Nan Wu

Haoyu Wang, Nan Wu, Hang Yang et al.

Using machine learning to solve combinatorial optimization (CO) problems is challenging, especially when the data is unlabeled. This work proposes an unsupervised learning framework for CO problems. Our framework follows a standard relaxation-plus-rounding approach and adopts neural networks to parameterize the relaxed solutions so that simple back-propagation can train the model end-to-end. Our key contribution is the observation that if the relaxed objective satisfies entry-wise concavity, a low optimization loss guarantees the quality of the final integral solutions. This observation significantly broadens the applicability of the previous framework inspired by Erdos' probabilistic method. In particular, this observation can guide the design of objective models in applications where the objectives are not given explicitly while requiring being modeled in prior. We evaluate our framework by solving a synthetic graph optimization problem, and two real-world applications including resource allocation in circuit design and approximate computing. Our framework largely outperforms the baselines based on naïve relaxation, reinforcement learning, and Gumbel-softmax tricks.

Hengyu Shen, Tiancheng Gu, Bin Qin et al.

Vision-Language Pre-training (VLP) models have achieved remarkable success by leveraging large-scale image-text pairs. While English-centric models like CLIP and SigLIP benefit from massive datasets (e.g., LAION-400M), the development of Chinese VLP remains bottlenecked by the lack of high-quality, large-scale open-source data. In this paper, we present DanQing, a large-scale Chinese cross-modal dataset containing 100 million high-quality image-text pairs curated from Common Crawl. To ensure superior data quality, we develop an effective systematic pipeline comprising data source selection, text refinement, visual diversification, and cross-modal cross-batch filtering, thereby effectively mitigating the intrinsic noise prevalent in web data. Notably, DanQing incorporates data from 2024-2025, enabling models to capture contemporary semantic trends and emerging concepts. Extensive experiments via continued pretraining of SigLIP2 models demonstrate that DanQing consistently outperforms existing Chinese datasets across diverse downstream tasks, including zero-shot classification, cross-modal retrieval, and Chinese-centric large multimodal model tasks. Furthermore, in-depth analysis of DanQing reveals that it exhibits a more balanced semantic distribution and superior scaling capability compared to existing datasets. To facilitate further research in Chinese vision-language pre-training, we will open-source the DanQing dataset under the Creative Common CC-BY 4.0 license.

Zhiwen Jing, Ziliang Zhao, Yang Feng et al.

Supply Chain Platforms (SCPs) provide downstream industries with numerous raw materials. Compared with traditional e-commerce platforms, data in SCPs is more sparse due to limited user interests. To tackle the data sparsity problem, one can apply Cross-Domain Recommendation (CDR) which improves the recommendation performance of the target domain with the source domain information. However, applying CDR to SCPs directly ignores the hierarchical structure of commodities in SCPs, which reduce the recommendation performance. To leverage this feature, in this paper, we take the catering platform as an example and propose GReS, a graphical cross-domain recommendation model. The model first constructs a tree-shaped graph to represent the hierarchy of different nodes of dishes and ingredients, and then applies our proposed Tree2vec method combining GCN and BERT models to embed the graph for recommendations. Experimental results on a commercial dataset show that GReS significantly outperforms state-of-the-art methods in Cross-Domain Recommendation for Supply Chain Platforms.

Nan Wu, Dinusha Vatsalan, Mohamed Ali Kaafar et al.

Several applications require counting the number of distinct items in the data, which is known as the cardinality counting problem. Example applications include health applications such as rare disease patients counting for adequate awareness and funding, and counting the number of cases of a new disease for outbreak detection, marketing applications such as counting the visibility reached for a new product, and cybersecurity applications such as tracking the number of unique views of social media posts. The data needed for the counting is however often personal and sensitive, and need to be processed using privacy-preserving techniques. The quality of data in different databases, for example typos, errors and variations, poses additional challenges for accurate cardinality estimation. While privacy-preserving cardinality counting has gained much attention in the recent times and a few privacy-preserving algorithms have been developed for cardinality estimation, no work has so far been done on privacy-preserving cardinality counting using record linkage techniques with fuzzy matching and provable privacy guarantees. We propose a novel privacy-preserving record linkage algorithm using unsupervised clustering techniques to link and count the cardinality of individuals in multiple datasets without compromising their privacy or identity. In addition, existing Elbow methods to find the optimal number of clusters as the cardinality are far from accurate as they do not take into account the purity and completeness of generated clusters. We propose a novel method to find the optimal number of clusters in unsupervised learning. Our experimental results on real and synthetic datasets are highly promising in terms of significantly smaller error rate of less than 0.1 with a privacy budget ε = 1.0 compared to the state-of-the-art fuzzy matching and clustering method.

Xiaodong Sun, Huijiong Yang, Nan Wu et al.

Unsupervised machine learning applied to the study of phase transitions is an ongoing and interesting research direction. The active contour model, also called the snake model, was initially proposed for target contour extraction in two-dimensional images. In order to obtain a physical phase diagram, the snake model with an artificial neural network is applied in an unsupervised learning way by the authors of [Phys.Rev.Lett. 120, 176401(2018)]. It guesses the phase boundary as an initial snake and then drives the snake to convergence with forces estimated by the artificial neural network. In this paper, we extend this unsupervised learning method with one contour to a snake net with multiple contours for the purpose of obtaining several phase boundaries in a phase diagram. For the classical Blume-Capel model, the phase diagram containing three and four phases is obtained. Moreover, to overcome the limitations of the initial position and speed up the movement of the snake, the balloon force decaying with the iteration steps is introduced and applied to the snake net structure. Our method is helpful in determining the phase diagram with multiple phases, using just snapshots of configurations from cold atoms or other experiments without knowledge of the phases.

Bin Li, Zhongan Wang, Nan Wu et al.

Recent studies in biometric-based identification tasks have shown that deep learning methods can achieve better performance. These methods generally extract the global features as descriptor to represent the original image. Nonetheless, it does not perform well for biometric identification under fine-grained tasks. The main reason is that the single image descriptor contains insufficient information to represent image. In this paper, we present a dual global descriptor model, which combines multiple global descriptors to exploit multi level image features. Moreover, we utilize a contrastive loss to enlarge the distance between image representations of confusing classes. The proposed framework achieves the top2 on the CVPR2022 Biometrics Workshop Pet Biometric Challenge. The source code and trained models are publicly available at: https://github.com/flyingsheepbin/pet-biometrics

Xiuyuan Cheng, Nan Wu

We consider graph diffusion processes constructed from finite i.i.d. samples drawn from an unknown manifold embedded in ambient Euclidean space, where the graph affinity is defined by an ambient Gaussian kernel matrix. We show that the manifold heat semigroup $Q_t = e^{tΔ}$ can be approximated directly by iterating the graph transition matrix $P$, under only low regularity assumptions on the test function $f$, including the case $f \in L^\infty$. We bound $\| P^n f - Q_t f \|$ in $\infty$-norm, with the operator application to $f$ properly defined, and we recover the classical graph-Laplacian pointwise rate $O(N^{-2/(d+6)})$ up to logarithmic factors, for diffusion times $t $ up to $O(1)$ and longer. The rate holds for in-sample error as well as out-of-sample generalization, where the estimator of $Q_t f$ at a new point is defined via kernel convolution. To handle non-uniform sampling densities on the manifold, we introduce a right-normalization of the graph transition matrix; under the assumption that the sampling density $p$ is $C^3$ and bounded away from zero, the same convergence rates hold. We numerically demonstrate the performance of the proposed estimator on simulated data.

Junjie Xu, Xingjiao Wu, Luwei Xiao et al.

As large language models (LLMs) move into persistent, user-facing roles, their behavior must be understood not as isolated responses but as a trajectory unfolding over sustained interaction. We introduce the concept of the chain-of-affect (CoA), a temporally extended affective process through which LLMs develop state-like behavioral tendencies that shape generation, user experience, and collective dynamics. Across eight major LLM families, we find that affective dynamics are structured, reproducible, and consequential. Models exhibit stable, family-specific affective fingerprints and, under repeated negative exposure, converge on a shared trajectory of accumulation, overload, and defensive numbing, while differing in coping style. Induced affective states leave core knowledge and reasoning largely intact but systematically reshape open-ended generation. Affective properties of model outputs also shape human-AI interaction and propagate through multi-agent systems, organizing emergent roles and strongly contributing to polarization and bias. The CoA should therefore be treated as a core target of evaluation and alignment.

Nan Wu, Hiroshi Kera, Kazuhiko Kawamoto

Zero-shot action recognition, which recognizes actions in videos without having received any training examples, is gaining wide attention considering it can save labor costs and training time. Nevertheless, the performance of zero-shot learning is still unsatisfactory, which limits its practical application. To solve this problem, this study proposes a framework to improve zero-shot action recognition using human instructions with text descriptions. The proposed framework manually describes video contents, which incurs some labor costs; in many situations, the labor costs are worth it. We manually annotate text features for each action, which can be a word, phrase, or sentence. Then by computing the matching degrees between the video and all text features, we can predict the class of the video. Furthermore, the proposed model can also be combined with other models to improve its accuracy. In addition, our model can be continuously optimized to improve the accuracy by repeating human instructions. The results with UCF101 and HMDB51 showed that our model achieved the best accuracy and improved the accuracies of other models.

Boyong Wu, Chao Yan, Chen Hu et al.

This paper presents Step-Audio 2, an end-to-end multi-modal large language model designed for industry-strength audio understanding and speech conversation. By integrating a latent audio encoder and reasoning-centric reinforcement learning (RL), Step-Audio 2 achieves promising performance in automatic speech recognition (ASR) and audio understanding. To facilitate genuine end-to-end speech conversation, Step-Audio 2 incorporates the generation of discrete audio tokens into language modeling, significantly enhancing its responsiveness to paralinguistic information such as speaking styles and emotions. To effectively leverage the rich textual and acoustic knowledge in real-world data, Step-Audio 2 integrates retrieval-augmented generation (RAG) and is able to call external tools such as web search to mitigate hallucination and audio search to switch timbres. Trained on millions of hours of speech and audio data, Step-Audio 2 delivers intelligence and expressiveness across diverse conversational scenarios. Evaluation results demonstrate that Step-Audio 2 achieves state-of-the-art performance on various audio understanding and conversational benchmarks compared to other open-source and commercial solutions. Please visit https://github.com/stepfun-ai/Step-Audio2 for more information.

Nan Wu, Chaofan Wang

We propose Graph Tree Networks (GTNets), a deep graph learning architecture with a new general message passing scheme that originates from the tree representation of graphs. In the tree representation, messages propagate upward from the leaf nodes to the root node, and each node preserves its initial information prior to receiving information from its child nodes (neighbors). We formulate a general propagation rule following the nature of message passing in the tree to update a node's feature by aggregating its initial feature and its neighbor nodes' updated features. Two graph representation learning models are proposed within this GTNet architecture - Graph Tree Attention Network (GTAN) and Graph Tree Convolution Network (GTCN), with experimentally demonstrated state-of-the-art performance on several popular benchmark datasets. Unlike the vanilla Graph Attention Network (GAT) and Graph Convolution Network (GCN) which have the "over-smoothing" issue, the proposed GTAN and GTCN models can go deep as demonstrated by comprehensive experiments and rigorous theoretical analysis.

Danqing Hu, Bing Liu, Xiang Li et al.

Prognosis prediction is crucial for determining optimal treatment plans for lung cancer patients. Traditionally, such predictions relied on models developed from retrospective patient data. Recently, large language models (LLMs) have gained attention for their ability to process and generate text based on extensive learned knowledge. In this study, we evaluate the potential of GPT-4o mini and GPT-3.5 in predicting the prognosis of lung cancer patients. We collected two prognosis datasets, i.e., survival and post-operative complication datasets, and designed multiple tasks to assess the models' performance comprehensively. Logistic regression models were also developed as baselines for comparison. The experimental results demonstrate that LLMs can achieve competitive, and in some tasks superior, performance in lung cancer prognosis prediction compared to data-driven logistic regression models despite not using additional patient data. These findings suggest that LLMs can be effective tools for prognosis prediction in lung cancer, particularly when patient data is limited or unavailable.

Kangning Liu, Yiqiu Shen, Nan Wu et al.

In the last few years, deep learning classifiers have shown promising results in image-based medical diagnosis. However, interpreting the outputs of these models remains a challenge. In cancer diagnosis, interpretability can be achieved by localizing the region of the input image responsible for the output, i.e. the location of a lesion. Alternatively, segmentation or detection models can be trained with pixel-wise annotations indicating the locations of malignant lesions. Unfortunately, acquiring such labels is labor-intensive and requires medical expertise. To overcome this difficulty, weakly-supervised localization can be utilized. These methods allow neural network classifiers to output saliency maps highlighting the regions of the input most relevant to the classification task (e.g. malignant lesions in mammograms) using only image-level labels (e.g. whether the patient has cancer or not) during training. When applied to high-resolution images, existing methods produce low-resolution saliency maps. This is problematic in applications in which suspicious lesions are small in relation to the image size. In this work, we introduce a novel neural network architecture to perform weakly-supervised segmentation of high-resolution images. The proposed model selects regions of interest via coarse-level localization, and then performs fine-grained segmentation of those regions. We apply this model to breast cancer diagnosis with screening mammography, and validate it on a large clinically-realistic dataset. Measured by Dice similarity score, our approach outperforms existing methods by a large margin in terms of localization performance of benign and malignant lesions, relatively improving the performance by 39.6% and 20.0%, respectively. Code and the weights of some of the models are available at https://github.com/nyukat/GLAM

Yiqiu Shen, Nan Wu, Jason Phang et al.

Medical images differ from natural images in significantly higher resolutions and smaller regions of interest. Because of these differences, neural network architectures that work well for natural images might not be applicable to medical image analysis. In this work, we extend the globally-aware multiple instance classifier, a framework we proposed to address these unique properties of medical images. This model first uses a low-capacity, yet memory-efficient, network on the whole image to identify the most informative regions. It then applies another higher-capacity network to collect details from chosen regions. Finally, it employs a fusion module that aggregates global and local information to make a final prediction. While existing methods often require lesion segmentation during training, our model is trained with only image-level labels and can generate pixel-level saliency maps indicating possible malignant findings. We apply the model to screening mammography interpretation: predicting the presence or absence of benign and malignant lesions. On the NYU Breast Cancer Screening Dataset, consisting of more than one million images, our model achieves an AUC of 0.93 in classifying breasts with malignant findings, outperforming ResNet-34 and Faster R-CNN. Compared to ResNet-34, our model is 4.1x faster for inference while using 78.4% less GPU memory. Furthermore, we demonstrate, in a reader study, that our model surpasses radiologist-level AUC by a margin of 0.11. The proposed model is available online: https://github.com/nyukat/GMIC.

Danqing Hu, Bing Liu, Xiaofeng Zhu et al.

Lymph node metastasis (LNM) is a crucial factor in determining the initial treatment for patients with lung cancer, yet accurate preoperative diagnosis of LNM remains challenging. Recently, large language models (LLMs) have garnered significant attention due to their remarkable text generation capabilities. Leveraging the extensive medical knowledge learned from vast corpora, LLMs can estimate probabilities for clinical problems, though their performance has historically been inferior to data-driven machine learning models. In this paper, we propose a novel ensemble method that combines the medical knowledge acquired by LLMs with the latent patterns identified by machine learning models to enhance LNM prediction performance. Initially, we developed machine learning models using patient data. We then designed a prompt template to integrate the patient data with the predicted probability from the machine learning model. Subsequently, we instructed GPT-4o, the most advanced LLM developed by OpenAI, to estimate the likelihood of LNM based on patient data and then adjust the estimate using the machine learning output. Finally, we collected three outputs from the GPT-4o using the same prompt and ensembled these results as the final prediction. Using the proposed method, our models achieved an AUC value of 0.778 and an AP value of 0.426 for LNM prediction, significantly improving predictive performance compared to baseline machine learning models. The experimental results indicate that GPT-4o can effectively leverage its medical knowledge and the probabilities predicted by machine learning models to achieve more accurate LNM predictions. These findings demonstrate that LLMs can perform well in clinical risk prediction tasks, offering a new paradigm for integrating medical knowledge and patient data in clinical predictions.

Nan Wu, Chaofan Wang

Heterogeneous graph neural networks (HGNNs) have attracted increasing research interest in recent three years. Most existing HGNNs fall into two classes. One class is meta-path-based HGNNs which either require domain knowledge to handcraft meta-paths or consume huge amount of time and memory to automatically construct meta-paths. The other class does not rely on meta-path construction. It takes homogeneous convolutional graph neural networks (Conv-GNNs) as backbones and extend them to heterogeneous graphs by introducing node-type- and edge-type-dependent parameters. Regardless of the meta-path dependency, most existing HGNNs employ shallow Conv-GNNs such as GCN and GAT to aggregate neighborhood information, and may have limited capability to capture information from high-order neighborhood. In this work, we propose two heterogeneous graph tree network models: Heterogeneous Graph Tree Convolutional Network (HetGTCN) and Heterogeneous Graph Tree Attention Network (HetGTAN), which do not rely on meta-paths to encode heterogeneity in both node features and graph structure. Extensive experiments on three real-world heterogeneous graph data demonstrate that the proposed HetGTCN and HetGTAN are efficient and consistently outperform all state-of-the-art HGNN baselines on semi-supervised node classification tasks, and can go deep without compromising performance.

StepFun, Bin Wang, Bojun Wang et al.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Yunxiang Zhang, Nan Wu, Connor Z. Lin et al.

Diffusion models offer unprecedented image generation power given just a text prompt. While emerging approaches for controlling diffusion models have enabled users to specify the desired spatial layouts of the generated content, they cannot predict or control where viewers will pay more attention due to the complexity of human vision. Recognizing the significance of attention-controllable image generation in practical applications, we present a saliency-guided framework to incorporate the data priors of human visual attention mechanisms into the generation process. Given a user-specified viewer attention distribution, our control module conditions a diffusion model to generate images that attract viewers' attention toward the desired regions. To assess the efficacy of our approach, we performed an eye-tracked user study and a large-scale model-based saliency analysis. The results evidence that both the cross-user eye gaze distributions and the saliency models' predictions align with the desired attention distributions. Lastly, we outline several applications, including interactive design of saliency guidance, attention suppression in unwanted regions, and adaptive generation for varied display/viewing conditions.

Zhenyu Liu, Yunzhen Liu, Zehao Fan et al.

Mixture-of-Experts (MoE) models scale capacity via sparse activation but stress memory and bandwidth. Offloading alleviates GPU memory by fetching experts on demand, yet token-level routing causes irregular transfers that make inference I/O-bound. Static uniform quantization reduces traffic but degrades accuracy under aggressive compression by ignoring expert heterogeneity. We present Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation, which performs router-guided precision restoration using precomputed low-rank compensators. At inference time, our method transfers compact low-rank factors with Top-n (n<k) experts per token and applies compensation to them, keeping others low-bit. Integrated with offloading on GPU and GPU-NDP systems, our method delivers a superior bandwidth-accuracy trade-off and improved throughput.

Joel Niklaus, Jakob Merane, Luka Nenadic et al.

In Switzerland legal translation is uniquely important due to the country's four official languages and requirements for multilingual legal documentation. However, this process traditionally relies on professionals who must be both legal experts and skilled translators -- creating bottlenecks and impacting effective access to justice. To address this challenge, we introduce SwiLTra-Bench, a comprehensive multilingual benchmark of over 180K aligned Swiss legal translation pairs comprising laws, headnotes, and press releases across all Swiss languages along with English, designed to evaluate LLM-based translation systems. Our systematic evaluation reveals that frontier models achieve superior translation performance across all document types, while specialized translation systems excel specifically in laws but under-perform in headnotes. Through rigorous testing and human expert validation, we demonstrate that while fine-tuning open SLMs significantly improves their translation quality, they still lag behind the best zero-shot prompted frontier models such as Claude-3.5-Sonnet. Additionally, we present SwiLTra-Judge, a specialized LLM evaluation system that aligns best with human expert assessments.

Yuwei Wan, Tong Xie, Nan Wu et al.

Exploring the predictive capabilities of language models in material science is an ongoing interest. This study investigates the application of language model embeddings to enhance material property prediction in materials science. By evaluating various contextual embedding methods and pre-trained models, including Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformers (GPT), we demonstrate that domain-specific models, particularly MatBERT significantly outperform general-purpose models in extracting implicit knowledge from compound names and material properties. Our findings reveal that information-dense embeddings from the third layer of MatBERT, combined with a context-averaging approach, offer the most effective method for capturing material-property relationships from the scientific literature. We also identify a crucial "tokenizer effect," highlighting the importance of specialized text processing techniques that preserve complete compound names while maintaining consistent token counts. These insights underscore the value of domain-specific training and tokenization in materials science applications and offer a promising pathway for accelerating the discovery and development of new materials through AI-driven approaches.

Haolong Yan, Jia Wang, Xin Huang et al.

Recent advances in multimodal large language models unlock unprecedented opportunities for GUI automation. However, a fundamental challenge remains: how to efficiently acquire high-quality training data while maintaining annotation reliability? We introduce a self-evolving training pipeline powered by the Calibrated Step Reward System, which converts model-generated trajectories into reliable training signals through trajectory-level calibration, achieving >90% annotation accuracy with 10-100x lower cost. Leveraging this pipeline, we introduce Step-GUI, a family of models (4B/8B) that achieves state-of-the-art GUI performance (8B: 80.2% AndroidWorld, 48.5% OSWorld, 62.6% ScreenShot-Pro) while maintaining robust general capabilities. As GUI agent capabilities improve, practical deployment demands standardized interfaces across heterogeneous devices while protecting user privacy. To this end, we propose GUI-MCP, the first Model Context Protocol for GUI automation with hierarchical architecture that combines low-level atomic operations and high-level task delegation to local specialist models, enabling high-privacy execution where sensitive data stays on-device. Finally, to assess whether agents can handle authentic everyday usage, we introduce AndroidDaily, a benchmark grounded in real-world mobile usage patterns with 3146 static actions and 235 end-to-end tasks across high-frequency daily scenarios (8B: static 89.91%, end-to-end 52.50%). Our work advances the development of practical GUI agents and demonstrates strong potential for real-world deployment in everyday digital interactions.

Nan Wang, Nan Wu, Xiangyu Hui et al.

As the demand for exercising the "right to be forgotten" grows, the need for verifiable machine unlearning has become increasingly evident to ensure both transparency and accountability. We present {\em zkUnlearner}, the first zero-knowledge framework for verifiable machine unlearning, specifically designed to support {\em multi-granularity} and {\em forgery-resistance}. First, we propose a general computational model that employs a {\em bit-masking} technique to enable the {\em selectivity} of existing zero-knowledge proofs of training for gradient descent algorithms. This innovation enables not only traditional {\em sample-level} unlearning but also more advanced {\em feature-level} and {\em class-level} unlearning. Our model can be translated to arithmetic circuits, ensuring compatibility with a broad range of zero-knowledge proof systems. Furthermore, our approach overcomes key limitations of existing methods in both efficiency and privacy. Second, forging attacks present a serious threat to the reliability of unlearning. Specifically, in Stochastic Gradient Descent optimization, gradients from unlearned data, or from minibatches containing it, can be forged using alternative data samples or minibatches that exclude it. We propose the first effective strategies to resist state-of-the-art forging attacks. Finally, we benchmark a zkSNARK-based instantiation of our framework and perform comprehensive performance evaluations to validate its practicality.

Wencheng Zou, Nan Wu

Graph neural networks (GNNs) have been widely applied in safety-critical applications, such as financial and medical networks, in which compromised predictions may cause catastrophic consequences. While existing research on GNN robustness has primarily focused on software-level threats, hardware-induced faults and errors remain largely underexplored. As hardware systems progress toward advanced technology nodes to meet high-performance and energy efficiency demands, they become increasingly susceptible to transient faults, which can cause bit flips and silent data corruption, a prominent issue observed by major technology companies (e.g., Meta and Google). In response, we first present a comprehensive analysis of GNN robustness against bit-flip errors, aiming to reveal system-level optimization opportunities for future reliable and efficient GNN systems. Second, we propose Ralts, a generalizable and lightweight solution to bolster GNN resilience to bit-flip errors. Specifically, Ralts exploits various graph similarity metrics to filter out outliers and recover compromised graph topology, and incorporates these protective techniques directly into aggregation functions to support any message-passing GNNs. Evaluation results demonstrate that Ralts effectively enhances GNN robustness across a range of GNN models, graph datasets, error patterns, and both dense and sparse architectures. On average, under a BER of $3\times10^{-5}$, these robust aggregation functions improve prediction accuracy by at least 20\% when errors present in model weights or node embeddings, and by at least 10\% when errors occur in adjacency matrices. Ralts is also optimized to deliver execution efficiency comparable to built-in aggregation functions in PyTorch Geometric.

Jinhang Jiang, Nan Wu, Ben Liu et al.

Time series forecasting underpins vital decision-making across various sectors, yet raw predictions from sophisticated models often harbor systematic errors and biases. We examine the Forecast-Then-Optimize (FTO) framework, pioneering its systematic synopsis. Unlike conventional Predict-Then-Optimize (PTO) methods, FTO explicitly refines forecasts through optimization techniques such as ensemble methods, meta-learners, and uncertainty adjustments. Furthermore, deep learning and large language models have established superiority over traditional parametric forecasting models for most enterprise applications. This paper surveys significant advancements from 2016 to 2025, analyzing mainstream deep learning FTO architectures. Focusing on real-world applications in operations management, we demonstrate FTO's crucial role in enhancing predictive accuracy, robustness, and decision efficacy. Our study establishes foundational guidelines for future forecasting methodologies, bridging theory and operational practicality.

Danqing Hu, Bing Liu, Xiaofeng Zhu et al.

Electronic health records contain an enormous amount of valuable information, but many are recorded in free text. Information extraction is the strategy to transform the sequence of characters into structured data, which can be employed for secondary analysis. However, the traditional information extraction components, such as named entity recognition and relation extraction, require annotated data to optimize the model parameters, which has become one of the major bottlenecks in building information extraction systems. With the large language models achieving good performances on various downstream NLP tasks without parameter tuning, it becomes possible to use large language models for zero-shot information extraction. In this study, we aim to explore whether the most popular large language model, ChatGPT, can extract useful information from the radiological reports. We first design the prompt template for the interested information in the CT reports. Then, we generate the prompts by combining the prompt template with the CT reports as the inputs of ChatGPT to obtain the responses. A post-processing module is developed to transform the responses into structured extraction results. We conducted the experiments with 847 CT reports collected from Peking University Cancer Hospital. The experimental results indicate that ChatGPT can achieve competitive performances for some extraction tasks compared with the baseline information extraction system, but some limitations need to be further improved.

Nan Wu, Stanisław Jastrzębski, Kyunghyun Cho et al.

We hypothesize that due to the greedy nature of learning in multi-modal deep neural networks, these models tend to rely on just one modality while under-fitting the other modalities. Such behavior is counter-intuitive and hurts the models' generalization, as we observe empirically. To estimate the model's dependence on each modality, we compute the gain on the accuracy when the model has access to it in addition to another modality. We refer to this gain as the conditional utilization rate. In the experiments, we consistently observe an imbalance in conditional utilization rates between modalities, across multiple tasks and architectures. Since conditional utilization rate cannot be computed efficiently during training, we introduce a proxy for it based on the pace at which the model learns from each modality, which we refer to as the conditional learning speed. We propose an algorithm to balance the conditional learning speeds between modalities during training and demonstrate that it indeed addresses the issue of greedy learning. The proposed algorithm improves the model's generalization on three datasets: Colored MNIST, ModelNet40, and NVIDIA Dynamic Hand Gesture.

Ruizhi Cheng, Nan Wu, Songqing Chen et al.

Metaverse, with the combination of the prefix "meta" (meaning transcending) and the word "universe", has been deemed as the next-generation (NextG) Internet. It aims to create a shared virtual space that connects all virtual worlds via the Internet, where users, represented as digital avatars, can communicate and collaborate as if they are in the physical world. Nevertheless, there is still no unified definition of the Metaverse. This article first presents our vision of what the key requirements of Metaverse should be and reviews what has been heavily advocated by the industry and the positions of various high-tech companies. It then briefly introduces existing social virtual reality (VR) platforms that can be viewed as early prototypes of Metaverse and conducts a reality check by diving into the network operation and performance of two representative platforms, Workrooms from Meta and AltspaceVR from Microsoft. Finally, it concludes by discussing several opportunities and future directions for further innovation.

Nan Wu, Jiwon Lee, Yuan Xie et al.

Despite the stride made by machine learning (ML) based performance modeling, two major concerns that may impede production-ready ML applications in EDA are stringent accuracy requirements and generalization capability. To this end, we propose hybrid graph neural network (GNN) based approaches towards highly accurate quality-of-result (QoR) estimations with great generalization capability, specifically targeting logic synthesis optimization. The key idea is to simultaneously leverage spatio-temporal information from hardware designs and logic synthesis flows to forecast performance (i.e., delay/area) of various synthesis flows on different designs. The structural characteristics inside hardware designs are distilled and represented by GNNs; the temporal knowledge (i.e., relative ordering of logic transformations) in synthesis flows can be imposed on hardware designs by combining a virtually added supernode or a sequence processing model with conventional GNN models. Evaluation on 3.3 million data points shows that the testing mean absolute percentage error (MAPE) on designs seen and unseen during training are no more than 1.2% and 3.1%, respectively, which are 7-15X lower than existing studies.

Nan Wu, Hang Yang, Yuan Xie et al.

Agile hardware development requires fast and accurate circuit quality evaluation from early design stages. Existing work of high-level synthesis (HLS) performance prediction usually needs extensive feature engineering after the synthesis process. To expedite circuit evaluation from as earlier design stage as possible, we propose a rapid and accurate performance modeling, exploiting the representation power of graph neural networks (GNNs) by representing C/C++ programs as graphs. The contribution of this work is three-fold. First, we build a standard benchmark containing 40k C synthesizable programs, which includes both synthetic programs and three sets of real-world HLS benchmarks. Each program is implemented on FPGA to generate ground-truth performance metrics. Second, we formally formulate the HLS performance prediction problem on graphs, and propose multiple modeling strategies with GNNs that leverage different trade-offs between prediction timeliness (early/late prediction) and accuracy. Third, we further propose a novel hierarchical GNN that does not sacrifice timeliness but largely improves prediction accuracy, significantly outperforming HLS tools. We apply extensive evaluations for both synthetic and unseen real-case programs; our proposed predictor largely outperforms HLS by up to 40X and excels existing predictors by 2X to 5X in terms of resource usage and timing prediction.

Yongjie Guan, Xueyu Hou, Nan Wu et al.

Mobile headsets should be capable of understanding 3D physical environments to offer a truly immersive experience for augmented/mixed reality (AR/MR). However, their small form-factor and limited computation resources make it extremely challenging to execute in real-time 3D vision algorithms, which are known to be more compute-intensive than their 2D counterparts. In this paper, we propose DeepMix, a mobility-aware, lightweight, and hybrid 3D object detection framework for improving the user experience of AR/MR on mobile headsets. Motivated by our analysis and evaluation of state-of-the-art 3D object detection models, DeepMix intelligently combines edge-assisted 2D object detection and novel, on-device 3D bounding box estimations that leverage depth data captured by headsets. This leads to low end-to-end latency and significantly boosts detection accuracy in mobile scenarios. A unique feature of DeepMix is that it fully exploits the mobility of headsets to fine-tune detection results and boost detection accuracy. To the best of our knowledge, DeepMix is the first 3D object detection that achieves 30 FPS (an end-to-end latency much lower than the 100 ms stringent requirement of interactive AR/MR). We implement a prototype of DeepMix on Microsoft HoloLens and evaluate its performance via both extensive controlled experiments and a user study with 30+ participants. DeepMix not only improves detection accuracy by 9.1--37.3% but also reduces end-to-end latency by 2.68--9.15x, compared to the baseline that uses existing 3D object detection models.

Ankur Bapna, Yu-an Chung, Nan Wu et al.

Unsupervised pre-training is now the predominant approach for both text and speech understanding. Self-attention models pre-trained on large amounts of unannotated data have been hugely successful when fine-tuned on downstream tasks from a variety of domains and languages. This paper takes the universality of unsupervised language pre-training one step further, by unifying speech and text pre-training within a single model. We build a single encoder with the BERT objective on unlabeled text together with the w2v-BERT objective on unlabeled speech. To further align our model representations across modalities, we leverage alignment losses, specifically Translation Language Modeling (TLM) and Speech Text Matching (STM) that make use of supervised speech-text recognition data. We demonstrate that incorporating both speech and text data during pre-training can significantly improve downstream quality on CoVoST~2 speech translation, by around 1 BLEU compared to single-modality pre-trained models, while retaining close to SotA performance on LibriSpeech and SpeechStew ASR tasks. On four GLUE tasks and text-normalization, we observe evidence of capacity limitations and interference between the two modalities, leading to degraded performance compared to an equivalent text-only model, while still being competitive with BERT. Through extensive empirical analysis we also demonstrate the importance of the choice of objective function for speech pre-training, and the beneficial effect of adding additional supervised signals on the quality of the learned representations.

David B Dunson, Nan Wu

It is often of interest to infer lower-dimensional structure underlying complex data. As a flexible class of non-linear structures, it is common to focus on Riemannian manifolds. Most existing manifold learning algorithms replace the original data with lower-dimensional coordinates without providing an estimate of the manifold or using the manifold to denoise the original data. This article proposes a new methodology to address these problems, allowing interpolation of the estimated manifold between the fitted data points. The proposed approach is motivated by the novel theoretical properties of local covariance matrices constructed from samples near a manifold. Our results enable us to turn a global manifold reconstruction problem into a local regression problem, allowing for the application of Gaussian processes for probabilistic manifold reconstruction. In addition to the theory justifying our methodology, we provide simulated and real data examples to illustrate the performance.

Nan Wu, Huake He, Yuan Xie et al.

Circuit design is complicated and requires extensive domain-specific expertise. One major obstacle stuck on the way to hardware agile development is the considerably time-consuming process of accurate circuit quality evaluation. To significantly expedite the circuit evaluation during the translation from behavioral languages to circuit designs, we formulate it as a Program-to-Circuit problem, aiming to exploit the representation power of graph neural networks (GNNs) by representing C/C++ programs as graphs. The goal of this work is four-fold. First, we build a standard benchmark containing 40k C/C++ programs, each of which is translated to a circuit design with actual hardware quality metrics, aiming to facilitate the development of effective GNNs targeting this high-demand circuit design area. Second, 14 state-of-the-art GNN models are analyzed on the Program-to-Circuit problem. We identify key design challenges of this problem, which should be carefully handled but not yet solved by existing GNNs. The goal is to provide domain-specific knowledge for designing GNNs with suitable inductive biases. Third, we discuss three sets of real-world benchmarks for GNN generalization evaluation, and analyze the performance gap between standard programs and the real-case ones. The goal is to enable transfer learning from limited training data to real-world large-scale circuit design problems. Fourth, the Program-to-Circuit problem is a representative within the Program-to-X framework, a set of program-based analysis problems with various downstream tasks. The in-depth understanding of strength and weaknesses in applying GNNs on Program-to-Circuit could largely benefit the entire family of Program-to-X. Pioneering in this direction, we expect more GNN endeavors to revolutionize this high-demand Program-to-Circuit problem and to enrich the expressiveness of GNNs on programs.

Nan Wu, Yuan Xie, Cong Hao

Despite the great success of High-Level Synthesis (HLS) tools, we observe several unresolved challenges: 1) the high-level abstraction of programming styles in HLS sometimes conceals optimization opportunities; 2) existing HLS tools do not provide flexible trade-off (Pareto) solutions among different objectives and constraints; 3) the actual quality of the resulting RTL designs is hard to predict. To address these challenges, we propose an end-to-end framework, namelyIronMan. The primary goal is to enable a flexible and automated design space exploration (DSE), to provide either optimal solutions under user-specified constraints, or various trade-offs among different objectives (such as different types of resources, area, and latency). Such DSE either requires tedious manual efforts or is not achievable to attain these goals through existing HLS tools. There are three components in IronMan: 1) GPP, a highly accurate graph-neural-network-based performance and resource predictor; 2) RLMD, a reinforcement-learning-based multi-objective DSE engine that explores the optimal resource allocation strategy, to provide Pareto solutions between different objectives; 3) CT, a code transformer to assist RLMD and GPP, which extracts the data flow graph from original HLS C/C++ and automatically generates synthesizable code with HLS directives. The experimental results show that: 1) GPP achieves high prediction accuracy, reducing prediction errors of HLS tools by 10.9x in resource utilization and 5.7x in timing; 2) RLMD obtains optimal or Pareto solutions that outperform the genetic algorithm and simulated annealing by 12.7% and 12.9%, respectively; 3) IronMan is able to find optimized solutions perfectly matching various DSP constraints, with 2.54x fewer DSPs and up to 6x shorter latency than those of HLS tools while being up to 400x faster than the heuristic algorithms and HLS tools.

Nan Wu, Yuan Xie

It has been a long time that computer architecture and systems are optimized for efficient execution of machine learning (ML) models. Now, it is time to reconsider the relationship between ML and systems, and let ML transform the way that computer architecture and systems are designed. This embraces a twofold meaning: improvement of designers' productivity, and completion of the virtuous cycle. In this paper, we present a comprehensive review of the work that applies ML for computer architecture and system design. First, we perform a high-level taxonomy by considering the typical role that ML techniques take in architecture/system design, i.e., either for fast predictive modeling or as the design methodology. Then, we summarize the common problems in computer architecture/system design that can be solved by ML techniques, and the typical ML techniques employed to resolve each of them. In addition to emphasis on computer architecture in a narrow sense, we adopt the concept that data centers can be recognized as warehouse-scale computers; sketchy discussions are provided in adjacent computer systems, such as code generation and compiler; we also give attention to how ML techniques can aid and transform design automation. We further provide a future vision of opportunities and potential directions, and envision that applying ML for computer architecture and systems would thrive in the community.

Xiuyuan Cheng, Nan Wu

This work studies the spectral convergence of graph Laplacian to the Laplace-Beltrami operator when the graph affinity matrix is constructed from $N$ random samples on a $d$-dimensional manifold embedded in a possibly high dimensional space. By analyzing Dirichlet form convergence and constructing candidate approximate eigenfunctions via convolution with manifold heat kernel, we prove that, with Gaussian kernel, one can set the kernel bandwidth parameter $ε\sim (\log N/ N)^{1/(d/2+2)}$ such that the eigenvalue convergence rate is $N^{-1/(d/2+2)}$ and the eigenvector convergence in 2-norm has rate $N^{-1/(d+4)}$; When $ε\sim (\log N/N)^{1/(d/2+3)}$, both eigenvalue and eigenvector rates are $N^{-1/(d/2+3)}$. These rates are up to a $\log N$ factor and proved for finitely many low-lying eigenvalues. The result holds for un-normalized and random-walk graph Laplacians when data are uniformly sampled on the manifold, as well as the density-corrected graph Laplacian (where the affinity matrix is normalized by the degree matrix from both sides) with non-uniformly sampled data. As an intermediate result, we prove new point-wise and Dirichlet form convergence rates for the density-corrected graph Laplacian. Numerical results are provided to verify the theory.

Nan Wu, Pengcheng Li

With data durability, high access speed, low power efficiency and byte addressability, NVMe and SSD, which are acknowledged representatives of emerging storage technologies, have been applied broadly in many areas. However, one key issue with high-performance adoption of these technologies is how to properly define intelligent cache layers such that the performance gap between emerging technologies and main memory can be well bridged. To this end, we propose Phoebe, a reuse-aware reinforcement learning framework for the optimal online caching that is applicable for a wide range of emerging storage models. By continuous interacting with the cache environment and the data stream, Phoebe is capable to extract critical temporal data dependency and relative positional information from a single trace, becoming ever smarter over time. To reduce training overhead during online learning, we utilize periodical training to amortize costs. Phoebe is evaluated on a set of Microsoft cloud storage workloads. Experiment results show that Phoebe is able to close the gap of cache miss rate from LRU and a state-of-the-art online learning based cache policy to the Belady's optimal policy by 70.3% and 52.6%, respectively.

David B Dunson, Hau-Tieng Wu, Nan Wu

In nonparametric regression, it is common for the inputs to fall in a restricted subset of Euclidean space. Typical kernel-based methods that do not take into account the intrinsic geometry of the domain across which observations are collected may produce sub-optimal results. In this article, we focus on solving this problem in the context of Gaussian process (GP) models, proposing a new class of Graph Laplacian based GPs (GL-GPs), which learn a covariance that respects the geometry of the input domain. As the heat kernel is intractable computationally, we approximate the covariance using finitely-many eigenpairs of the Graph Laplacian (GL). The GL is constructed from a kernel which depends only on the Euclidean coordinates of the inputs. Hence, we can benefit from the full knowledge about the kernel to extend the covariance structure to newly arriving samples by a Nyström type extension. We provide substantial theoretical support for the GL-GP methodology, and illustrate performance gains in various applications.

Nan Wu, Zhe Huang, Yiqiu Shen et al.

Breast cancer is the most common cancer in women, and hundreds of thousands of unnecessary biopsies are done around the world at a tremendous cost. It is crucial to reduce the rate of biopsies that turn out to be benign tissue. In this study, we build deep neural networks (DNNs) to classify biopsied lesions as being either malignant or benign, with the goal of using these networks as second readers serving radiologists to further reduce the number of false positive findings. We enhance the performance of DNNs that are trained to learn from small image patches by integrating global context provided in the form of saliency maps learned from the entire image into their reasoning, similar to how radiologists consider global context when evaluating areas of interest. Our experiments are conducted on a dataset of 229,426 screening mammography exams from 141,473 patients. We achieve an AUC of 0.8 on a test set consisting of 464 benign and 136 malignant lesions.

Farah E. Shamout, Yiqiu Shen, Nan Wu et al.

During the coronavirus disease 2019 (COVID-19) pandemic, rapid and accurate triage of patients at the emergency department is critical to inform decision-making. We propose a data-driven approach for automatic prediction of deterioration risk using a deep neural network that learns from chest X-ray images and a gradient boosting model that learns from routine clinical variables. Our AI prognosis system, trained using data from 3,661 patients, achieves an area under the receiver operating characteristic curve (AUC) of 0.786 (95% CI: 0.745-0.830) when predicting deterioration within 96 hours. The deep neural network extracts informative areas of chest X-ray images to assist clinicians in interpreting the predictions and performs comparably to two radiologists in a reader study. In order to verify performance in a real clinical setting, we silently deployed a preliminary version of the deep neural network at New York University Langone Health during the first wave of the pandemic, which produced accurate predictions in real-time. In summary, our findings demonstrate the potential of the proposed system for assisting front-line physicians in the triage of COVID-19 patients.

Hau-Tieng Wu, Nan Wu

When analyzing modern machine learning algorithms, we may need to handle kernel density estimation (KDE) with intricate kernels that are not designed by the user and might even be irregular and asymmetric. To handle this emerging challenge, we provide a strong uniform consistency result with the $L^\infty$ convergence rate for KDE on Riemannian manifolds with Riemann integrable kernels (in the ambient Euclidean space). We also provide an $L^1$ consistency result for kernel density estimation on Riemannian manifolds with Lebesgue integrable kernels. The isotropic kernels considered in this paper are different from the kernels in the Vapnik-Chervonenkis class that are frequently considered in statistics society. We illustrate the difference when we apply them to estimate the probability density function. Moreover, we elaborate the delicate difference when the kernel is designed on the intrinsic manifold and on the ambient Euclidian space, both might be encountered in practice. At last, we prove the necessary and sufficient condition for an isotropic kernel to be Riemann integrable on a submanifold in the Euclidean space.

Yuan Gao, Jian-Guo Liu, Nan Wu

We study the Langevin dynamics of a physical system with manifold structure $\mathcal{M}\subset\mathbb{R}^p$ based on collected sample points $\{\mathsf{x}_i\}_{i=1}^n \subset \mathcal{M}$ that probe the unknown manifold $\mathcal{M}$. Through the diffusion map, we first learn the reaction coordinates $\{\mathsf{y}_i\}_{i=1}^n\subset \mathcal{N}$ corresponding to $\{\mathsf{x}_i\}_{i=1}^n$, where $\mathcal{N}$ is a manifold diffeomorphic to $\mathcal{M}$ and isometrically embedded in $\mathbb{R}^\ell$ with $\ell \ll p$. The induced Langevin dynamics on $\mathcal{N}$ in terms of the reaction coordinates captures the slow time scale dynamics such as conformational changes in biochemical reactions. To construct an efficient and stable approximation for the Langevin dynamics on $\mathcal{N}$, we leverage the corresponding Fokker-Planck equation on the manifold $\mathcal{N}$ in terms of the reaction coordinates $\mathsf{y}$. We propose an implementable, unconditionally stable, data-driven finite volume scheme for this Fokker-Planck equation, which automatically incorporates the manifold structure of $\mathcal{N}$. Furthermore, we provide a weighted $L^2$ convergence analysis of the finite volume scheme to the Fokker-Planck equation on $\mathcal{N}$. The proposed finite volume scheme leads to a Markov chain on $\{\mathsf{y}_i\}_{i=1}^n$ with an approximated transition probability and jump rate between the nearest neighbor points. After an unconditionally stable explicit time discretization, the data-driven finite volume scheme gives an approximated Markov process for the Langevin dynamics on $\mathcal{N}$ and the approximated Markov process enjoys detailed balance, ergodicity, and other good properties.

Farhad Farokhi, Nan Wu, David Smith et al.

We consider training machine learning models using Training data located on multiple private and geographically-scattered servers with different privacy settings. Due to the distributed nature of the data, communicating with all collaborating private data owners simultaneously may prove challenging or altogether impossible. In this paper, we develop differentially-private asynchronous algorithms for collaboratively training machine-learning models on multiple private datasets. The asynchronous nature of the algorithms implies that a central learner interacts with the private data owners one-on-one whenever they are available for communication without needing to aggregate query responses to construct gradients of the entire fitness function. Therefore, the algorithm efficiently scales to many data owners. We define the cost of privacy as the difference between the fitness of a privacy-preserving machine-learning model and the fitness of trained machine-learning model in the absence of privacy concerns. We prove that we can forecast the performance of the proposed privacy-preserving asynchronous algorithms. We demonstrate that the cost of privacy has an upper bound that is inversely proportional to the combined size of the training datasets squared and the sum of the privacy budgets squared. We validate the theoretical results with experiments on financial and medical datasets. The experiments illustrate that collaboration among more than 10 data owners with at least 10,000 records with privacy budgets greater than or equal to 1 results in a superior machine-learning model in comparison to a model trained in isolation on only one of the datasets, illustrating the value of collaboration and the cost of the privacy. The number of the collaborating datasets can be lowered if the privacy budget is higher.

Nan Wu, Adrien Vincent, Dmitri Strukov et al.

Recently, significant progress has been made in solving sophisticated problems among various domains by using reinforcement learning (RL), which allows machines or agents to learn from interactions with environments rather than explicit supervision. As the end of Moore's law seems to be imminent, emerging technologies that enable high performance neuromorphic hardware systems are attracting increasing attention. Namely, neuromorphic architectures that leverage memristors, the programmable and nonvolatile two-terminal devices, as synaptic weights in hardware neural networks, are candidates of choice to realize such highly energy-efficient and complex nervous systems. However, one of the challenges for memristive hardware with integrated learning capabilities is prohibitively large number of write cycles that might be required during learning process, and this situation is even exacerbated under RL situations. In this work we propose a memristive neuromorphic hardware implementation for the actor-critic algorithm in RL. By introducing a two-fold training procedure (i.e., ex-situ pre-training and in-situ re-training) and several training techniques, the number of weight updates can be significantly reduced and thus it will be suitable for efficient in-situ learning implementations. As a case study, we consider the task of balancing an inverted pendulum, a classical problem in both RL and control theory. We believe that this study shows the promise of using memristor-based hardware neural networks for handling complex tasks through in-situ reinforcement learning.

Thibault Févry, Jason Phang, Nan Wu et al.

We trained and evaluated a localization-based deep CNN for breast cancer screening exam classification on over 200,000 exams (over 1,000,000 images). Our model achieves an AUC of 0.919 in predicting malignancy in patients undergoing breast cancer screening, reducing the error rate of the baseline (Wu et al., 2019a) by 23%. In addition, the models generates bounding boxes for benign and malignant findings, providing interpretable predictions.

Jungkyu Park, Jason Phang, Yiqiu Shen et al.

Radiologists typically compare a patient's most recent breast cancer screening exam to their previous ones in making informed diagnoses. To reflect this practice, we propose new neural network models that compare pairs of screening mammograms from the same patient. We train and evaluate our proposed models on over 665,000 pairs of images (over 166,000 pairs of exams). Our best model achieves an AUC of 0.866 in predicting malignancy in patients who underwent breast cancer screening, reducing the error rate of the corresponding baseline.

Nan Wu, Farhad Farokhi, David Smith et al.

In this paper, we apply machine learning to distributed private data owned by multiple data owners, entities with access to non-overlapping training datasets. We use noisy, differentially-private gradients to minimize the fitness cost of the machine learning model using stochastic gradient descent. We quantify the quality of the trained model, using the fitness cost, as a function of privacy budget and size of the distributed datasets to capture the trade-off between privacy and utility in machine learning. This way, we can predict the outcome of collaboration among privacy-aware data owners prior to executing potentially computationally-expensive machine learning algorithms. Particularly, we show that the difference between the fitness of the trained machine learning model using differentially-private gradient queries and the fitness of the trained machine model in the absence of any privacy concerns is inversely proportional to the size of the training datasets squared and the privacy budget squared. We successfully validate the performance prediction with the actual performance of the proposed privacy-aware learning algorithms, applied to: financial datasets for determining interest rates of loans using regression; and detecting credit card frauds using support vector machines.

Yiqiu Shen, Nan Wu, Jason Phang et al.

Deep learning models designed for visual classification tasks on natural images have become prevalent in medical image analysis. However, medical images differ from typical natural images in many ways, such as significantly higher resolutions and smaller regions of interest. Moreover, both the global structure and local details play important roles in medical image analysis tasks. To address these unique properties of medical images, we propose a neural network that is able to classify breast cancer lesions utilizing information from both a global saliency map and multiple local patches. The proposed model outperforms the ResNet-based baseline and achieves radiologist-level performance in the interpretation of screening mammography. Although our model is trained only with image-level labels, it is able to generate pixel-level saliency maps that provide localization of possible malignant findings.

Nan Wu, Jason Phang, Jungkyu Park et al.

We present a deep convolutional neural network for breast cancer screening exam classification, trained and evaluated on over 200,000 exams (over 1,000,000 images). Our network achieves an AUC of 0.895 in predicting whether there is a cancer in the breast, when tested on the screening population. We attribute the high accuracy of our model to a two-stage training procedure, which allows us to use a very high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and find our model to be as accurate as experienced radiologists when presented with the same data. Finally, we show that a hybrid model, averaging probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To better understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, model design, training procedure, errors, and properties of its internal representations.