Chien-Ming Chen

AAAI Conference 2026 Conference Paper

Efficient Multimodal Large Language Model via Dynamic KV Cache Quantization

• Jiahao Fan
• Chien-Ming Chen

Multimodal large language models (LMMs) have demonstrated remarkable capabilities across diverse vision-language tasks, including image captioning, visual question answering, and text-image retrieval. However, their computational complexity and memory footprint, particularly in the key-value (KV) cache during inference, pose significant challenges for real-time deployment, especially on resource-constrained devices. In this paper, we propose Dynamic KV Cache Quantization, a novel quantization strategy tailored for multimodal LMMs. Our approach applies per-channel quantization to (K) and per-token quantization to (V), leveraging their respective statistical distributions to optimize precision allocation. Additionally, we introduce an adaptive token and channel recording mechanism that dynamically adjusts quantization parameters based on real-time distribution tracking, effectively mitigating the impact of outliers. To further enhance compression efficiency, we implement fine-grained grouping, which partitions KV tensors into localized subgroups, enabling more adaptive quantization. Experimental results on LLaVA-1.5 (7B/13B) and Qwen-VL across multiple multimodal benchmarks demonstrate that our method significantly outperforms existing KV-cache quantization approaches, achieving a superior trade-off between memory efficiency and model accuracy.

EAAI Journal 2025 Journal Article

Advancing explainability of adversarial trained Convolutional Neural Networks for robust engineering applications

• Dehua Zhou
• Ziyu Song
• Zicong Chen
• Xianting Huang
• Congming Ji
• Saru Kumari
• Chien-Ming Chen
• Sachin Kumar

The great success of Convolutional Neural Networks has driven their widespread use in Industry 5. 0, especially in safety-critical applications where model robustness is essential. While adversarial training is commonly used to make models more robust, it often changes how the model fuses features, leading to complex trade-offs and making its behavior harder to understand. This lack of transparency contradicts the demand for interpretable and trustworthy artificial intelligence systems in modern industry, where applications like autonomous driving and intelligent medical systems require model interpretability to assign post-accident responsibility and support legal and liability decisions. Our work introduces an innovative approach to enhance the interpretability of adversarially trained Convolutional Neural Networks, providing detailed and quantifiable insights through a six-dimensional semantic framework. This framework, which examines objects, parts, scenes, materials, textures, and colors, offers fine-grained explanations that closely mirror human cognitive processes. We conduct extensive experiments across multiple classical models, and the results reveal that adversarial training induces structure-dependent changes in feature preferences, offering insights into the trade-off between standard and robust accuracy in adversarially trained Convolutional Neural Networks. These findings provide rational support for the use of artificial intelligence in real-world applications and promote the development and implementation of explainable artificial intelligence in modern industries.

JBHI Journal 2025 Journal Article

Enhancing Large Medical Equipment Health-Aware Control with Bayesian Graph Attention Transformer-Based Probabilistic Remaining Useful Life Prediction

• Huamao Jiang
• Keqin Li
• Zhonghua Liu
• Chien-Ming Chen
• Jianhui Lv

Large medical equipment such as MRI, CT scanners, and linear accelerators are critical healthcare infrastructure. The health status of these systems directly impacts diagnHealth-Aware Control osis accuracy, treatment efficacy, and overall healthcare quality. However, current maintenance approaches often rely on fixed schedules or reactive strategies, leading to unnecessary downtime, increased costs, and potential patient care disruptions. This lack of uncertain information hinders robust decision-making in health-aware control (HAC) and maintenance scheduling. Therefore, this paper proposes a novel probabilistic remaining useful life (RUL) prediction method for large medical equipment based on a Bayesian graph attention transformer. First, we construct a graph attention network to extract complex non-Euclidean spatial relationships between multiple sensor signals from medical equipment. We then integrate this graph attention mechanism into the transformer's temporal multi-head attention module, combining graph attention networks' spatial feature extraction advantages with transformer models' temporal feature extraction capabilities. This integration enables joint extraction of spatiotemporal relationships in sensor data. Simultaneously, we employ an improved Bayesian network to quantify prediction uncertainty, providing point estimates of RUL and corresponding confidence intervals. The resulting framework supports HAC decisions that balance large medical equipment performance and longevity. Experiments conducted on two medical equipment operation datasets (CT and MRI) demonstrate the effectiveness and superiority of our approach compared to existing methods, achieving more accurate and reliable RUL predictions that enable optimized maintenance scheduling and enhanced HAC strategies.

JBHI Journal 2025 Journal Article

Guest Editorial: On Advancing Healthcare Informatics With Large Language Models

• Saru Kumari
• Chien-Ming Chen
• Mohammad Shojafar
• Muhammad Naveed Aman

JBHI Journal 2025 Journal Article

SeqNovo: De Novo Peptide Sequencing Prediction in IoMT via Seq2Seq

• Ke Wang
• Mingjia Zhu
• Wadii Boulila
• Maha Driss
• Thippa Reddy Gadekallu
• Chien-Ming Chen
• Lei Wang
• Saru Kumari

In the Internet of Medical Things (IoMT), de novo peptide sequencing prediction is one of the most important techniques for the fields of disease prediction, diagnosis, and treatment. Recently, deep-learning-based peptide sequencing prediction has been a new trend. However, most popular deep learning models for peptide sequencing prediction suffer from poor interpretability and poor ability to capture long-range dependencies. To solve these issues, we propose a model named SeqNovo, which has the encoding-decoding structure of sequence to sequence (Seq2Seq), the highly nonlinear properties of multilayer perceptron (MLP), and the ability of the attention mechanism to capture long-range dependencies. SeqNovo use MLP to improve the feature extraction and utilize the attention mechanism to discover key information. A series of experiments have been conducted to show that the SeqNovo is superior to the Seq2Seq benchmark model, DeepNovo. SeqNovo improves both the accuracy and interpretability of the predictions, which will be expected to support more related research.

EAAI Journal 2022 Journal Article

On-shelf utility mining from transaction database

• Jiahui Chen
• Xu Guo
• Wensheng Gan
• Chien-Ming Chen
• Weiping Ding
• Guoting Chen

As an important technique for dealing with transaction database in the field of data mining, utility-driven mining can be used to discover useful patterns (i. e. , itemsets, sequences) which have a high utility. However, it has a bias towards the item/object combinations which have more exhibition period since they have more opportunity to generate a high utility. To address this, the on-shelf time period of items need to be considered, thus on-shelf utility mining (OSUM) can be applied in the application which is more closer to the actual situation. Currently several models have been proposed to deal with the OSUM problem, but they still suffer from the requirement that it needs to maintain a massive candidates in memory and to scan database many times. In this paper, we propose two effective one-phase algorithms named OSUMI (On-Shelf Utility Mining from transactIon database) and OSUMI + (the improve version of OSUMI). Both OSUMI and OSUMI + search all itemsets as a set-enumeration tree and discover the on-shelf itemsets with high utility in a more practical way. More precisely, in order to avoid the problems of high memory consumption, two algorithms apply some properties of the concept of on-shelf utility. Besides, two upper-bounds named subtree utility and local utility are applied to early filter out unpromising patterns and then prune the search space. Finally, an extensive experimental study on several real on-shelf datasets shows that our proposed algorithms can be significantly faster than the state-of-the-art algorithm.