Skip to main content
Back

UC Irvine Fixes QuantHD's 48-Point Accuracy Collapse on Analog AI Chips

A joint optimization scheme for hyperdimensional computing recovers 48% accuracy on QuantHD and 5.4x on RelHD when analog compute-in-memory hardware distorts similarity operations.

Pro

Paper, Full Access
  • UC Irvine team proposes hardware-aware optimization for hyperdimensional computing on noisy compute-in-memory chips (paper).
  • Encoding is reformulated as minimizing the Frobenius norm between an ideal kernel and the hardware-distorted one.
  • Optimized QuantHD reaches 84% on ISOLET versus 36% for naive QuantHD under the same nonlinearity.
  • Optimized RelHD on the Cora citation graph delivers a 5.4x accuracy gain over the naive baseline.
  • Graph reconstruction with GrapHD, which fails entirely under analog distortion, becomes reliable again after joint optimization.
  • Published at GLSVLSI 2025; no public code repository has been released with the paper.

Analog compute-in-memory chips promise large energy wins by doing math directly inside memory arrays, but they are noisy, nonlinear, and hostile to the clean floating-point assumptions that modern machine learning depends on. A team from UC Irvine has proposed a way to make brain-inspired hyperdimensional computing survive those distortions, with results showing it can recover almost all the accuracy that naive implementations lose on realistic analog hardware.

The paper, Robust Reasoning and Learning, reframes hypervector encoding as an optimization problem tuned to the specific quirks of the chip it will run on. When applied to QuantHD, the method reaches 84% accuracy under severe hardware-induced perturbations, a 48-point jump over naive QuantHD in the same conditions.

Overview diagram showing how hyperdimensional computing operations map onto compute-in-memory hardware

Why analog memory keeps breaking ML

Compute-in-memory (CIM) architectures skip the shuttle of data between DRAM and processor by performing multiply-accumulate operations directly in the memory array using physical laws like Ohm's and Kirchhoff's. Great for energy, but the analog dot products saturate, drift, and quantize in ways that a GPU never does. Traditional machine learning depends on high-precision arithmetic and near-ideal hardware assumptions, which sits awkwardly with the variability of aggressively scaled semiconductor devices, and CIM architectures pile on nonlinear distortions and reliability concerns.

Hyperdimensional Computing (HDC) is a candidate workaround. It represents data as very long random vectors (thousands of dimensions) and does learning through simple algebraic operations: bundling (elementwise addition to superpose concepts) and binding (elementwise multiplication to associate concepts, like linking a node with its neighbors in a graph). Because information spreads across thousands of components, small per-component errors typically wash out. That noise tolerance is why the field has become attractive for FeFET, RRAM, and phase-change memory accelerators.