Author name cluster

MingYu Lu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

10 papers

2 author rows

ICLR Conference 2025 Conference Paper

An Efficient Framework for Crediting Data Contributors of Diffusion Models

Mingyu Lu
Chris Lin
Chanwoo Kim 0002
Su-In Lee

As diffusion models are deployed in real-world settings and their performance driven by training data, appraising the contribution of data contributors is crucial to creating incentives for sharing quality data and to implementing policies for data compensation. Depending on the use case, model performance corresponds to various global properties of the distribution learned by a diffusion model (e.g., overall aesthetic quality). Hence, here we address the problem of attributing global properties of diffusion models to data contributors. The Shapley value provides a principled approach to valuation by uniquely satisfying game-theoretic axioms of fairness. However, estimating Shapley values for diffusion models is computationally impractical because it requires retraining and rerunning inference on many subsets of data contributors. We introduce a method to efficiently retrain and rerun inference for Shapley value estimation, by leveraging model pruning and fine-tuning. We evaluate the utility of our method with three use cases: (i) image quality for a DDPM trained on a CIFAR dataset, (ii) demographic diversity for an LDM trained on CelebA-HQ, and (iii) aesthetic quality for a Stable Diffusion model LoRA-finetuned on Post-Impressionist artworks. Our results empirically demonstrate that our framework can identify important data contributors across global properties, outperforming existing attribution methods for diffusion models.

Details

NeurIPS Conference 2025 Conference Paper

CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning

MingYu Lu
Ethan Weinberger
Chanwoo Kim
Su-In Lee

High-content screening (HCS) assays based on high-throughput microscopy techniques such as Cell Painting have enabled the interrogation of cells' morphological responses to perturbations at an unprecedented scale. The collection of such data promises to facilitate a better understanding of the relationships between different perturbations and their effects on cellular state. Towards achieving this goal, recent advances in cross-modal contrastive learning could, in theory, be leveraged to learn a unified latent space that aligns perturbations with their corresponding morphological effects. However, the application of such methods to HCS data is not straightforward due to substantial differences in the semantics of Cell Painting images compared to natural images, and the difficulty of representing different classes of perturbations (e. g. small molecule vs CRISPR gene knockout) in a single latent space. In response to these challenges, here we introduce CellCLIP, a cross-modal contrastive learning framework for HCS data. CellCLIP leverages pre-trained image encoders coupled with a novel channel encoding scheme to better capture relationships between different microscopy channels in image embeddings, along with natural language encoders for representing perturbations. Our framework outperforms current open-source models, demonstrating the best performance in both cross-modal retrieval and biologically meaningful downstream tasks while also achieving significant reductions in computation time. Code for our reproducing our experiments is available at https: //github. com/suinleelab/CellCLIP.

PDF Details

AIIM Journal 2025 Journal Article

Fuzzy-DDI: A robust fuzzy logic query model for complex drug–drug interaction prediction

Junkai Cheng
Yijia Zhang
Hengyi Zhang
MingYu Lu

Drug–drug interactions (DDI) refer to the compound effects that occur when patients take multiple drugs simultaneously, which may reduce the drug efficacy and even harm the patient’s health. Therefore, DDI prediction is significant for drug development and safe medication. Despite the great efforts of researchers, existing methods mainly focus on predicting interactions between drug pairs, cannot contain more biomedical information, and have poor robustness, limiting their application in real-world scenarios. Therefore, we propose a new robust fuzzy logic query model, Fuzzy-DDI, to predict DDI under various complex conditions. Specifically, Fuzzy-DDI decomposes DDI predictions into relational projections and logical operations on rough sets during inference. Fuzzy logic makes it more fault-tolerant than binary logic models. We explore the reasoning ability of the model in a more realistic and meaningful DDI prediction task with target cell type information and explore the robustness of Fuzzy-DDI in noisy environments and missing sample environments. Experiments on three benchmark datasets show that Fuzzy-DDI significantly outperforms state-of-the-art methods on various DDI prediction tasks, demonstrating its capabilities in inference and robustness. The data and code are available at https: //github. com/Cheng0829/Fuzzy-DDI.

Details DOI

IJCAI Conference 2025 Conference Paper

RRG-Mamba: Efficient Radiology Report Generation with State Space Model

Xiaodi Hou
Xiaobo Li
MingYu Lu
Simiao Wang
Yijia Zhang

Recent advancements in radiology report generation have utilized deep neural networks such as CNNs and Transformers, achieving notable improvements in generating accurate and detailed reports. However, their practical adoption is hindered by the challenge of balancing global dependency modeling with computational efficiency. The state space model, particularly its enhanced variant Mamba, offers promising linear-complexity solutions for long-range dependency modeling. Despite its strengths, Mamba’s fixed positional encoding limits its ability to effectively capture complex spatial dependencies. To address this gap, we propose RRG-Mamba, an advanced framework for efficient radiology report generation. Within the RRGMamba, we enhance the vanilla Mamba by integrating rotary position encoding (RoPE), enabling dynamic modeling of relative positional information in visual feature sequences. Furthermore, we design a global dependency learning module to optimize long-range visual feature sequence modeling. Extensive experiments on publicly available datasets, including IU X-Ray and MIMIC-CXR, demonstrate that RRG-Mamba achieves a 3. 7% improvement in BLEU-4 score over existing models, along with significant gains in computational and memory efficiency. Our code is available at https: //github. com/Eleanorhxd/RRG-Mamba.

PDF Details DOI

JBHI Journal 2025 Journal Article

Semantic-Enhanced Graph Contrastive Learning With Adaptive Denoising for Drug Repositioning

Huimin Yu
MingYu Lu
Zeqian Li
Yijia Zhang

The traditional drug development process requires a significant investment in workforce and financial resources. Drug repositioning as an efficient alternative has attracted much attention during the last few years. Despite the wide application and success of the method, there are still many shortcomings in the existing model. For example, sparse datasets will seriously affect the existing methods' performance. Additionally, these methods do not pay attention to the noise in datasets. In response to the above defects, we propose a semantic-enriched augmented graph contrastive learning with an adaptive denoising method, called SGCD. This method enhances data from the perspective of the embedding layer, deeply mines potential neighborhood relation-ships in semantic space, and combines similar drugs in the semantic neighborhoods into prototype comparison targets, thus effectively mitigating the impact of data sparsity on the model. Moreover, to enhance the model's robustness to noisy data, we use the adaptive denoising method, which can effectively identify noisy data in the training process. Exhaustive experiments on multiple real datasets show the effectiveness of the proposed model.

Details DOI

EAAI Journal 2024 Journal Article

Latent domain knowledge distillation for nighttime semantic segmentation

Yunan Liu
Simiao Wang
Chunpeng Wang
MingYu Lu
Yu Sang

Despite significant progress made in image semantic segmentation, most research has primarily focused on daytime scenes. Semantic segmentation of nighttime images is equally critical for autonomous driving; however, this task poses more challenges due to inadequate lighting and difficulties associated with obtaining accurate manual annotations. In this paper, we propose a novel method called Latent Domain Knowledge Distillation (LDKD) for accurate nighttime semantic segmentation. Within a teacher-student framework, our LDKD combines domain alignment and knowledge distillation in a mutually reinforcing manner. First, we introduce a bidirectional photometric alignment module capable of generating daytime-like and nighttime-like latent images in real-time. This module effectively bridges the appearance gap between the source domain (daytime) and the target domain (nighttime). Second, we enhance collaboration between the teacher and student networks. The student network unifies image-level and feature-level domain alignment, while the teacher network generates reliable pseudo-labels by distilling knowledge from the latent domain. Furthermore, to account for potential noise in pseudo-labels, we propose a noise-tolerant learning method aimed at mitigating the risks associated with overreliance on pseudo-labels during model training. Extensive experiments on Dark Zurich and ACDC datasets show that our LDKD achieves state-of-the-art performance, demonstrating the effectiveness of our method for nighttime semantic segmentation.

Details DOI

JBHI Journal 2024 Journal Article

TransFOL: A Logical Query Model for Complex Relational Reasoning in Drug-Drug Interaction

Junkai Cheng
Yijia Zhang
Hengyi Zhang
Shaoxiong Ji
MingYu Lu

Predicting drug-drug interaction (DDI) plays a crucial role in drug recommendation and discovery. However, wet lab methods are prohibitively expensive and time-consuming due to drug interactions. In recent years, deep learning methods have gained widespread use in drug reasoning. Although these methods have demonstrated effectiveness, they can only predict the interaction between a drug pair and do not contain any other information. However, DDI is greatly affected by various other biomedical factors (such as the dose of the drug). As a result, it is challenging to apply them to more complex and meaningful reasoning tasks. Therefore, this study regards DDI as a link prediction problem on knowledge graphs and proposes a DDI prediction model based on Cross-Transformer and Graph Convolutional Networks (GCNs) in first-order logical query form, TransFOL. In the model, a biomedical query graph is first built to learn the embedding representation. Subsequently, an enhancement module is designed to aggregate the semantics of entities and relations. Cross-Transformer is used for encoding to obtain semantic information between nodes, and GCN is used to gather neighbour information further and predict inference results. To evaluate the performance of TransFOL on common DDI tasks, we conduct experiments on two benchmark datasets. The experimental results indicate that our model outperforms state-of-the-art methods on traditional DDI tasks. Additionally, we introduce different biomedical information in the other two experiments to make the settings more realistic. Experimental results verify the strong drug reasoning ability and generalization of TransFOL in complex settings.

Details DOI

ICML Conference 2023 Conference Paper

Learning to Maximize Mutual Information for Dynamic Feature Selection

Ian Connick Covert
Wei Qiu
Mingyu Lu
Nayoon Kim
Nathan J. White
Su-In Lee

Feature selection helps reduce data acquisition costs in ML, but the standard approach is to train models with static feature subsets. Here, we consider the dynamic feature selection (DFS) problem where a model sequentially queries features based on the presently available information. DFS is often addressed with reinforcement learning, but we explore a simpler approach of greedily selecting features based on their conditional mutual information. This method is theoretically appealing but requires oracle access to the data distribution, so we develop a learning approach based on amortized optimization. The proposed method is shown to recover the greedy policy when trained to optimality, and it outperforms numerous existing feature selection methods in our experiments, thus validating it as a simple but powerful approach for this problem.

Details

EAAI Journal 2023 Journal Article

TCAMixer: A lightweight Mixer based on a novel triple concepts attention mechanism for NLP

Xiaoyan Liu
Huanling Tang
Jie Zhao
Quansheng Dou
MingYu Lu

Large-scale model sizes and expensive computing costs cause the challenge of deploying and applying large pre-trained models. Hence, this paper presents a novel Triple Concepts Attention Mechanism and a lightweight TCAMixer model for edge devices to classify texts. Furthermore, the TCAMixer abstracts textual concepts in a human way, which is unmatched by other counterparts such as pNLP-Mixer (a projection-based MLP-Mixer model for Nature Language Processing) and HyperMixer (a hyper network using dynamic token-mixing layers). Experimental results on several public datasets demonstrate that the TCAMixer outperforms the counterparts by a significant margin, for example, achieving 3% higher accuracy with a smaller model size of 0. 177M. Additionally, the TCAMixer achieves a performance of 85% to 98. 7% compared to that of large pre-trained models but only occupies 1/3000 to 1/2000 of their size on most test datasets.

Details DOI

EAAI Journal 2021 Journal Article

AONet: Active Offset Network for crowd flow prediction

Dafeng Wang
Qian Ma
Naiyao Wang
Xuanzhe Fan
MingYu Lu
Hongbo Liu

Predicting crowd flow is of great importance to public safety and traffic management. The crowd flow is difficult to predict accurately and timely due to the uncertainty of the future positions. In this paper, we propose a novel Active Offset Network (AONet), in which ActiveGRU (Active Gate Recurrent Unit) is designed to predict the variation of pedestrians’ positions in the crowd flow. Its inner location-variant recurrent structure is implemented by utilizing convolution operation on low dimensional spatio-temporal sequences to obtain fractional offset locations. Afterwards, the sampling locations are determined by bilinear interpolation on fractional offset locations. Moreover, a probabilistic sparse strategy is introduced to reduce the links between sampling locations during supervised training. Finally, the experiments over popular benchmarks demonstrate that our method can actively characterize the future positions of pedestrians. Meanwhile, the performance of the proposed AONet is superior over state-of-art baselines with regard to both accuracy and computational savings.

Details DOI