Selected work

Representative Papers

HAMP-LIC: Hessian-Aware Mixed-Precision Post-Training Quantization for Learned Image Compression

Aug 12, 2026

This work addresses the significant degradation in reconstruction quality and cross-platform codec inconsistency observed in existing neural image compression models when deployed at low bitrates, primarily due to the neglect of inter-layer quantization sensitivity variations. To overcome these limitations, the authors propose a four-stage mixed-precision post-training quantization (PTQ) framework that introduces Hessian trace-based block-level sensitivity estimation for the first time. This approach integrates task-aware fine-tuning, global constraint-guided bit-width allocation, and block-level reconstruction, thereby completely eliminating cross-platform discrepancies. Evaluated on the Minnen2018 and Cheng2020 models, the method achieves up to a 4.85× compression ratio with only a 0.59% BD-rate loss, substantially outperforming current fixed- and mixed-precision PTQ baselines.

0 citationsRead paper

DTMC-Based Analysis and Scheduling for Periodic Flows with Proactive HARQ

Aug 06, 2026

This work addresses the challenge of schedulability analysis for heterogeneous periodic traffic in ultra-reliable low-latency communication (URLLC) systems, where proactive HARQ introduces slot-level timing effects—such as feedback delay—that complicate resource allocation. To tackle this, the paper proposes a discrete-time Markov chain (DTMC)-based modeling framework that accurately captures cross-slot dynamics, including HARQ round-trip latency, through an expanded state space. Coupled with a two-stage genetic algorithm, the approach optimizes offset scheduling to meet diverse reliability and latency requirements. This study is the first to apply DTMCs to timing modeling of periodic flows under proactive HARQ, enabling precise schedulability analysis that explicitly accounts for feedback delay. Simulations demonstrate that the proposed method significantly improves schedulability compared to reactive HARQ, K-repetition schemes, and non-guaranteed proactive HARQ, while maintaining manageable computational overhead.

0 citationsRead paper

Recent publications

Latest Papers

HAMP-LIC: Hessian-Aware Mixed-Precision Post-Training Quantization for Learned Image Compression

Aug 12, 2026

This work addresses the significant degradation in reconstruction quality and cross-platform codec inconsistency observed in existing neural image compression models when deployed at low bitrates, primarily due to the neglect of inter-layer quantization sensitivity variations. To overcome these limitations, the authors propose a four-stage mixed-precision post-training quantization (PTQ) framework that introduces Hessian trace-based block-level sensitivity estimation for the first time. This approach integrates task-aware fine-tuning, global constraint-guided bit-width allocation, and block-level reconstruction, thereby completely eliminating cross-platform discrepancies. Evaluated on the Minnen2018 and Cheng2020 models, the method achieves up to a 4.85× compression ratio with only a 0.59% BD-rate loss, substantially outperforming current fixed- and mixed-precision PTQ baselines.

0 citationsRead paper

DTMC-Based Analysis and Scheduling for Periodic Flows with Proactive HARQ

Aug 06, 2026

This work addresses the challenge of schedulability analysis for heterogeneous periodic traffic in ultra-reliable low-latency communication (URLLC) systems, where proactive HARQ introduces slot-level timing effects—such as feedback delay—that complicate resource allocation. To tackle this, the paper proposes a discrete-time Markov chain (DTMC)-based modeling framework that accurately captures cross-slot dynamics, including HARQ round-trip latency, through an expanded state space. Coupled with a two-stage genetic algorithm, the approach optimizes offset scheduling to meet diverse reliability and latency requirements. This study is the first to apply DTMCs to timing modeling of periodic flows under proactive HARQ, enabling precise schedulability analysis that explicitly accounts for feedback delay. Simulations demonstrate that the proposed method significantly improves schedulability compared to reactive HARQ, K-repetition schemes, and non-guaranteed proactive HARQ, while maintaining manageable computational overhead.

0 citationsRead paper