Memristor-Based Approximate Query Architecture for In-Memory Hyperdimensional Computing

As a new computing paradigm, hyperdimensional computing (HDC) has gradually manifested its advantages in edge-side intelligent applications by virtue of its interpretability, hardware-friendliness and robustness. The core of HDC is to encode input samples into a hypervector, and then use it to query the class hypervector space. Compared with the conventional architecture that uses CMOS-based circuits to complete the computation in the query operation, the hyperdimensional associative memory (HAM) enables the query operation to be completed in memory, which significantly reduces the query delay and energy consumption. However, the existing HDC algorithms require the HAM to achieve high precision query in inference, which leads to the complex structure of the HAM, and thus makes the area and energy consumption of the HAM unable to be further reduced. In this paper, a novel efficient HAM architecture based on approximate query method is proposed, to simplify the existing architecture. Meanwhile, a training method of HDC which matches the proposed HAM architecture is proposed to compensate for the decrease in accuracy caused by approximate query. Experimental results show that the proposed HAM framework can save more than 60% of area and energy consumption, and achieve accuracy comparable to existing state-of-the-art methods by using the proposed training method.