Yukai Song

Yukai Song

Auto-generated profile, based on publicly available information

Yukai Song

University of Pittsburgh, Pittsburgh, USA

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
29 June 2025
Verified
Conference Paper
Open access
SJShixin Ji
Shixin Ji
XCXingzhen Chen
Xingzhen Chen
JZJinming Zhuang
Jinming Zhuang
WZWei Zhang
Wei Zhang
ZYZhuoping Yang
Zhuoping Yang
PZPeipei Zhou
Peipei Zhou

6 November 2024 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Verified
Journal Article
Closed access
YTYue Tang
Yue Tang
YSYukai Song
Yukai Song
NENaveena Elango
Naveena Elango
SPSheena Ratnam Priya
Sheena Ratnam Priya
AJAlex K. Jones
Alex K. Jones
Jingtong HuJingtong Hu
Jingtong Hu
Deep neural networks (DNNs) are rapidly evolving from streamlined single-modality single-task (SMST) to multimodality multitask (MMMT) with large variations for different layers and complex data dependencies among layers. To support such models, hardware systems also evolved to be heterogeneous. The heterogeneous system comes from the prevailing trend to integrate diverse accelerators into the system for lower latency. FPGAs have high-computation density and communication bandwidth and are configurable to be deployed with different designs of accelerators, which are widely used for various machine-learning applications. However, scaling from SMST to MMMT on heterogeneous FPGAs is challenging since MMMT has much larger layer variations, a massive number of layers, and complex data dependency among different backbones. Previous mapping algorithms are either inefficient or over-simplified which makes them impractical in general scenarios. In this work, we propose CHEF to enable efficient implementation of MMMT models in realistic heterogeneous FPGA clusters, i.e., deploying heterogeneous accelerators on heterogeneous FPGAs (A2F) and mapping the heterogeneous DNNs on the deployed heterogeneous accelerators (M2A). We propose CHEF-A2F, a two-stage accelerators-to-FPGAs deployment approach to co-optimize hardware deployment and accelerator mapping. In addition, we propose CHEF-M2A, which can support general and practical cases compared to previous mapping algorithms. To the best of our knowledge, this is the first attempt to implement MMMT models in real heterogeneous FPGA clusters. Experimental results show that the latency obtained with CHEF is near-optimal while the search time is 10 $000\times $ less than exhaustively searching the optimal solution.

27 February 2025
Verified
Conference Paper
Open access
SJShixin Ji
Shixin Ji
XCXingzhen Chen
Xingzhen Chen
WZWei Zhang
Wei Zhang
ZYZhuoping Yang
Zhuoping Yang
JZJinming Zhuang
Jinming Zhuang
PZPeipei Zhou
Peipei Zhou
Real-time requirements are essential for safety-critical systems such as autonomous driving and robotics. In these systems, schedulability - ensuring every task completes before its deadline - is critical, as missing even a single deadline can lead to catastrophic consequences. Unlike Quality-of-Service (QoS) or service-level agreement (SLA) metrics, where occasional deadline misses are tolerable, time-predictable safety-critical systems demand strict adherence to deadlines. Existing multi-task scheduling frameworks, optimized for QoS/SLA, fail to meet these requirements, even when incorporating real-time techniques like preemption and scheduling. In FPGA-based systems, accelerators are widely used to handle real-time computational tasks. While existing accelerators focus on low latency and high throughput, these optimizations are insufficient when multiple tasks share a single accelerator, as low latency alone does not ensure schedulability. For example in Fig.1(a), a system with tasks of 20 ms and 400 ms periods, requiring 10 ms and 200 ms execution times, will miss deadlines on a 1 TOPS accelerator. Even increasing throughput to 10 TOPS (Fig.1(b)) without proper scheduling fails to resolve the issue. Furthermore, without a theoretical analysis, designers cannot determine whether the system after latency optimization can meet the deadline. As a result, these latency-optimized accelerators can not be safely used in time-predictable, safety-critical real-time systems. To address these challenges, we propose the ART framework: Accelerator Customization for Real-Time Systems. ART enables FPGA accelerators to support Earliest Deadline First (EDF) scheduling and limited preemption mechanisms, specifically targeting real-time safety-critical systems. Alongside hardware support, ART provides a software framework to analyze schedulability and balance between schedulability and performance. As shown in Figure 1(c), a 1 TOPS accelerator with ART can satisfy the schedulability requirement. Whereas the latency-optimized accelerator without ART in Figure 1(b) can not satisfy and thus is not viable in this scenario. Our contributions are summarized as follows: Necessity analysis of scheduling and preemption mechanism: We conduct a detailed case study to understand the necessity of scheduling and preemption. ART hardware framework and accelerator system: We propose the ART hardware framework and accelerator system for EDF scheduling and limited preemption on FPGA. ART can transform a baseline HLS accelerator to support these features. ART software framework: To guarantee the schedulability in real-time scenarios, we propose the ART software framework. We migrate the schedulability analysis and optimal preemption point placements to hardware acceleration, explain their effectiveness and package these algorithms into an automated framework. ART implementations and experiments: We implement and evaluate the ART accelerator system in AMD Zynq ZCU102, and Versal VCK190 platforms. The onboard experiment shows that ART guarantees the schedulability whereas the corresponding baseline accelerators can not. ART is lightweight, utilizing < 0.4% of resources on ZCU102 and < 0.1% on VCK190. Necessity analysis of scheduling and preemption mechanism: We conduct a detailed case study to understand the necessity of scheduling and preemption. ART hardware framework and accelerator system: We propose the ART hardware framework and accelerator system for EDF scheduling and limited preemption on FPGA. ART can transform a baseline HLS accelerator to support these features. ART software framework: To guarantee the schedulability in real-time scenarios, we propose the ART software framework. We migrate the schedulability analysis and optimal preemption point placements to hardware acceleration, explain their effectiveness and package these algorithms into an automated framework. ART implementations and experiments: We implement and evaluate the ART accelerator system in AMD Zynq ZCU102, and Versal VCK190 platforms. The onboard experiment shows that ART guarantees the schedulability whereas the corresponding baseline accelerators can not. ART is lightweight, utilizing < 0.4% of resources on ZCU102 and < 0.1% on VCK190.

19 March 2026 in JMIR Preprints
Verified
Preprint
Open access
YSYukai Song
Yukai Song
PZPengfei Zhou
Pengfei Zhou
CECésar Gabriel Escobar-Viera
César Gabriel Escobar-Viera
CBCandice Biernesser
Candice Biernesser
WHWei Huang
Wei Huang
Jingtong HuJingtong Hu
Jingtong Hu
BACKGROUND Suicide rates have risen worldwide in recent years, underscoring the urgent need for proactive and scalable prevention strategies. Social media has emerged as a valuable source for early screening, as many at-risk individuals who avoid formal help due to stigma express distress online. However, current suicide risk detection systems face a critical limitation: while explicit suicidal expressions are relatively easy to identify, implicit suicidal ideation conveyed through metaphor, sarcasm, or subtle emotional cues remains difficult to detect and constitutes a major source of system failure in real-world deployment. OBJECTIVE This study aims to develop a robust and computationally efficient suicide risk detection framework that can reliably handle both explicit and implicit expressions while controlling inference cost and improving practical deployability for mental health screening. METHODS We propose a robust and efficient two-stage framework that explicitly accounts for heterogeneous signal strength through selective escalation. In Stage 1, a lightweight Bidirectional Encoder Representations from Transformers (BERT) classifier rapidly resolves high-confidence explicit cases. In Stage 2, only ambiguous inputs are escalated for further analysis through two complementary mechanisms: (i) a multi-perspective large language model (LLM) voting strategy designed to improve recall on implicit suicide risk, and (ii) a feature-based machine learning ensemble guided by psychologically grounded indicators extracted via prompt-engineered LLMs, enabling efficient inference and enhanced interpretability. RESULTS We evaluate the proposed framework on two complementary datasets representing distinct risk profiles: an explicit-dominant Reddit dataset and an implicit-only DeepSuiMind dataset. The framework achieves an F1 score of 98.0% on the Reddit benchmark and up to 99.94% recall on the implicit-only DeepSuiMind benchmark (positive-only; precision is not defined). It reduces the cross-domain performance gap to approximately 1.2-1.4% in recall and 1.7-2.2% in F1, and significantly lowers LLM invocation costs compared with single-model baselines. CONCLUSIONS These results demonstrate that system-level design with selective escalation offers a practical and effective pathway for suicide risk detection systems that are both robust to implicit expressions and efficient enough for real-world mental health screening applications. CLINICALTRIAL N/A

3 July 2026 in ACM Transactions on Cyber-Physical Systems
Verified
Journal Article
Open access
JCJingxian Chen
Jingxian Chen
YSYukai Song
Yukai Song
SXShengjie Xu
Shengjie Xu
KKKyle Kessler
Kyle Kessler
PGPrateek Ganguli
Prateek Ganguli
Jingtong HuJingtong Hu
Jingtong Hu
Traditional neural architecture pruning methods primarily focus on trade-offs between inference accuracy and computational cost ( e.g. , FLOPs), without accounting for system-level behaviors. In the context of cyber-physical systems (CPS), perception models and the downstream controllers that rely on their outputs are typically designed independently. In this paper, we study the joint design of neural architecture pruning strategies and the controllers that utilize the resulting inference outputs. We propose two key innovations: (i) incorporating system-level safety metrics into the pruning optimization loop, and (ii) a regression-based scaling law method that efficiently models the relationship between pruning ratios and system-level safety outcomes. This enables rapid pruning guided by system-level constraints, avoiding the inefficiency of exhaustive search or conventional neural architecture search approaches that focus on local accuracy or latency metrics. We demonstrate our approach on YOLO-based perception models deployed in control systems for distance estimation. To the best of our knowledge, this is the first work to introduce a scaling-law-guided pruning method for YOLO-based perception models, offering a principled design automation framework for safety-aware, machine learning–enabled CPS.