Search

Soft prompt tuning is a widely studied parameter-efficient fine-tuning method. However, it has a clear drawback: many soft tokens must be inserted into the input sequences to guarantee downstream performance. As a result, soft prompt tuning is less considered than Low-rank adaptation (LoRA) in the large language modeling (LLM) era. In this work, we propose a novel prompt tuning method, Instruction-Aware Prompt Tuning (IAPT), that requires only four soft tokens. First, we install a parameter-efficient soft prompt generator at each Transformer layer to generate idiosyncratic soft prompts for each input instruction. The generated soft prompts can be seen as a semantic summary of the input instructions and can effectively guide the output generation. Second, the soft prompt generators are modules with a bottleneck architecture consisting of a self-attention pooling operation, two linear projections, and an activation function. Pilot experiments show that prompt generators at different Transformer layers require different activation functions. Thus, we propose to learn the idiosyncratic activation functions for prompt generators automatically with the help of rational functions. We have conducted experiments on various tasks, and the experimental results demonstrate that (a) our IAPT method can outperform the recent baselines with comparable tunable parameters. (b) Our IAPT method is more efficient than LoRA under the single-backbone multi-tenant setting., Comment: Accepted by ACL-2024

Knowledge of the medical decision process, which can be modeled as medical decision trees (MDTs), is critical to build clinical decision support systems. However, the current MDT construction methods rely heavily on time-consuming and laborious manual annotation. In this work, we propose a novel task, Text2MDT, to explore the automatic extraction of MDTs from medical texts such as medical guidelines and textbooks. We normalize the form of the MDT and create an annotated Text-to-MDT dataset in Chinese with the participation of medical experts. We investigate two different methods for the Text2MDT tasks: (a) an end-to-end framework which only relies on a GPT style large language models (LLM) instruction tuning to generate all the node information and tree structures. (b) The pipeline framework which decomposes the Text2MDT task to three subtasks. Experiments on our Text2MDT dataset demonstrate that: (a) the end-to-end method basd on LLMs (7B parameters or larger) show promising results, and successfully outperform the pipeline methods. (b) The chain-of-thought (COT) prompting method \cite{Wei2022ChainOT} can improve the performance of the fine-tuned LLMs on the Text2MDT test set. (c) the lightweight pipelined method based on encoder-based pretrained models can perform comparably with LLMs with model complexity two magnititudes smaller. Our Text2MDT dataset is open-sourced at \url{https://tianchi.aliyun.com/dataset/95414}, and the source codes are open-sourced at \url{https://github.com/michael-wzhu/text2dt}.

Learning from human feedback has become a pivot technique in aligning large language models (LLMs) with human preferences. However, acquiring vast and premium human feedback is bottlenecked by time, labor, and human capability, resulting in small sizes or limited topics of current datasets. This further hinders feedback learning as well as alignment research within the open-source community. To address this issue, we explore how to go beyond human feedback and collect high-quality \textit{AI feedback} automatically for a scalable alternative. Specifically, we identify \textbf{scale and diversity} as the key factors for feedback data to take effect. Accordingly, we first broaden instructions and responses in both amount and breadth to encompass a wider range of user-assistant interactions. Then, we meticulously apply a series of techniques to mitigate annotation biases for more reliable AI feedback. We finally present \textsc{UltraFeedback}, a large-scale, high-quality, and diversified AI feedback dataset, which contains over 1 million GPT-4 feedback for 250k user-assistant conversations from various aspects. Built upon \textsc{UltraFeedback}, we align a LLaMA-based model by best-of-$n$ sampling and reinforcement learning, demonstrating its exceptional performance on chat benchmarks. Our work validates the effectiveness of scaled AI feedback data in constructing strong open-source chat language models, serving as a solid foundation for future feedback learning research. Our data and models are available at https://github.com/thunlp/UltraFeedback., Comment: ICML 2024 camera ready

Natural language is expected to be a key medium for various human-machine interactions in the era of large language models. When it comes to the biochemistry field, a series of tasks around molecules (e.g., property prediction, molecule mining, etc.) are of great significance while having a high technical threshold. Bridging the molecule expressions in natural language and chemical language can not only hugely improve the interpretability and reduce the operation difficulty of these tasks, but also fuse the chemical knowledge scattered in complementary materials for a deeper comprehension of molecules. Based on these benefits, we propose the conversational molecular design, a novel task adopting natural language for describing and editing target molecules. To better accomplish this task, we design ChatMol, a knowledgeable and versatile generative pre-trained model, enhanced by injecting experimental property information, molecular spatial knowledge, and the associations between natural and chemical languages into it. Several typical solutions including large language models (e.g., ChatGPT) are evaluated, proving the challenge of conversational molecular design and the effectiveness of our knowledge enhancement method. Case observations and analysis are conducted to provide directions for further exploration of natural-language interaction in molecular discovery.

Parameter-efficient tuning (PET) methods can effectively drive extremely large pre-trained language models (PLMs) by training only minimal parameters. Different PET methods utilize different manually designed tunable modules. In small PLMs, there are usually noticeable performance differences among PET methods. Nevertheless, as the model scale increases, the performance differences become marginal. Hence, we hypothesize that model scaling mitigates the impact of design differences on PET methods. To investigate this hypothesis, we introduce a more flexible PET method called Arbitrary PET (APET) method. The APET method is compatible with a tunable module, which consists of any number of parameters distributed in arbitrary positions. Then, we utilize it and conduct experiments on 11 NLP tasks across 3 representative PLMs. Our investigations reveal that model scaling (1) mitigates the effects of the positions of tunable parameters on performance, and (2) enables tuning methods to achieve performance comparable to full-parameter fine-tuning by optimizing fewer tunable parameters. Intriguingly, we also observe that tuning methods optimize the similar number of tunable parameters to exceed random guess performance on different tasks. We collectively discuss this phenomenon and the two aforementioned findings from an optimization perspective to understand the underlying mechanisms. These conclusions enhance our understanding of the impact of model scaling on PET and assist in designing more effective and efficient PET methods for PLMs of different scales. The source code can be obtained from this GitHub repository: \url{https://github.com/yushengsu-thu/PET_Scaling}.

In-context learning is a new learning paradigm where a language model conditions on a few input-output pairs (demonstrations) and a test input, and directly outputs the prediction. It has been shown highly dependent on the provided demonstrations and thus promotes the research of demonstration retrieval: given a test input, relevant examples are retrieved from the training set to serve as informative demonstrations for in-context learning. While previous works focus on training task-specific retrievers for several tasks separately, these methods are often hard to transfer and scale on various tasks, and separately trained retrievers incur a lot of parameter storage and deployment cost. In this paper, we propose Unified Demonstration Retriever (\textbf{UDR}), a single model to retrieve demonstrations for a wide range of tasks. To train UDR, we cast various tasks' training signals into a unified list-wise ranking formulation by language model's feedback. Then we propose a multi-task list-wise ranking training framework, with an iterative mining strategy to find high-quality candidates, which can help UDR fully incorporate various tasks' signals. Experiments on 30+ tasks across 13 task families and multiple data domains show that UDR significantly outperforms baselines. Further analyses show the effectiveness of each proposed component and UDR's strong ability in various scenarios including different LMs (1.3B - 175B), unseen datasets, varying demonstration quantities, etc., Comment: ACL 2023 camera ready version

Filter pruning is widely adopted to compress and accelerate the Convolutional Neural Networks (CNNs), but most previous works ignore the relationship between filters and channels in different layers. Processing each layer independently fails to utilize the collaborative relationship across layers. In this paper, we intuitively propose a novel pruning method by explicitly leveraging the Filters Similarity in Consecutive Layers (FSCL). FSCL compresses models by pruning filters whose corresponding features are more worthless in the model. The extensive experiments demonstrate the effectiveness of FSCL, and it yields remarkable improvement over state-of-the-art on accuracy, FLOPs and parameter reduction on several benchmark models and datasets., Comment: Accepted by ICASSP 2023 (oral)

Contrastive learning has emerged as a powerful tool for graph representation learning. However, most contrastive learning methods learn features of graphs with fixed coarse-grained scale, which might underestimate either local or global information. To capture more hierarchical and richer representation, we propose a novel Hierarchical Contrastive Learning (HCL) framework that explicitly learns graph representation in a hierarchical manner. Specifically, HCL includes two key components: a novel adaptive Learning to Pool (L2Pool) method to construct more reasonable multi-scale graph topology for more comprehensive contrastive objective, a novel multi-channel pseudo-siamese network to further enable more expressive learning of mutual information within each scale. Comprehensive experimental results show HCL achieves competitive performance on 12 datasets involving node classification, node clustering and graph classification. In addition, the visualization of learned representation reveals that HCL successfully captures meaningful characteristics of graphs., Comment: published at The 21st International Semantic Web Conference ( ISWC 2022 )

Large scale pre-training models have been widely used in named entity recognition (NER) tasks. However, model ensemble through parameter averaging or voting can not give full play to the differentiation advantages of different models, especially in the open domain. This paper describes our NER system in the SemEval 2022 task11: MultiCoNER. We proposed an effective system to adaptively ensemble pre-trained language models by a Transformer layer. By assigning different weights to each model for different inputs, we adopted the Transformer layer to integrate the advantages of diverse models effectively. Experimental results show that our method achieves superior performances in Farsi and Dutch.

This paper describes the PASH participation in TREC 2021 Deep Learning Track. In the recall stage, we adopt a scheme combining sparse and dense retrieval method. In the multi-stage ranking phase, point-wise and pair-wise ranking strategies are used one after another based on model continual pre-trained on general knowledge and document-level data. Compared to TREC 2020 Deep Learning Track, we have additionally introduced the generative model T5 to further enhance the performance., Comment: TREC 2021

Recognition of glomeruli lesions is the key for diagnosis and treatment planning in kidney pathology; however, the coexisting glomerular structures such as mesangial regions exacerbate the difficulties of this task. In this paper, we introduce a scheme to recognize fine-grained glomeruli lesions from whole slide images. First, a focal instance structural similarity loss is proposed to drive the model to locate all types of glomeruli precisely. Then an Uncertainty Aided Apportionment Network is designed to carry out the fine-grained visual classification without bounding-box annotations. This double branch-shaped structure extracts common features of the child class from the parent class and produces the uncertainty factor for reconstituting the training dataset. Results of slide-wise evaluation illustrate the effectiveness of the entire scheme, with an 8-22% improvement of the mean Average Precision compared with remarkable detection methods. The comprehensive results clearly demonstrate the effectiveness of the proposed method., Comment: 33 pages, 6 figures, accepted by the Pattern Recognition journal

Early exiting allows instances to exit at different layers according to the estimation of difficulty. Previous works usually adopt heuristic metrics such as the entropy of internal outputs to measure instance difficulty, which suffers from generalization and threshold-tuning. In contrast, learning to exit, or learning to predict instance difficulty is a more appealing way. Though some effort has been devoted to employing such "learn-to-exit" modules, it is still unknown whether and how well the instance difficulty can be learned. As a response, we first conduct experiments on the learnability of instance difficulty, which demonstrates that modern neural models perform poorly on predicting instance difficulty. Based on this observation, we propose a simple-yet-effective Hash-based Early Exiting approach (HashEE) that replaces the learn-to-exit modules with hash functions to assign each token to a fixed exiting layer. Different from previous methods, HashEE requires no internal classifiers nor extra parameters, and therefore is more efficient. Experimental results on classification, regression, and generation tasks demonstrate that HashEE can achieve higher performance with fewer FLOPs and inference time compared with previous state-of-the-art early exiting methods., Comment: Accepted to Findings of ACL 2022

Anti-cancer drug discoveries have been serendipitous, we sought to present the Open Molecular Graph Learning Benchmark, named CandidateDrug4Cancer, a challenging and realistic benchmark dataset to facilitate scalable, robust, and reproducible graph machine learning research for anti-cancer drug discovery. CandidateDrug4Cancer dataset encompasses multiple most-mentioned 29 targets for cancer, covering 54869 cancer-related drug molecules which are ranged from pre-clinical, clinical and FDA-approved. Besides building the datasets, we also perform benchmark experiments with effective Drug Target Interaction (DTI) prediction baselines using descriptors and expressive graph neural networks. Experimental results suggest that CandidateDrug4Cancer presents significant challenges for learning molecular graphs and targets in practical application, indicating opportunities for future researches on developing candidate drugs for treating cancers., Comment: Accepted by Workshop on Graph Learning Benchmarks, The Web Conference 2021

Building damage detection after natural disasters like earthquakes is crucial for initiating effective emergency response actions. Remotely sensed very high spatial resolution (VHR) imagery can provide vital information due to their ability to map the affected buildings with high geometric precision. Many approaches have been developed to detect damaged buildings due to earthquakes. However, little attention has been paid to exploiting rich features represented in VHR images using Deep Neural Networks (DNN). This paper presents a novel superpixel based approach combining DNN and a modified segmentation method, to detect damaged buildings from VHR imagery. Firstly, a modified Fast Scanning and Adaptive Merging method is extended to create initial over-segmentation. Secondly, the segments are merged based on the Region Adjacent Graph (RAG), considered an improved semantic similarity criterion composed of Local Binary Patterns (LBP) texture, spectral, and shape features. Thirdly, a pre-trained DNN using Stacked Denoising Auto-Encoders called SDAE-DNN is presented, to exploit the rich semantic features for building damage detection. Deep-layer feature abstraction of SDAE-DNN could boost detection accuracy through learning more intrinsic and discriminative features, which outperformed other methods using state-of-the-art alternative classifiers. We demonstrate the feasibility and effectiveness of our method using a subset of WorldView-2 imagery, in the complex urban areas of Bhaktapur, Nepal, which was affected by the Nepal Earthquake of April 25, 2015.

Self-supervised learning has gradually emerged as a powerful technique for graph representation learning. However, transferable, generalizable, and robust representation learning on graph data still remains a challenge for pre-training graph neural networks. In this paper, we propose a simple and effective self-supervised pre-training strategy, named Pairwise Half-graph Discrimination (PHD), that explicitly pre-trains a graph neural network at graph-level. PHD is designed as a simple binary classification task to discriminate whether two half-graphs come from the same source. Experiments demonstrate that the PHD is an effective pre-training strategy that offers comparable or superior performance on 13 graph classification tasks compared with state-of-the-art strategies, and achieves notable improvements when combined with node-level strategies. Moreover, the visualization of learned representation revealed that PHD strategy indeed empowers the model to learn graph-level knowledge like the molecular scaffold. These results have established PHD as a powerful and effective self-supervised learning strategy in graph-level representation learning., Comment: accepted by the 30th International Joint Conference on Artificial Intelligence (IJCAI-21)

Background: The current clinical workflow for esophageal gross tumor volume (GTV) contouring relies on manual delineation of high labor-costs and interuser variability. Purpose: To validate the clinical applicability of a deep learning (DL) multi-modality esophageal GTV contouring model, developed at 1 institution whereas tested at multiple ones. Methods and Materials: We collected 606 esophageal cancer patients from four institutions. 252 institution-1 patients had a treatment planning-CT (pCT) and a pair of diagnostic FDG-PETCT; 354 patients from other 3 institutions had only pCT. A two-streamed DL model for GTV segmentation was developed using pCT and PETCT scans of a 148 patient institution-1 subset. This built model had the flexibility of segmenting GTVs via only pCT or pCT+PETCT combined. For independent evaluation, the rest 104 institution-1 patients behaved as unseen internal testing, and 354 institutions 2-4 patients were used for external testing. We evaluated manual revision degrees by human experts to assess the contour-editing effort. The performance of the deep model was compared against 4 radiation oncologists in a multiuser study with 20 random external patients. Contouring accuracy and time were recorded for the pre-and post-DL assisted delineation process. Results: Our model achieved high segmentation accuracy in internal testing (mean Dice score: 0.81 using pCT and 0.83 using pCT+PET) and generalized well to external evaluation (mean DSC: 0.80). Expert assessment showed that the predicted contours of 88% patients need only minor or no revision. In multi-user evaluation, with the assistance of a deep model, inter-observer variation and required contouring time were reduced by 37.6% and 48.0%, respectively. Conclusions: Deep learning predicted GTV contours were in close agreement with the ground truth and could be adopted clinically with mostly minor or no changes., Comment: 36 pages, 10 figures

In this work, we introduce a fast and accurate method for unsupervised 3D medical image registration. This work is built on top of a recent algorithm SAM, which is capable of computing dense anatomical/semantic correspondences between two images at the pixel level. Our method is named SAME, which breaks down image registration into three steps: affine transformation, coarse deformation, and deep deformable registration. Using SAM embeddings, we enhance these steps by finding more coherent correspondences, and providing features and a loss function with better semantic guidance. We collect a multi-phase chest computed tomography dataset with 35 annotated organs for each patient and conduct inter-subject registration for quantitative evaluation. Results show that SAME outperforms widely-used traditional registration techniques (Elastix FFD, ANTs SyN) and learning based VoxelMorph method by at least 4.7% and 2.7% in Dice scores for two separate tasks of within-contrast-phase and across-contrast-phase registration, respectively. SAME achieves the comparable performance to the best traditional registration method, DEEDS (from our evaluation), while being orders of magnitude faster (from 45 seconds to 1.2 seconds).

Lymph node station (LNS) delineation from computed tomography (CT) scans is an indispensable step in radiation oncology workflow. High inter-user variabilities across oncologists and prohibitive laboring costs motivated the automated approach. Previous works exploit anatomical priors to infer LNS based on predefined ad-hoc margins. However, without voxel-level supervision, the performance is severely limited. LNS is highly context-dependent - LNS boundaries are constrained by anatomical organs - we formulate it as a deep spatial and contextual parsing problem via encoded anatomical organs. This permits the deep network to better learn from both CT appearance and organ context. We develop a stratified referencing organ segmentation protocol that divides the organs into anchor and non-anchor categories and uses the former's predictions to guide the later segmentation. We further develop an auto-search module to identify the key organs that opt for the optimal LNS parsing performance. Extensive four-fold cross-validation experiments on a dataset of 98 esophageal cancer patients (with the most comprehensive set of 12 LNSs + 22 organs in thoracic region to date) are conducted. Our LNS parsing model produces significant performance improvements, with an average Dice score of 81.1% +/- 6.1%, which is 5.0% and 19.2% higher over the pure CT-based deep model and the previous representative approach, respectively.

TextVQA requires models to read and reason about text in images to answer questions about them. Specifically, models need to incorporate a new modality of text present in the images and reason over it to answer TextVQA questions. In this challenge, we use generative model T5 for TextVQA task. Based on pre-trained checkpoint T5-3B from HuggingFace repository, two other pre-training tasks including masked language modeling(MLM) and relative position prediction(RPP) are designed to better align object feature and scene text. In the stage of pre-training, encoder is dedicate to handle the fusion among multiple modalities: question text, object text labels, scene text labels, object visual features, scene visual features. After that decoder generates the text sequence step-by-step, cross entropy loss is required by default. We use a large-scale scene text dataset in pre-training and then fine-tune the T5-3B with the TextVQA dataset only., Comment: Winner of TextVQA 2021

Artificial Intelligence (AI), along with the recent progress in biomedical language understanding, is gradually changing medical practice. With the development of biomedical language understanding benchmarks, AI applications are widely used in the medical field. However, most benchmarks are limited to English, which makes it challenging to replicate many of the successes in English for other languages. To facilitate research in this direction, we collect real-world biomedical data and present the first Chinese Biomedical Language Understanding Evaluation (CBLUE) benchmark: a collection of natural language understanding tasks including named entity recognition, information extraction, clinical diagnosis normalization, single-sentence/sentence-pair classification, and an associated online platform for model evaluation, comparison, and analysis. To establish evaluation on these tasks, we report empirical results with the current 11 pre-trained Chinese models, and experimental results show that state-of-the-art neural models perform by far worse than the human ceiling. Our benchmark is released at \url{https://tianchi.aliyun.com/dataset/dataDetail?dataId=95414&lang=en-us}., Comment: Accepted by ACL 2022

Measuring lesion size is an important step to assess tumor growth and monitor disease progression and therapy response in oncology image analysis. Although it is tedious and highly time-consuming, radiologists have to work on this task by using RECIST criteria (Response Evaluation Criteria In Solid Tumors) routinely and manually. Even though lesion segmentation may be the more accurate and clinically more valuable means, physicians can not manually segment lesions as now since much more heavy laboring will be required. In this paper, we present a prior-guided dual-path network (PDNet) to segment common types of lesions throughout the whole body and predict their RECIST diameters accurately and automatically. Similar to [1], a click guidance from radiologists is the only requirement. There are two key characteristics in PDNet: 1) Learning lesion-specific attention matrices in parallel from the click prior information by the proposed prior encoder, named click-driven attention; 2) Aggregating the extracted multi-scale features comprehensively by introducing top-down and bottom-up connections in the proposed decoder, named dual-path connection. Experiments show the superiority of our proposed PDNet in lesion segmentation and RECIST diameter prediction using the DeepLesion dataset and an external test set. PDNet learns comprehensive and representative deep image features for our tasks and produces more accurate results on both lesion segmentation and RECIST diameter prediction.

Accurately segmenting a variety of clinically significant lesions from whole body computed tomography (CT) scans is a critical task on precision oncology imaging, denoted as universal lesion segmentation (ULS). Manual annotation is the current clinical practice, being highly time-consuming and inconsistent on tumor's longitudinal assessment. Effectively training an automatic segmentation model is desirable but relies heavily on a large number of pixel-wise labelled data. Existing weakly-supervised segmentation approaches often struggle with regions nearby the lesion boundaries. In this paper, we present a novel weakly-supervised universal lesion segmentation method by building an attention enhanced model based on the High-Resolution Network (HRNet), named AHRNet, and propose a regional level set (RLS) loss for optimizing lesion boundary delineation. AHRNet provides advanced high-resolution deep image features by involving a decoder, dual-attention and scale attention mechanisms, which are crucial to performing accurate lesion segmentation. RLS can optimize the model reliably and effectively in a weakly-supervised fashion, forcing the segmentation close to lesion boundary. Extensive experimental results demonstrate that our method achieves the best performance on the publicly large-scale DeepLesion dataset and a hold-out test set.

Landmark localization plays an important role in medical image analysis. Learning based methods, including CNN and GCN, have demonstrated the state-of-the-art performance. However, most of these methods are fully-supervised and heavily rely on manual labeling of a large training dataset. In this paper, based on a fully-supervised graph-based method, DAG, we proposed a semi-supervised extension of it, termed few-shot DAG, \ie five-shot DAG. It first trains a DAG model on the labeled data and then fine-tunes the pre-trained model on the unlabeled data with a teacher-student SSL mechanism. In addition to the semi-supervised loss, we propose another loss using JS divergence to regulate the consistency of the intermediate feature maps. We extensively evaluated our method on pelvis, hand and chest landmark detection tasks. Our experiment results demonstrate consistent and significant improvements over previous methods., Comment: 10 pages

Depending on the application, radiological diagnoses can be associated with high inter- and intra-rater variabilities. Most computer-aided diagnosis (CAD) solutions treat such data as incontrovertible, exposing learning algorithms to considerable and possibly contradictory label noise and biases. Thus, managing subjectivity in labels is a fundamental problem in medical imaging analysis. To address this challenge, we introduce auto-decoded deep latent embeddings (ADDLE), which explicitly models the tendencies of each rater using an auto-decoder framework. After a simple linear transformation, the latent variables can be injected into any backbone at any and multiple points, allowing the model to account for rater-specific effects on the diagnosis. Importantly, ADDLE does not expect multiple raters per image in training, meaning it can readily learn from data mined from hospital archives. Moreover, the complexity of training ADDLE does not increase as more raters are added. During inference each rater can be simulated and a 'mean' or 'greedy' virtual rating can be produced. We test ADDLE on the problem of liver steatosis diagnosis from 2D ultrasound (US) by collecting 46 084 studies along with clinical US diagnoses originating from 65 different raters. We evaluated diagnostic performance using a separate dataset with gold-standard biopsy diagnoses. ADDLE can improve the partial areas under the curve (AUCs) for diagnosing severe steatosis by 10.5% over standard classifiers while outperforming other annotator-noise approaches, including those requiring 65 times the parameters., Comment: Main body includes 10 pages and 3 figures

Bone mineral density (BMD) is a clinically critical indicator of osteoporosis, usually measured by dual-energy X-ray absorptiometry (DEXA). Due to the limited accessibility of DEXA machines and examinations, osteoporosis is often under-diagnosed and under-treated, leading to increased fragility fracture risks. Thus it is highly desirable to obtain BMDs with alternative cost-effective and more accessible medical imaging examinations such as X-ray plain films. In this work, we formulate the BMD estimation from plain hip X-ray images as a regression problem. Specifically, we propose a new semi-supervised self-training algorithm to train the BMD regression model using images coupled with DEXA measured BMDs and unlabeled images with pseudo BMDs. Pseudo BMDs are generated and refined iteratively for unlabeled images during self-training. We also present a novel adaptive triplet loss to improve the model's regression accuracy. On an in-house dataset of 1,090 images (819 unique patients), our BMD estimation method achieves a high Pearson correlation coefficient of 0.8805 to ground-truth BMDs. It offers good feasibility to use the more accessible and cheaper X-ray imaging for opportunistic osteoporosis screening.

How to produce expressive molecular representations is a fundamental challenge in AI-driven drug discovery. Graph neural network (GNN) has emerged as a powerful technique for modeling molecular data. However, previous supervised approaches usually suffer from the scarcity of labeled data and have poor generalization capability. Here, we proposed a novel Molecular Pre-training Graph-based deep learning framework, named MPG, that leans molecular representations from large-scale unlabeled molecules. In MPG, we proposed a powerful MolGNet model and an effective self-supervised strategy for pre-training the model at both the node and graph-level. After pre-training on 11 million unlabeled molecules, we revealed that MolGNet can capture valuable chemistry insights to produce interpretable representation. The pre-trained MolGNet can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of drug discovery tasks, including molecular properties prediction, drug-drug interaction, and drug-target interaction, involving 13 benchmark datasets. Our work demonstrates that MPG is promising to become a novel approach in the drug discovery pipeline.

Active learning generally involves querying the most representative samples for human labeling, which has been widely studied in many fields such as image classification and object detection. However, its potential has not been explored in the more complex instance segmentation task that usually has relatively higher annotation cost. In this paper, we propose a novel and principled semi-supervised active learning framework for instance segmentation. Specifically, we present an uncertainty sampling strategy named Triplet Scoring Predictions (TSP) to explicitly incorporate samples ranking clues from classes, bounding boxes and masks. Moreover, we devise a progressive pseudo labeling regime using the above TSP in semi-supervised manner, it can leverage both the labeled and unlabeled data to minimize labeling effort while maximize performance of instance segmentation. Results on medical images datasets demonstrate that the proposed method results in the embodiment of knowledge from available data in a meaningful way. The extensive quantitatively and qualitatively experiments show that, our method can yield the best-performing model with notable less annotation costs, compared with state-of-the-arts., Comment: 13 pages, 7 figures, accepted for presentation at BMVC2020

Although BERT based relation classification (RC) models have achieved significant improvements over the traditional deep learning models, it seems that no consensus can be reached on what is the optimal architecture. Firstly, there are multiple alternatives for entity span identification. Second, there are a collection of pooling operations to aggregate the representations of entities and contexts into fixed length vectors. Third, it is difficult to manually decide which feature vectors, including their interactions, are beneficial for classifying the relation types. In this work, we design a comprehensive search space for BERT based RC models and employ neural architecture search (NAS) method to automatically discover the design choices mentioned above. Experiments on seven benchmark RC tasks show that our method is efficient and effective in finding better architectures than the baseline BERT based RC model. Ablation study demonstrates the necessity of our search space design and the effectiveness of our search method.

Though the transformer architectures have shown dominance in many natural language understanding tasks, there are still unsolved issues for the training of transformer models, especially the need for a principled way of warm-up which has shown importance for stable training of a transformer, as well as whether the task at hand prefer to scale the attention product or not. In this paper, we empirically explore automating the design choices in the transformer model, i.e., how to set layer-norm, whether to scale, number of layers, number of heads, activation function, etc, so that one can obtain a transformer architecture that better suits the tasks at hand. RL is employed to navigate along search space, and special parameter sharing strategies are designed to accelerate the search. It is shown that sampling a proportion of training data per epoch during search help to improve the search quality. Experiments on the CoNLL03, Multi-30k, IWSLT14 and WMT-14 shows that the searched transformer model can outperform the standard transformers. In particular, we show that our learned model can be trained more robustly with large learning rates without warm-up., Comment: will add new technical contents

Representation learning (RL) plays an important role in extracting proper representations from complex medical data for various analyzing tasks, such as patient grouping, clinical endpoint prediction and medication recommendation. Medical data can be divided into two typical categories, outpatient and inpatient, that have different data characteristics. However, few of existing RL methods are specially designed for inpatients data, which have strong temporal relations and consistent diagnosis. In addition, for unordered medical activity set, existing medical RL methods utilize a simple pooling strategy, which would result in indistinguishable contributions among the activities for learning. In this work, weproposeInpatient2Vec, anovelmodel for learning three kinds of representations for inpatient, including medical activity, hospital day and diagnosis. A multi-layer self-attention mechanism with two training tasks is designed to capture the inpatient data characteristics and process the unordered set. Using a real-world dataset, we demonstrate that the proposed approach outperforms the competitive baselines on semantic similarity measurement and clinical events prediction tasks.

Precision medicine requires the precision disease risk prediction models. In literature, there have been a lot well-established (inter-)national risk models, but when applying them into the local population, the prediction performance becomes unsatisfactory. To address the localization issue, this paper exploits the way to develop knowledge-enhanced localized risk models. On the one hand, we tune models by learning from regional Electronic Health Record (EHR) repositories, and on the other hand, we propose knowledge injection into the EHR data learning process. For experiments, we leverage the Pooled Cohort Equations (PCE, as recommended in ACC/AHA guidelines to estimate the risk of ASCVD) to develop a localized ASCVD risk prediction model in diabetes. The experimental results show that, if directly using the PCE algorithm on our cohort, the AUC is only 0.653, while our knowledge-enhanced localized risk model can achieve higher prediction performance with AUC of 0.723 (improved by 10.7%).