dimensional (HD) computing.

The goal is to guide a deep network controller to conform with HD computing by assigning quasi-orthogonal HD vectors to unrelated objects in the key memory. The HD vectors are then directed by adjusting the controller weights during meta-learning in such a way that the query vector gets near the set of correct class vectors, and the vectors from different classes move away from each other to produce mutually quasi-orthogonal vectors in the key memory (demonstrated in a 3D space for sake of visualization). This is achieved by using proper similarity and sharpening functions, regularizer, and expanding the vector dimensionality at the interface of controller and key memory. Then, the resulting real-valued representations can be readily transformed to dense binary/bipolar HD vectors for efficient and robust inference in a key memory using in-memory computing.