Qinyu Chen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Qinyu Chen has a total of 115 co-authors. Here are the top co-authors by number of shared publications:
- Chang Gao
Department of Microelectronics, Delft University of Technology, Delft, The Netherlands
7 shared publications Chang Gao
4 shared publications- Li Li
Nanjing University
3 shared publications - Yuxiang Fu
Nanjing University
3 shared publications - Bojun Cheng
The Hong Kong University of Science and Technology (Guangzhou)
2 shared publications - Congyi Sun
School of Electronic Science and Engineering, Nanjing University, Nanjing, China
2 shared publications
Latest Publications
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. PP (10638210) , 1-14
QSNNA: An Energy-Efficient Quaternary Spiking Neural Network Accelerator for Seizure Detection
This work proposes a novel quaternary spiking neural network (QSNN) paradigm and its corresponding energy-efficient hardware accelerator to address the efficiency–accuracy tradeoffs in edge computing. Unlike traditional binary spiking neural networks (SNNs), the proposed architecture leverages multilevel spike magnitudes to enhance temporal representation density. To optimize hardware implementation, we introduce a low-complexity temporal feature fusion encoding (TFFE) scheme and an adaptive spike suppression mechanism (ASSM). Specifically, the TFFE scheme extracts time-domain features in time-series signal processing tasks to achieve efficient spike encoding, thereby circumventing complex encoding overheads. Furthermore, the ASSM dynamically exploits intrinsic neural sparsity to skip redundant computations, allowing for configurable trade-offs between inference latency and power consumption. Based on this algorithm–hardware codesign, the QSNN accelerator (QSNNA) is implemented in a 40-nm CMOS process. We validate the reliability and efficiency of this architecture through an epileptic seizure detection task using the CHB-MIT database. Experimental results demonstrate that the accelerator achieves a minimum energy consumption of $0.066~\mu $ J per inference and occupies a compact core area of 0.21 mm2, representing a $\gt\!3.4\times $ reduction in energy compared to state-of-the-art (SOTA) accelerators. Simultaneously, the system attains high classification performance with accuracy $\geq 97.40$ %, sensitivity $\geq 96.74$ %, and specificity $\geq 97.88$ %, outperforming baseline methods.
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. PP (10638210) , 1-14
QSNNA: An Energy-Efficient Quaternary Spiking Neural Network Accelerator for Seizure Detection
This work proposes a novel quaternary spiking neural network (QSNN) paradigm and its corresponding energy-efficient hardware accelerator to address the efficiency–accuracy tradeoffs in edge computing. Unlike traditional binary spiking neural networks (SNNs), the proposed architecture leverages multilevel spike magnitudes to enhance temporal representation density. To optimize hardware implementation, we introduce a low-complexity temporal feature fusion encoding (TFFE) scheme and an adaptive spike suppression mechanism (ASSM). Specifically, the TFFE scheme extracts time-domain features in time-series signal processing tasks to achieve efficient spike encoding, thereby circumventing complex encoding overheads. Furthermore, the ASSM dynamically exploits intrinsic neural sparsity to skip redundant computations, allowing for configurable trade-offs between inference latency and power consumption. Based on this algorithm–hardware codesign, the QSNN accelerator (QSNNA) is implemented in a 40-nm CMOS process. We validate the reliability and efficiency of this architecture through an epileptic seizure detection task using the CHB-MIT database. Experimental results demonstrate that the accelerator achieves a minimum energy consumption of $0.066~\mu $ J per inference and occupies a compact core area of 0.21 mm2, representing a $\gt\!3.4\times $ reduction in energy compared to state-of-the-art (SOTA) accelerators. Simultaneously, the system attains high classification performance with accuracy $\geq 97.40$ %, sensitivity $\geq 96.74$ %, and specificity $\geq 97.88$ %, outperforming baseline methods.
- Open Access
Eye Gaze Tells You Where to Compute: Gaze-Driven Efficient VLMs
Vision-Language Models (VLMs) deliver impressive performance in understanding visual content with language instructions. However, redundancy in vision tokens results in the degenerated inference efficiency of VLMs, which hinders real-time use on edge consumer devices such as AR/VR devices. Existing efficiency methods commonly prune visual tokens using learned saliency, sparse attention schedules, or controller policies, but they often require architectural modification or access to intermediate activations. These pipelines add inference-time modules that increase compute and memory and often lead to an accuracy trade-off. Moreover, they also suffer from misalignment between the prompts and the region of interest in the images. Without human guidance, the model may focus on the wrong regions and miss small, high-frequency details when prompts or scenes change. In this paper, we propose GazeVLM, a training-free framework that uses the human eye gaze as a natural supervisory signal to allocate computation where it matters. By extracting gaze-driven regions of interest (ROIs) and optionally combining them with a low-resolution global view, GazeVLM mimics fovea-periphery perception to cut redundant visual tokens while preserving task-relevant details. We evaluate the visual question answering tasks on Qwen2.5-VL-3B/7B on the VOILA-COCO benchmark with human gaze. Quality of the answer is assessed by GPT-4o pairwise judging and a weighted score over coverage, accuracy, details, and fluency. Efficiency is measured by token counts and FLOPs. GazeVLM reduces visual tokens by up to 93.1%, total tokens by up to 59.6%, and FLOPs by 50%, while keeping better answer quality relative to full-resolution baselines. Our results show that aligning model computation with human gaze offers a simple, plug-and-play path toward efficient VLM inference on consumer devices.
Open AccessEye Gaze Tells You Where to Compute: Gaze-Driven Efficient VLMs
Vision-Language Models (VLMs) deliver impressive performance in understanding visual content with language instructions. However, redundancy in vision tokens results in the degenerated inference efficiency of VLMs, which hinders real-time use on edge consumer devices such as AR/VR devices. Existing efficiency methods commonly prune visual tokens using learned saliency, sparse attention schedules, or controller policies, but they often require architectural modification or access to intermediate activations. These pipelines add inference-time modules that increase compute and memory and often lead to an accuracy trade-off. Moreover, they also suffer from misalignment between the prompts and the region of interest in the images. Without human guidance, the model may focus on the wrong regions and miss small, high-frequency details when prompts or scenes change. In this paper, we propose GazeVLM, a training-free framework that uses the human eye gaze as a natural supervisory signal to allocate computation where it matters. By extracting gaze-driven regions of interest (ROIs) and optionally combining them with a low-resolution global view, GazeVLM mimics fovea-periphery perception to cut redundant visual tokens while preserving task-relevant details. We evaluate the visual question answering tasks on Qwen2.5-VL-3B/7B on the VOILA-COCO benchmark with human gaze. Quality of the answer is assessed by GPT-4o pairwise judging and a weighted score over coverage, accuracy, details, and fluency. Efficiency is measured by token counts and FLOPs. GazeVLM reduces visual tokens by up to 93.1%, total tokens by up to 59.6%, and FLOPs by 50%, while keeping better answer quality relative to full-resolution baselines. Our results show that aligning model computation with human gaze offers a simple, plug-and-play path toward efficient VLM inference on consumer devices.
- 11 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (11) , 4002-4017
An 8.62-μW 75-dB DR SoC Fully Integrated SoC for Spoken Language Understanding
We present a sub-10- $\mu $ W fully integrated SoC for on-device spoken language understanding (SLU). Its analog feature extractor (FEx) applies global and per-channel automatic gain control (AGC) to extend the system’s dynamic range (DR)—a critical requirement for real-world scenarios, including far-field operations. The on-chip streaming-mode recurrent neural network (RNN) accelerator exploits temporal sparsity and pooling, reducing its power by $2.3\times $ . By combining hardware-aware training with a behavioral model of the FEx that captures circuit nonidealities, the network is trained to maintain SLU accuracy despite chip-to-chip variation. Fabricated in a 65-nm CMOS process, the SoC occupies 2.23 mm2 and consumes 8.62 $\mu $ W for end-to-end SLU. The 16-channel FEx achieves 93-dB DR while dissipating 1.85 $\mu $ W at 100-Hz feature frame rate. The SoC is evaluated on the 32-class Fluent Speech Commands dataset (FSCD), achieving 92.9% accuracy for 2.8-mVrms inputs while maintaining >85% accuracy over a 75-dB input range.
- 11 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (11) , 4002-4017
An 8.62-μW 75-dB DR SoC Fully Integrated SoC for Spoken Language Understanding
We present a sub-10- $\mu $ W fully integrated SoC for on-device spoken language understanding (SLU). Its analog feature extractor (FEx) applies global and per-channel automatic gain control (AGC) to extend the system’s dynamic range (DR)—a critical requirement for real-world scenarios, including far-field operations. The on-chip streaming-mode recurrent neural network (RNN) accelerator exploits temporal sparsity and pooling, reducing its power by $2.3\times $ . By combining hardware-aware training with a behavioral model of the FEx that captures circuit nonidealities, the network is trained to maintain SLU accuracy despite chip-to-chip variation. Fabricated in a 65-nm CMOS process, the SoC occupies 2.23 mm2 and consumes 8.62 $\mu $ W for end-to-end SLU. The 16-channel FEx achieves 93-dB DR while dissipating 1.85 $\mu $ W at 100-Hz feature frame rate. The SoC is evaluated on the 32-class Fluent Speech Commands dataset (FSCD), achieving 92.9% accuracy for 2.8-mVrms inputs while maintaining >85% accuracy over a 75-dB input range.
- 18 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (2) , 1093-1106
A 1.1 μJ/Inference Binary Spiking Neural Network Accelerator for DVS Gesture Recognition
Dynamic vision sensors (DVS) are bioinspired sensors that can generate sparse data streams with low latency, low power consumption and high dynamic range. Spiking neural networks (SNNs), which are inspired by biological brains, are event-based models, and therefore they can be used to process the binary data streams produced by such sensors naturally. However, SNN accelerators usually require more memory and longer time for inference, due to the extra time dimension in SNNs. In this paper, an energy efficient binary spiking neural network (BSNN) accelerator for DVS gesture recognition is proposed with algorithm and hardware codesign. We integrate the binary neural network (BNN) training method into SNN to directly train a BSNN model, which significantly reduces memory consumption. A temporal pooling (TP) layer is further proposed to reduce the time steps in SNNs while maintaining competitive accuracy. The proposed BSNN accelerator can achieve high parallelism with high resource utilization, and the sparsity of input spikes is utilized to further reduce power consumption. The proposed BSNN model achieves an accuracy of 95.49% on IBM DVS Gesture dataset. The implementation results show that the BSNN accelerator can achieve 38.2k inference per second with $1.1~\mu $ J/inference energy consumption and 216.9 TOPS/W energy efficiency.
- 18 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (2) , 1093-1106
A 1.1 μJ/Inference Binary Spiking Neural Network Accelerator for DVS Gesture Recognition
Dynamic vision sensors (DVS) are bioinspired sensors that can generate sparse data streams with low latency, low power consumption and high dynamic range. Spiking neural networks (SNNs), which are inspired by biological brains, are event-based models, and therefore they can be used to process the binary data streams produced by such sensors naturally. However, SNN accelerators usually require more memory and longer time for inference, due to the extra time dimension in SNNs. In this paper, an energy efficient binary spiking neural network (BSNN) accelerator for DVS gesture recognition is proposed with algorithm and hardware codesign. We integrate the binary neural network (BNN) training method into SNN to directly train a BSNN model, which significantly reduces memory consumption. A temporal pooling (TP) layer is further proposed to reduce the time steps in SNNs while maintaining competitive accuracy. The proposed BSNN accelerator can achieve high parallelism with high resource utilization, and the sparsity of input spikes is utilized to further reduce power consumption. The proposed BSNN model achieves an accuracy of 95.49% on IBM DVS Gesture dataset. The implementation results show that the BSNN accelerator can achieve 38.2k inference per second with $1.1~\mu $ J/inference energy consumption and 216.9 TOPS/W energy efficiency.
- Open Access
DeltaLLM: A Training-Free Framework Exploiting Temporal Sparsity for Efficient Edge LLM Inference
Deploying Large Language Models (LLMs) on edge devices remains challenging due to their quadratically increasing computations with the sequence length. Existing studies for dynamic attention pruning are designed for hardware with massively parallel computation capabilities, such as GPUs or TPUs, and aim at long context lengths (e.g., 64K), making them unsuitable for edge scenarios. We present DeltaLLM, a training-free framework that exploits temporal sparsity in attention patterns to enable efficient LLM inference across both the prefilling and decoding stages, on resource-constrained edge devices. DeltaLLM introduces an accuracy- and memory-aware delta matrix construction strategy that introduces temporal sparsity, and a context-aware hybrid attention mechanism that combines full attention in a local context window with delta approximation outside it to increase accuracy. We evaluate our framework on the edge-device-friendly BitNet-b1.58-2B-4T model and Llama3.2-1B-Instruct model across diverse language tasks. The results show that on BitNet, our framework increases the attention sparsity from 0% to 60% during the prefilling stage with slight accuracy improvement on the WG task, and 0% to 57% across both the prefilling and decoding stages, with even higher F1 score from 29.63 to 30.97 on SQuAD-v2 task. On the Llama model, it can also achieve up to 60% sparsity during the prefilling stage and around 57% across both stages with negligible accuracy drop. These results demonstrate that DeltaLLM offers a promising solution for efficient edge deployment, requiring no fine-tuning and seamlessly integrating with existing inference pipelines.
Open AccessDeltaLLM: A Training-Free Framework Exploiting Temporal Sparsity for Efficient Edge LLM Inference
Deploying Large Language Models (LLMs) on edge devices remains challenging due to their quadratically increasing computations with the sequence length. Existing studies for dynamic attention pruning are designed for hardware with massively parallel computation capabilities, such as GPUs or TPUs, and aim at long context lengths (e.g., 64K), making them unsuitable for edge scenarios. We present DeltaLLM, a training-free framework that exploits temporal sparsity in attention patterns to enable efficient LLM inference across both the prefilling and decoding stages, on resource-constrained edge devices. DeltaLLM introduces an accuracy- and memory-aware delta matrix construction strategy that introduces temporal sparsity, and a context-aware hybrid attention mechanism that combines full attention in a local context window with delta approximation outside it to increase accuracy. We evaluate our framework on the edge-device-friendly BitNet-b1.58-2B-4T model and Llama3.2-1B-Instruct model across diverse language tasks. The results show that on BitNet, our framework increases the attention sparsity from 0% to 60% during the prefilling stage with slight accuracy improvement on the WG task, and 0% to 57% across both the prefilling and decoding stages, with even higher F1 score from 29.63 to 30.97 on SQuAD-v2 task. On the Llama model, it can also achieve up to 60% sparsity during the prefilling stage and around 57% across both stages with negligible accuracy drop. These results demonstrate that DeltaLLM offers a promising solution for efficient edge deployment, requiring no fine-tuning and seamlessly integrating with existing inference pipelines.
- Open Access
SHAP-AAD: DeepSHAP-Guided Channel Reduction for EEG Auditory Attention Detection
Electroencephalography (EEG)-based auditory attention detection (AAD) offers a non-invasive way to enhance hearing aids, but conventional methods rely on too many electrodes, limiting wearability and comfort. This paper presents SHAP-AAD, a two-stage framework that combines DeepSHAP-based channel selection with a lightweight temporal convolutional network (TCN) for efficient AAD using fewer channels.DeepSHAP, an explainable AI technique, is applied to a Convolutional Neural Network (CNN) trained on topographic alpha-power maps to rank channel importance, and the top-k EEG channels are used to train a compact TCN. Experiments on the DTU dataset show that using 32 channels yields comparable accuracy to the full 64-channel setup (79.21% vs. 81.06%) on average. In some cases, even 8 channels can deliver satisfactory accuracy. These results demonstrate the effectiveness of SHAP-AAD in reducing complexity while preserving high detection performance.
Open AccessSHAP-AAD: DeepSHAP-Guided Channel Reduction for EEG Auditory Attention Detection
Electroencephalography (EEG)-based auditory attention detection (AAD) offers a non-invasive way to enhance hearing aids, but conventional methods rely on too many electrodes, limiting wearability and comfort. This paper presents SHAP-AAD, a two-stage framework that combines DeepSHAP-based channel selection with a lightweight temporal convolutional network (TCN) for efficient AAD using fewer channels.DeepSHAP, an explainable AI technique, is applied to a Convolutional Neural Network (CNN) trained on topographic alpha-power maps to rank channel importance, and the top-k EEG channels are used to train a compact TCN. Experiments on the DTU dataset show that using 32 channels yields comparable accuracy to the full 64-channel setup (79.21% vs. 81.06%) on average. In some cases, even 8 channels can deliver satisfactory accuracy. These results demonstrate the effectiveness of SHAP-AAD in reducing complexity while preserving high detection performance.
- 16 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Microwave and Wireless Technology Letters
- Vol. 35 (6) , 772-775
DeltaDPD: Exploiting Dynamic Temporal Sparsity in Recurrent Neural Networks for Energy-Efficient Wideband Digital Predistortion
Digital predistortion (DPD) is a popular technique to enhance signal quality in wideband radio frequency (RF) power amplifiers (PAs). With increasing bandwidth and data rates, DPD faces significant energy consumption challenges during deployment, contrasting with its efficiency goals. State-of-the-art DPD models rely on recurrent neural networks (RNNs), whose computational complexity hinders system efficiency. This letter introduces DeltaDPD, exploring the dynamic temporal sparsity of input signals and neuronal hidden states in RNNs for energy-efficient DPD, reducing arithmetic operations and memory accesses while preserving satisfactory linearization performance. Applying a TM3.1a 200 MHz-BW 256-QAM OFDM signal to a 3.5-GHz GaN Doherty RF PA, DeltaDPD achieves −50.03 dBc in adjacent channel power ratio (ACPR), −37.22dB in normalized mean square error (NMSE) and −38.52 dB in error vector magnitude (EVM) with 52% temporal sparsity, leading to a $1.8\times $ reduction in estimated inference power. The DeltaDPD code is available at https://www.opendpd.com.
- 16 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Microwave and Wireless Technology Letters
- Vol. 35 (6) , 772-775
DeltaDPD: Exploiting Dynamic Temporal Sparsity in Recurrent Neural Networks for Energy-Efficient Wideband Digital Predistortion
Digital predistortion (DPD) is a popular technique to enhance signal quality in wideband radio frequency (RF) power amplifiers (PAs). With increasing bandwidth and data rates, DPD faces significant energy consumption challenges during deployment, contrasting with its efficiency goals. State-of-the-art DPD models rely on recurrent neural networks (RNNs), whose computational complexity hinders system efficiency. This letter introduces DeltaDPD, exploring the dynamic temporal sparsity of input signals and neuronal hidden states in RNNs for energy-efficient DPD, reducing arithmetic operations and memory accesses while preserving satisfactory linearization performance. Applying a TM3.1a 200 MHz-BW 256-QAM OFDM signal to a 3.5-GHz GaN Doherty RF PA, DeltaDPD achieves −50.03 dBc in adjacent channel power ratio (ACPR), −37.22dB in normalized mean square error (NMSE) and −38.52 dB in error vector magnitude (EVM) with 52% temporal sparsity, leading to a $1.8\times $ reduction in estimated inference power. The DeltaDPD code is available at https://www.opendpd.com.
- Open Access
Event-Based Eye Tracking. 2025 Event-based Vision Workshop
This survey serves as a review for the 2025 Event-Based Eye Tracking Challenge organized as part of the 2025 CVPR event-based vision workshop. This challenge focuses on the task of predicting the pupil center by processing event camera recorded eye movement. We review and summarize the innovative methods from teams rank the top in the challenge to advance future event-based eye tracking research. In each method, accuracy, model size, and number of operations are reported. In this survey, we also discuss event-based eye tracking from the perspective of hardware design.
Open AccessEvent-Based Eye Tracking. 2025 Event-based Vision Workshop
This survey serves as a review for the 2025 Event-Based Eye Tracking Challenge organized as part of the 2025 CVPR event-based vision workshop. This challenge focuses on the task of predicting the pupil center by processing event camera recorded eye movement. We review and summarize the innovative methods from teams rank the top in the challenge to advance future event-based eye tracking research. In each method, accuracy, model size, and number of operations are reported. In this survey, we also discuss event-based eye tracking from the perspective of hardware design.
- 1 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72 (10) , 5748-5760
Toward Efficient Eye Tracking in AR/VR Devices: A Near-Eye DVS-Based Processor for Real-Time Gaze Estimation
This paper presents an efficient near-eye dynamic vision sensor (DVS)-based processor for real-time eye tracking in augmented reality/virtual reality (AR/VR) devices. The processor takes advantage of the sparse event data with fine time resolution from the DVS, addressing the need for high frame-rate, low-power, and accurate eye tracking on wearable devices with extended battery life. Exploiting the inherent sparsity of event data, we propose an event-density-based region of interest (ROI) determination method that operates directly on event stream, which requires 47× fewer operations than the traditional methods, effectively overcoming the latency problem caused by the heavy computational loads. To eliminate the issue of decreasing accuracy at the edges of the field of view (FoV), we customized and fine-tuned a neural network for gaze estimation, ensuring uniformly distributed sub-degree accuracy. An estimator with a streamlined output mapping strategy and an adaptive window-sliding convolution scheme is implemented for gaze estimation acceleration. The processor is designed and fabricated in UMC 40-nm LP technology with a core area of 2.52 mm2 and performs end-to-end eye tracking exclusively with the raw event stream from DVS, achieving an average accuracy of 0.91∘ within a 96∘×64∘ FoV. Operating at 200 MHz, it achieves a dynamic frame rate of up to 1.2 kHz and requires only 12.7 μJ of energy per gaze estimation. By integrating the DVS, the processor enables real-time, low-power, and accurate eye tracking, enhancing the immersive experience on AR/VR devices and offering intuitive and seamless interactions.
- 1 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72 (10) , 5748-5760
Toward Efficient Eye Tracking in AR/VR Devices: A Near-Eye DVS-Based Processor for Real-Time Gaze Estimation
This paper presents an efficient near-eye dynamic vision sensor (DVS)-based processor for real-time eye tracking in augmented reality/virtual reality (AR/VR) devices. The processor takes advantage of the sparse event data with fine time resolution from the DVS, addressing the need for high frame-rate, low-power, and accurate eye tracking on wearable devices with extended battery life. Exploiting the inherent sparsity of event data, we propose an event-density-based region of interest (ROI) determination method that operates directly on event stream, which requires 47× fewer operations than the traditional methods, effectively overcoming the latency problem caused by the heavy computational loads. To eliminate the issue of decreasing accuracy at the edges of the field of view (FoV), we customized and fine-tuned a neural network for gaze estimation, ensuring uniformly distributed sub-degree accuracy. An estimator with a streamlined output mapping strategy and an adaptive window-sliding convolution scheme is implemented for gaze estimation acceleration. The processor is designed and fabricated in UMC 40-nm LP technology with a core area of 2.52 mm2 and performs end-to-end eye tracking exclusively with the raw event stream from DVS, achieving an average accuracy of 0.91∘ within a 96∘×64∘ FoV. Operating at 200 MHz, it achieves a dynamic frame rate of up to 1.2 kHz and requires only 12.7 μJ of energy per gaze estimation. By integrating the DVS, the processor enables real-time, low-power, and accurate eye tracking, enhancing the immersive experience on AR/VR devices and offering intuitive and seamless interactions.