A Computing-in-Memory-based One-Class Hyperdimensional Computing Model for Outlier Detection

TL;DR

ODHD introduces a one-class outlier detector based on hyperdimensional computing, representing inliers with a single one-class hypervector and detecting outliers via cosine similarity to a data-derived threshold. The authors extend this idea with IM-ODHD, a SRAM-based computing-in-memory platform, and a HW/SW co-design to accelerate both training and inference in memory. Across six ODDS datasets, ODHD outperforms OCSVM, isolation forest, autoencoder, and HDAD on accuracy, F1, and ROC-AUC, while IM-ODHD dramatically reduces training/testing latency (up to 331.5x/889x) and energy (up to 14x/36.9x) relative to GPU implementations. The work demonstrates a viable, energy-efficient path for high-dimensional, memory-centric anomaly detection and elucidates design tradeoffs via a thorough design-space exploration. It also introduces CiM-friendly thresholding using mean/ MAD and details near-memory permutation and bundling to support scalable HDC-based learning on silicon.

Abstract

In this work, we present ODHD, an algorithm for outlier detection based on hyperdimensional computing (HDC), a non-classical learning paradigm. Along with the HDC-based algorithm, we propose IM-ODHD, a computing-in-memory (CiM) implementation based on hardware/software (HW/SW) codesign for improved latency and energy efficiency. The training and testing phases of ODHD may be performed with conventional CPU/GPU hardware or our IM-ODHD, SRAM-based CiM architecture using the proposed HW/SW codesign techniques. We evaluate the performance of ODHD on six datasets from different application domains using three metrics, namely accuracy, F1 score, and ROC-AUC, and compare it with multiple baseline methods such as OCSVM, isolation forest, and autoencoder. The experimental results indicate that ODHD outperforms all the baseline methods in terms of these three metrics on every dataset for both CPU/GPU and CiM implementations. Furthermore, we perform an extensive design space exploration to demonstrate the tradeoff between delay, energy efficiency, and performance of ODHD. We demonstrate that the HW/SW codesign implementation of the outlier detection on IM-ODHD is able to outperform the GPU-based implementation of ODHD by at least 331.5x/889x in terms of training/testing latency (and on average 14.0x/36.9x in terms of training/testing energy consumption.

Figures (6)

• Figure 1: The algorithmic flow of ODHD with six key phases.
• Figure 2: IMC architecture for outlier detection. (a) ($P \times Q$) mat-level architecture. The PEs are marked green for source and blue for destination. (b) Detail of one ($M \times N$) subarray.
• Figure 3: An example for the permutation with our CiM architecture; (a-f) corresponds to the steps of round 1; (g) depicts round 2.
• Figure 4: Threshold trend on MNIST dataset with both standard deviation and mean absolute deviation metrics.
• Figure 5: Comparison between ODHD , IM-ODHD and four baseline methods based on three metrics, ACC, F1 and ROC.