Search

The combination of Large Language Models (LLM) and Automatic Speech Recognition (ASR), when deployed on edge devices (called edge ASR-LLM), can serve as a powerful personalized assistant to enable audio-based interaction for users. Compared to text-based interaction, edge ASR-LLM allows accessible and natural audio interactions. Unfortunately, existing ASR-LLM models are mainly trained in high-performance computing environments and produce substantial model weights, making them difficult to deploy on edge devices. More importantly, to better serve users' personalized needs, the ASR-LLM must be able to learn from each distinct user, given that audio input often contains highly personalized characteristics that necessitate personalized on-device training. Since individually fine-tuning the ASR or LLM often leads to suboptimal results due to modality-specific limitations, end-to-end training ensures seamless integration of audio features and language understanding (cross-modal alignment), ultimately enabling a more personalized and efficient adaptation on edge devices. However, due to the complex training requirements and substantial computational demands of existing approaches, cross-modal alignment between ASR audio and LLM can be challenging on edge devices. In this work, we propose a resource-efficient cross-modal alignment framework that bridges ASR and LLMs on edge devices to handle personalized audio input. Our framework enables efficient ASR-LLM alignment on resource-constrained devices like NVIDIA Jetson Orin (8GB RAM), achieving 50x training time speedup while improving the alignment quality by more than 50\%. To the best of our knowledge, this is the first work to study efficient ASR-LLM alignment on resource-constrained edge devices., Comment: 7 pages, 8 figures

This paper proposes an ultra-low power, mixed-bit-width sparse convolutional neural network (CNN) accelerator to accelerate ventricular arrhythmia (VA) detection. The chip achieves 50% sparsity in a quantized 1D CNN using a sparse processing element (SPE) architecture. Measurement on the prototype chip TSMC 40nm CMOS low-power (LP) process for the VA classification task demonstrates that it consumes 10.60 $\mu$W of power while achieving a performance of 150 GOPS and a diagnostic accuracy of 99.95%. The computation power density is only 0.57 $\mu$W/mm$^2$, which is 14.23X smaller than state-of-the-art works, making it highly suitable for implantable and wearable medical devices., Comment: 2 pages, accepted to The 30th Asia and South Pacific Design Automation Conference (ASP-DAC 2025)

Content addressable memory (CAM) stands out as an efficient hardware solution for memory-intensive search operations by supporting parallel computation in memory. However, developing a CAM-based accelerator architecture that achieves acceptable accuracy, while minimizing hardware cost and catering to both exact and approximate search, still presents a significant challenge especially when considering a broader spectrum of applications. This complexity stems from CAM's rapid evolution across multiple levels--algorithms, architectures, circuits, and underlying devices. This paper introduces CAMASim, a first comprehensive CAM accelerator simulation framework, emphasizing modularity, flexibility, and generality. CAMASim establishes the detailed design space for CAM-based accelerators, incorporates automated functional simulation for accuracy, and enables hardware performance prediction, by leveraging a circuit-level CAM modeling tool. This work streamlines the design space exploration for CAM-based accelerator, aiding researchers in developing effective CAM-based accelerators for various search-intensive applications.

Earth Mover's Distance (EMD) is an important similarity measure between two distributions, used in computer vision and many other application domains. However, its exact calculation is computationally and memory intensive, which hinders its scalability and applicability for large-scale problems. Various approximate EMD algorithms have been proposed to reduce computational costs, but they suffer lower accuracy and may require additional memory usage or manual parameter tuning. In this paper, we present a novel approach, NNS-EMD, to approximate EMD using Nearest Neighbor Search (NNS), in order to achieve high accuracy, low time complexity, and high memory efficiency. The NNS operation reduces the number of data points compared in each NNS iteration and offers opportunities for parallel processing. We further accelerate NNS-EMD via vectorization on GPU, which is especially beneficial for large datasets. We compare NNS-EMD with both the exact EMD and state-of-the-art approximate EMD algorithms on image classification and retrieval tasks. We also apply NNS-EMD to calculate transport mapping and realize color transfer between images. NNS-EMD can be 44x to 135x faster than the exact EMD implementation, and achieves superior accuracy, speedup, and memory efficiency over existing approximate EMD methods.

In this work, we propose SEE-MCAM, scalable and compact multi-bit CAM (MCAM) designs that utilize the three-terminal ferroelectric FET (FeFET) as the proxy. By exploiting the multi-level-cell characteristics of FeFETs, our proposed SEE-MCAM designs enable multi-bit associative search functions and achieve better energy efficiency and performance than existing FeFET-based CAM designs. We validated the functionality of our proposed designs by achieving 3 bits per cell CAM functionality, resulting in 3x improvement in storage density. The area per bit of the proposed SEE-MCAM cell is 8% of the conventional CMOS CAM. We thoroughly investigated the scalability and robustness of the proposed design. Evaluation results suggest that the proposed 2FeFET-1T SEE-MCAM achieves 9.8x more energy efficiency and 1.6x less search latency compared to the CMOS CAM, respectively. When compared to existing MCAM designs, the proposed SEE-MCAM can achieve 8.7x and 4.9x more energy efficiency than ReRAM-based and FeFET-based MCAMs, respectively. Benchmarking results show that our approach provides up to 3 orders of magnitude improvement in speedup and energy efficiency over a GPU implementation in accelerating a novel quantized hyperdimensional computing (HDC) application., Comment: Accepted by Internation Conference on Computer-Aided Design (ICCAD), 2023

Field Programmable Gate Array (FPGA) is widely used in acceleration of deep learning applications because of its reconfigurability, flexibility, and fast time-to-market. However, conventional FPGA suffers from the tradeoff between chip area and reconfiguration latency, making efficient FPGA accelerations that require switching between multiple configurations still elusive. In this paper, we perform technology-circuit-architecture co-design to break this tradeoff with no additional area cost and lower power consumption compared with conventional designs while providing dynamic reconfiguration, which can hide the reconfiguration time behind the execution time. Leveraging the intrinsic transistor structure and non-volatility of ferroelectric FET (FeFET), compact FPGA primitives are proposed and experimentally verified, including 1FeFET look-up table (LUT) cell, 1FeFET routing cell for connection blocks (CBs) and switch boxes (SBs). To support dynamic reconfiguration, two local copies of primitives are placed in parallel, which enables loading of arbitrary configuration without interrupting the active configuration execution. A comprehensive evaluation shows that compared with the SRAM-based FPGA, our dynamic reconfiguration design shows 63.0%/71.1% reduction in LUT/CB area and 82.7%/53.6% reduction in CB/SB power consumption with minimal penalty in the critical path delay (9.6%). We further implement a Super-Sub network model to show the benefit from the context-switching capability of our design. We also evaluate the timing performance of our design over conventional FPGA in various application scenarios. In one scenario that users switch between two preloaded configurations, our design yields significant time saving by 78.7% on average. In the other scenario of implementing multiple configurations with dynamic reconfiguration, our design offers time saving of 20.3% on average., Comment: 54 pages, 15 figures

Automatic classification of pigmented, non-pigmented, and depigmented non-melanocytic skin lesions have garnered lots of attention in recent years. However, imaging variations in skin texture, lesion shape, depigmentation contrast, lighting condition, etc. hinder robust feature extraction, affecting classification accuracy. In this paper, we propose a new deep neural network that exploits input data for robust feature extraction. Specifically, we analyze the convolutional network's behavior (field-of-view) to find the location of deep supervision for improved feature extraction. To achieve this, first, we perform activation mapping to generate an object mask, highlighting the input regions most critical for classification output generation. Then the network layer whose layer-wise effective receptive field matches the approximated object shape in the object mask is selected as our focus for deep supervision. Utilizing different types of convolutional feature extractors and classifiers on three melanoma detection datasets and two vitiligo detection datasets, we verify the effectiveness of our new method., Comment: MICCAI 2022

Experience replay is an essential component in deep reinforcement learning (DRL), which stores the experiences and generates experiences for the agent to learn in real time. Recently, prioritized experience replay (PER) has been proven to be powerful and widely deployed in DRL agents. However, implementing PER on traditional CPU or GPU architectures incurs significant latency overhead due to its frequent and irregular memory accesses. This paper proposes a hardware-software co-design approach to design an associative memory (AM) based PER, AMPER, with an AM-friendly priority sampling operation. AMPER replaces the widely-used time-costly tree-traversal-based priority sampling in PER while preserving the learning performance. Further, we design an in-memory computing hardware architecture based on AM to support AMPER by leveraging parallel in-memory search operations. AMPER shows comparable learning performance while achieving 55x to 270x latency improvement when running on the proposed hardware compared to the state-of-the-art PER running on GPU., Comment: 9 pages, 9 figures. The work was accepted by the 41st International Conference on Computer-Aided Design (ICCAD), 2022, San Diego

Lifelong on-device learning is a key challenge for machine intelligence, and this requires learning from few, often single, samples. Memory augmented neural network has been proposed to achieve the goal, but the memory module has to be stored in an off-chip memory due to its size. Therefore the practical use has been heavily limited. Previous works on emerging memory-based implementation have difficulties in scaling up because different modules with various structures are difficult to integrate on the same chip and the small sense margin of the content addressable memory for the memory module heavily limited the degree of mismatch calculation. In this work, we implement the entire memory augmented neural network architecture in a fully integrated memristive crossbar platform and achieve an accuracy that closely matches standard software on digital hardware for the Omniglot dataset. The successful demonstration is supported by implementing new functions in crossbars in addition to widely reported matrix multiplications. For example, the locality-sensitive hashing operation is implemented in crossbar arrays by exploiting the intrinsic stochasticity of memristor devices. Besides, the content-addressable memory module is realized in crossbars, which also supports the degree of mismatches. Simulations based on experimentally validated models show such an implementation can be efficiently scaled up for one-shot learning on the Mini-ImageNet dataset. The successful demonstration paves the way for practical on-device lifelong learning and opens possibilities for novel attention-based algorithms not possible in conventional hardware., Comment: 54 pages, 21 figures, 3 tables

Compression is a standard procedure for making convolutional neural networks (CNNs) adhere to some specific computing resource constraints. However, searching for a compressed architecture typically involves a series of time-consuming training/validation experiments to determine a good compromise between network size and performance accuracy. To address this, we propose an image complexity-guided network compression technique for biomedical image segmentation. Given any resource constraints, our framework utilizes data complexity and network architecture to quickly estimate a compressed model which does not require network training. Specifically, we map the dataset complexity to the target network accuracy degradation caused by compression. Such mapping enables us to predict the final accuracy for different network sizes, based on the computed dataset complexity. Thus, one may choose a solution that meets both the network size and segmentation accuracy requirements. Finally, the mapping is used to determine the convolutional layer-wise multiplicative factor for generating a compressed network. We conduct experiments using 5 datasets, employing 3 commonly-used CNN architectures for biomedical image segmentation as representative networks. Our proposed framework is shown to be effective for generating compressed segmentation networks, retaining up to $\approx 95\%$ of the full-sized network segmentation accuracy, and at the same time, utilizing $\approx 32x$ fewer network trainable weights (average reduction) of the full-sized networks., Comment: ACM JETC

Hyperdimensional Computing (HDC) is an emerging computational framework that mimics important brain functions by operating over high-dimensional vectors, called hypervectors (HVs). In-memory computing implementations of HDC are desirable since they can significantly reduce data transfer overheads. All existing in-memory HDC platforms consider binary HVs where each dimension is represented with a single bit. However, utilizing multi-bit HVs allows HDC to achieve acceptable accuracies in lower dimensions which in turn leads to higher energy efficiencies. Thus, we propose a highly accurate and efficient multi-bit in-memory HDC inference platform called MIMHD. MIMHD supports multi-bit operations using ferroelectric field-effect transistor (FeFET) crossbar arrays for multiply-and-add and FeFET multi-bit content-addressable memories for associative search. We also introduce a novel hardware-aware retraining framework (HWART) that trains the HDC model to learn to work with MIMHD. For six popular datasets and 4000 dimension HVs, MIMHD using 3-bit (2-bit) precision HVs achieves (i) average accuracies of 92.6% (88.9%) which is 8.5% (4.8%) higher than binary implementations; (ii) 84.1x (78.6x) energy improvement over a GPU, and (iii) 38.4x (34.3x) speedup over a GPU, respectively. The 3-bit $\times$ is 4.3x and 13x faster and more energy-efficient than binary HDC accelerators while achieving similar accuracies., Comment: Accepted at ISLPED 2021

The memory wall bottleneck is a key challenge across many data-intensive applications. Multi-level FeFET-based embedded non-volatile memories are a promising solution for denser and more energy-efficient on-chip memory. However, reliable multi-level cell storage requires careful optimizations to minimize the design overhead costs. In this work, we investigate the interplay between FeFET device characteristics, programming schemes, and memory array architecture, and explore different design choices to optimize performance, energy, area, and accuracy metrics for critical data-intensive workloads. From our cross-stack design exploration, we find that we can store DNN weights and social network graphs at a density of over 8MB/mm^2 and sub-2ns read access latency without loss in application accuracy., Comment: Accepted at ISLPED 2021

From diagnosing neovascular diseases to detecting white matter lesions, accurate tiny vessel segmentation in fundus images is critical. Promising results for accurate vessel segmentation have been known. However, their effectiveness in segmenting tiny vessels is still limited. In this paper, we study retinal vessel segmentation by incorporating tiny vessel segmentation into our framework for the overall accurate vessel segmentation. To achieve this, we propose a new deep convolutional neural network (CNN) which divides vessel segmentation into two separate objectives. Specifically, we consider the overall accurate vessel segmentation and tiny vessel segmentation as two individual objectives. Then, by exploiting the objective-dependent (homoscedastic) uncertainty, we enable the network to learn both objectives simultaneously. Further, to improve the individual objectives, we propose: (a) a vessel weight map based auxiliary loss for enhancing tiny vessel connectivity (i.e., improving tiny vessel segmentation), and (b) an enhanced encoder-decoder architecture for improved localization (i.e., for accurate vessel segmentation). Using 3 public retinal vessel segmentation datasets (CHASE_DB1, DRIVE, and STARE), we verify the superiority of our proposed framework in segmenting tiny vessels (8.3% average improvement in sensitivity) while achieving better area under the receiver operating characteristic curve (AUC) compared to state-of-the-art methods., Comment: ISBI 2021

Nearest neighbor (NN) search is an essential operation in many applications, such as one/few-shot learning and image classification. As such, fast and low-energy hardware support for accurate NN search is highly desirable. Ternary content-addressable memories (TCAMs) have been proposed to accelerate NN search for few-shot learning tasks by implementing $L_\infty$ and Hamming distance metrics, but they cannot achieve software-comparable accuracies. This paper proposes a novel distance function that can be natively evaluated with multi-bit content-addressable memories (MCAMs) based on ferroelectric FETs (FeFETs) to perform a single-step, in-memory NN search. Moreover, this approach achieves accuracies comparable to floating-point precision implementations in software for NN classification and one/few-shot learning tasks. As an example, the proposed method achieves a 98.34% accuracy for a 5-way, 5-shot classification task for the Omniglot dataset (only 0.8% lower than software-based implementations) with a 3-bit MCAM. This represents a 13% accuracy improvement over state-of-the-art TCAM-based implementations at iso-energy and iso-delay. The presented distance function is resilient to the effects of FeFET device-to-device variations. Furthermore, this work experimentally demonstrates a 2-bit implementation of FeFET MCAM using AND arrays from GLOBALFOUNDRIES to further validate proof of concept., Comment: To be published in DATE'21

With the advance in mobile computing, Internet of Things, and ubiquitous wireless connectivity, social sensing based edge computing (SSEC) has emerged as a new computation paradigm where people and their personally owned devices collect sensor measurements from the physical world and process them at the edge of the network. This paper focuses on a privacy-aware task allocation problem where the goal is to optimize the computation task allocation in SSEC systems while respecting the users' customized privacy settings. It introduces a novel Game-theoretic Privacy-aware Task Allocation (G-PATA) framework to achieve the goal. G-PATA includes (i) a bottom-up game-theoretic model to generate the maximum payoffs at end devices while satisfying the end user's privacy settings; (ii) a top-down incentive scheme to adjust the rewards for the tasks to ensure that the task allocation decisions made by end devices meet the Quality of Service (QoS) requirements of the applications. Furthermore, the framework incorporates an efficient load balancing and iteration reduction component to adapt to the dynamic changes in status and privacy configurations of end devices. The G-PATA framework was implemented on a real-world edge computing platform that consists of heterogeneous end devices (Jetson TX1 and TK1 boards, and Raspberry Pi3). We compare G-PATA with state-of-the-art task allocation schemes through two real-world social sensing applications. The results show that G-PATA significantly outperforms existing approaches under various privacy settings (our scheme achieved as much as 47% improvements in delay reduction for the application and 15% more payoffs for end devices compared to the baselines.).

An analog synapse circuit based on ferroelectric-metal field-effect transistors is proposed, that offers 6-bit weight precision. The circuit is comprised of volatile least significant bits (LSBs) used solely during training, and non-volatile most significant bits (MSBs) used for both training and inference. The design works at a 1.8V logic-compatible voltage, provides 10^10 endurance cycles, and requires only 250ps update pulses. A variant of LeNet trained with the proposed synapse achieves 98.2% accuracy on MNIST, which is only 0.4% lower than an ideal implementation of the same network with the same bit precision. Furthermore, the proposed synapse offers improvements of up to 26% in area, 44.8% in leakage power, 16.7% in LSB update pulse duration, and two orders of magnitude in endurance cycles, when compared to state-of-the-art hybrid synaptic circuits. Our proposed synapse can be extended to an 8-bit design, enabling a VGG-like network to achieve 88.8% accuracy on CIFAR-10 (only 0.8% lower than an ideal implementation of the same network)., Comment: Accepted at ISCAS'20 for oral presentation

We propose a technology-independent method, referred to as adjacent connection matrix (ACM), to efficiently map signed weight matrices to non-negative crossbar arrays. When compared to same-hardware-overhead mapping methods, using ACM leads to improvements of up to 20% in training accuracy for ResNet-20 with the CIFAR-10 dataset when training with 5-bit precision crossbar arrays or lower. When compared with strategies that use two elements to represent a weight, ACM achieves comparable training accuracies, while also offering area and read energy reductions of 2.3x and 7x, respectively. ACM also has a mild regularization effect that improves inference accuracy in crossbar arrays without any retraining or costly device/variation-aware training., Comment: Accepted at DAC'20

A new spintronic nonvolatile memory cell analogous to 1T DRAM with non-destructive read is proposed. The cells can be used as neural computing units. A dual-circuit neural network architecture is proposed to leverage these devices against the complex operations involved in convolutional networks. Simulations based on HSPICE and Matlab were performed to study the performance of this architecture when classifying images as well as the effect of varying the size and stability of the nanomagnets. The spintronic cells outperform a purely charge-based implementation of the same network, consuming about 100 pJ total per image processed.

As cost and performance benefits associated with Moore's Law scaling slow, researchers are studying alternative architectures (e.g., based on analog and/or spiking circuits) and/or computational models (e.g., convolutional and recurrent neural networks) to perform application-level tasks faster, more energy efficiently, and/or more accurately. We investigate cellular neural network (CeNN)-based co-processors at the application-level for these metrics. While it is well-known that CeNNs can be well-suited for spatio-temporal information processing, few (if any) studies have quantified the energy/delay/accuracy of a CeNN-friendly algorithm and compared the CeNN-based approach to the best von Neumann algorithm at the application level. We present an evaluation framework for such studies. As a case study, a CeNN-friendly target-tracking algorithm was developed and mapped to an array architecture developed in conjunction with the algorithm. We compare the energy, delay, and accuracy of our architecture/algorithm (assuming all overheads) to the most accurate von Neumann algorithm (Struck). Von Neumann CPU data is measured on an Intel i5 chip. The CeNN approach is capable of matching the accuracy of Struck, and can offer approximately 1000x improvements in energy-delay product.

Convolutional neural networks (CNNs) for biomedical image analysis are often of very large size, resulting in high memory requirement and high latency of operations. Searching for an acceptable compressed representation of the base CNN for a specific imaging application typically involves a series of time-consuming training/validation experiments to achieve a good compromise between network size and accuracy. To address this challenge, we propose CC-Net, a new image complexity-guided CNN compression scheme for biomedical image segmentation. Given a CNN model, CC-Net predicts the final accuracy of networks of different sizes based on the average image complexity computed from the training data. It then selects a multiplicative factor for producing a desired network with acceptable network accuracy and size. Experiments show that CC-Net is effective for generating compressed segmentation networks, retaining up to 95% of the base network segmentation accuracy and utilizing only 0.1% of trainable parameters of the full-sized networks in the best case., Comment: Updated FM energy dist. figure

Deep neural network (DNN) accelerators with improved energy and delay are desirable for meeting the requirements of hardware targeted for IoT and edge computing systems. Convolutional neural networks (CoNNs) belong to one of the most popular types of DNN architectures. This paper presents the design and evaluation of an accelerator for CoNNs. The system-level architecture is based on mixed-signal, cellular neural networks (CeNNs). Specifically, we present (i) the implementation of different layers, including convolution, ReLU, and pooling, in a CoNN using CeNN, (ii) modified CoNN structures with CeNN-friendly layers to reduce computational overheads typically associated with a CoNN, (iii) a mixed-signal CeNN architecture that performs CoNN computations in the analog and mixed signal domain, and (iv) design space exploration that identifies what CeNN-based algorithm and architectural features fare best compared to existing algorithms and architectures when evaluated over common datasets -- MNIST and CIFAR-10. Notably, the proposed approach can lead to 8.7$\times$ improvements in energy-delay product (EDP) per digit classification for the MNIST dataset at iso-accuracy when compared with the state-of-the-art DNN engine, while our approach could offer 4.3$\times$ improvements in EDP when compared to other network implementations for the CIFAR-10 dataset., Comment: 25 pages

The 55th Design Automation Conference (DAC) held its first System Design Contest (SDC) in 2018. SDC'18 features a lower power object detection challenge (LPODC) on designing and implementing novel algorithms based object detection in images taken from unmanned aerial vehicles (UAV). The dataset includes 95 categories and 150k images, and the hardware platforms include Nvidia's TX2 and Xilinx's PYNQ Z1. DAC-SDC'18 attracted more than 110 entries from 12 countries. This paper presents in detail the dataset and evaluation procedure. It further discusses the methods developed by some of the entries as well as representative results. The paper concludes with directions for future improvements., Comment: 12 pages, 21 figures