6 February 2023 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
Brain-inspired Hyperdimensional Computing (HDC) is an emerging framework in low-energy designs for solving classification tasks at the edge. Unlike mainstream neural networks (NNs) with high computational complexity, HDC offers a fast training strategy and an energy-efficient inference mechanism by computing data in the high-dimensional space with high parallelism. However, the original computing method of HDC results in weak accuracy due to its simplicity. Therefore, many works have proposed algorithmic innovations to improve the accuracy of HDC, while maintaining its efficiency with dedicated hardware architectures. Given this period of rapid evolution, this review paper presents a comprehensive study of recent developments improving the accuracy-efficiency trade-off of HDC, covering both algorithm and hardware-level approaches. Besides, as the amount of data generated by the Internet of the Things (IoT) devices keeps increasing, edge-cloud collaboration has become a key area of research. The emerging federated learning (FL) has become a key enabler in overcoming security and privacy issues by exchanging locally trained models for global aggregation. However, the energy constraint of edge devices hinders them from the real-time training of NN models. Alternatively, adopting HDC as a local model can be a potential solution to alleviating this issue since the computational requirements of HDC are affordable to most edge devices. In addition, HDC-based coding for wireless communication has been gaining traction due to its noise robustness. Therefore, this paper also covers recent works on HDC-based edge-cloud collaborative learning. Finally, we conclude this paper by highlighting the possible applications and future directions of HDC.
25 January 2021 in IEEE Open Journal of Circuits and Systems
Extreme learning machine (ELM) has shown to be an effective and low-power approach for real-time electrocardiography (ECG) anomaly detection. However, prior ELM inference chips are noise-prone and lacking in reconfigurability. In this article, we present an arbitrarily reconfigurable ELM inference engine fabricated in 40-nm CMOS technology for robust ECG anomaly detection. By combining Adaptive boosting (Adaboost) and Eigenspace denoising with ELM (AE-ELM), robust classification under noisy conditions is achieved and saves the number of required multiplications by 95.9%. For chip implementation, a reconfigurable VLSI architecture is designed to support arbitrary complexity of AE-ELM, accounting for dynamic change in application requirements. On the other hand, we propose to construct the input weight matrix of ELM as a Bernoulli random matrix, which further reduces the number of multiplications by 55.2%. For real-time detection, parallel computing is exploited to reduce the latency by up to 86.8%. Overall, the 0.21-mm 2 AE-ELM inference engine shows its robustness against noisy signals and achieves $1.83\times$ AEE compared with the state-of-the-art ELM design.
15 April 2021 in IEEE Transactions on Computers
Brain-inspired Hyperdimensional computing (HDC) has shown its effectiveness in low-power/energy designs for edge computing in the Internet of Things (IoT). Due to limited resources available on edge devices, multi-task learning (MTL), which accommodates multiple cognitive tasks in one model, is considered a more efficient deployment of HDC. However, as the number of tasks increases, MTL-based HDC (MTL-HDC) suffers from the huge overhead of associative memory (AM) and performance degradation. This hinders MTL-HDC from the practical realization on edge devices. This article aims to establish an MTL framework for HDC to achieve a flexible and efficient trade-off between memory overhead and performance degradation. For the shared-AM approach, we propose Dimension Ranking for Effective AM Sharing (DREAMS) to effectively merge multiple AMs while preserving as much information of each task as possible. For the independent-AM approach, we propose Dimension Ranking for Independent MEmory Retrieval (DRIMER) to extract and concatenate informative components of AMs while mitigating interferences among tasks. By leveraging both mechanisms, we propose a hybrid framework of Multi-Tasking HDC, called MulTa-HDC. To adapt an MTL-HDC system to an edge device given a memory resource budget, MulTa-HDC utilizes three parameters to flexibly adjust the proportion of the shared AM and independent AMs. The proposed MulTa-HDC is widely evaluated across three common benchmarks under two standard task protocols. The simulation results of ISOLET, UCIHAR, and MNIST datasets demonstrate that the proposed MulTa-HDC outperforms other state-of-the-art compressed HD models, including SparseHD and CompHD, by up to 8.23% in terms of classification accuracy.
7 December 2023 in IEEE Transactions on Circuits and Systems I: Regular Papers
Convolutional neural networks (CNNs) are known for their exceptional performance in various applications; however, their energy consumption during inference can be substantial. Analog Computing-In-Memory (CIM) has shown promise in enhancing the energy efficiency of CNNs, but the use of analog-to-digital converters (ADCs) remains a challenge. In analog CIM-based accelerators, ADCs convert analog partial sums from CIM crossbar arrays to digital values, with high-precision ADCs accounting for over 60% of the system’s energy consumption. To prevent ADCs from damaging the energy efficiency benefits of CIM, researchers have explored quantizing CNNs to use low-precision ADCs, trading off accuracy for energy efficiency. However, these approaches often necessitate data-dependent adjustments to minimize accuracy loss. Instead, we observe that the first most significant toggled bit indicates the optimal quantization range for each input value. Accordingly, we propose a range-aware rounding (RAR) method for runtime bit-width adjustment, eliminating the need for pre-deployment efforts. RAR can be easily integrated into a CIM accelerator using dynamic block-floating-point arithmetic. We also seamlessly incorporate a bit-level zero-skipping mechanism by dynamically forming input blocks. Experimental results demonstrate that our methods maintain accuracy while achieving up to 1.81 $\bm{\times }$ and 2.08 $\bm{\times }$ energy efficiency improvements on the CIFAR-10 and ImageNet datasets, respectively, compared with state-of-the-art techniques.
6 February 2023 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
The wide adoption of convolutional neural networks (CNNs) in many applications has given rise to unrelenting computational demand and memory requirements. Computing-in-Memory (CIM) architecture has demonstrated great potential to break the memory wall and effectively execute CNN workloads. Ongoing research focuses on pruning or quantizing CNNs to achieve higher efficiency on CIM. However, prior works preclude the possibility of integrating both techniques in the same framework. On the other hand, directly incorporating energy estimation during the model compression process has not been well explored in the literature. In this paper, we present an Energy-aware Unified Pruning-Quantization (E-UPQ) mechanism, a novel framework for automated compression (pruning + quantization) of CNNs while considering the energy-accuracy trade-off. Specifically, E-UPQ interweaves pruning and quantization seamlessly by viewing pruning as a special case of “0-bit” quantization during the mixed-precision search. In addition, E-UPQ introduces a set of trainable parameters to incorporate energy information during the compression process, closing the gap between compression policy and energy optimization. Experimental results evaluated on DNN+NeuroSim show that E-UPQ reduces energy consumption by up to 79.3% and 66.6% for VGG-16 and ResNet-18, respectively, compared with the state-of-the-art work, while achieving similar accuracy on CIFAR-100. Layer-wise analysis and ablation studies are provided to validate the effectiveness of the E-UPQ. We also present the corresponding CIM architecture to support the proposed E-UPQ framework.