Task Dataset Research Articles

Selective Layer Tuning and Performance Study of Pre-Trained Models Using Genetic Algorithm

Utilizing pre-trained models involves fully or partially using pre-trained parameters as initialization. In general, configuring a pre-trained model demands practitioners’ knowledge about problems or an exhaustive trial–error experiment according to a given task. In this paper, we propose tuning trainable layers using a genetic algorithm on a pre-trained model that is fine-tuned on single-channel image datasets for a classification task. The single-channel dataset comprises images from grayscale and preprocessed audio signals transformed into a log-Mel spectrogram. Four deep-learning models used in the experimental evaluation employed the pre-trained model with the ImageNet dataset. The proposed genetic algorithm was applied to find the highest fitness for every generation to determine the selective layer tuning of the pre-trained models. Compared to the conventional fine-tuning method and random layer search, our proposed selective layer search with a genetic algorithm achieves higher accuracy, on average, by 9.7% and 1.88% (MNIST-Fashion), 1.31% and 1.14% (UrbanSound8k), and 2.2% and 0.29% (HospitalAlarmSound), respectively. In addition, our searching method can naturally be applied to various datasets of the same task without prior knowledge about the dataset of interest.

Multi-way relation-enhanced hypergraph representation learning for anti-cancer drug synergy prediction.

Drug combinations have exhibited promise in treating cancers with less toxicity and fewer adverse reactions. However, in vitro screening of synergistic drug combinations is time-consuming and labor-intensive because of the combinatorial explosion. Although a number of computational methods have been developed for predicting synergistic drug combinations, the multi-way relations between drug combinations and cell lines existing in drug synergy data have not been well exploited. We propose a multi-way relation-enhanced hypergraph representation learning method to predict anti-cancer drug synergy, named HypergraphSynergy. HypergraphSynergy formulates synergistic drug combinations over cancer cell lines as a hypergraph, in which drugs and cell lines are represented by nodes and synergistic drug-drug-cell line triplets are represented by hyperedges, and leverages the biochemical features of drugs and cell lines as node attributes. Then, a hypergraph neural network is designed to learn the embeddings of drugs and cell lines from the hypergraph and predict drug synergy. Moreover, the auxiliary task of reconstructing the similarity networks of drugs and cell lines is considered to enhance the generalization ability of the model. In the computational experiments, HypergraphSynergy outperforms other state-of-the-art synergy prediction methods on two benchmark datasets for both classification and regression tasks and is applicable to unseen drug combinations or cell lines. The studies revealed that the hypergraph formulation allows us to capture and explain complex multi-way relations of drug combinations and cell lines, and also provides a flexible framework to make the best use of diverse information. The source data and codes of HypergraphSynergy can be freely downloaded from https://github.com/liuxuan666/HypergraphSynergy. Supplementary data are available at Bioinformatics online.

Robust Feature Selection-Based Speech Emotion Classification Using Deep Transfer Learning

Speech Emotion Classification (SEC) relies heavily on the quality of feature extraction and selection from the speech signal. Improvement on this to enhance the classification of emotion had attracted significant attention from researchers. Many primitives and algorithmic solutions for efficient SEC with minimum cost have been proposed; however, the accuracy and performance of these methods have not yet attained a satisfactory point. In this work, we proposed a novel deep transfer learning approach with distinctive emotional rich feature selection techniques for speech emotion classification. We adopt mel-spectrogram extracted from speech signal as the input to our deep convolutional neural network for efficient feature extraction. We froze 19 layers of our pretrained convolutional neural network from re-training to increase efficiency and minimize computational cost. One flattened layer and two dense layers were used. A ReLu activation function was used at the last layer of our feature extraction segment. To prevent misclassification and reduce feature dimensionality, we employed the Neighborhood Component Analysis (NCA) feature selection algorithm for picking out the most relevant features before the actual classification of emotion. Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) classifiers were utilized at the topmost layer of our model. Two popular datasets for speech emotion classification tasks were used, which are: Berling Emotional Speech Database (EMO-DB), and Toronto English Speech Set (TESS), and a combination of EMO-DB with TESS was used in our experiment. We obtained a state-of-the-art result with an accuracy rate of 94.3%, 100% specificity on EMO-DB, and 97.2%, 99.80% on TESS datasets, respectively. The performance of our proposed method outperformed some recent work in SEC after assessment on the three datasets.

A scoping review of publicly available language tasks in clinical natural language processing.

To provide a scoping review of papers on clinical natural language processing (NLP) shared tasks that use publicly available electronic health record data from a cohort of patients. We searched 6 databases, including biomedical research and computer science literature databases. A round of title/abstract screening and full-text screening were conducted by 2 reviewers. Our method followed the PRISMA-ScR guidelines. A total of 35 papers with 48 clinical NLP tasks met inclusion criteria between 2007 and 2021. We categorized the tasks by the type of NLP problems, including named entity recognition, summarization, and other NLP tasks. Some tasks were introduced as potential clinical decision support applications, such as substance abuse detection, and phenotyping. We summarized the tasks by publication venue and dataset type. The breadth of clinical NLP tasks continues to grow as the field of NLP evolves with advancements in language systems. However, gaps exist with divergent interests between the general domain NLP community and the clinical informatics community for task motivation and design, and in generalizability of the data sources. We also identified issues in data preparation. The existing clinical NLP tasks cover a wide range of topics and the field is expected to grow and attract more attention from both general domain NLP and clinical informatics community. We encourage future work to incorporate multidisciplinary collaboration, reporting transparency, and standardization in data preparation. We provide a listing of all the shared task papers and datasets from this review in a GitLab repository.

Reproducing FSL's fMRI data analysis via Nipype: Relevance, challenges, and solutions.

The "replication crisis" in neuroscientific research has led to calls for improving reproducibility. In traditional neuroscience analyses, irreproducibility may occur as a result of issues across various stages of the methodological process. For example, different operating systems, different software packages, and even different versions of the same package can lead to variable results. Nipype, an open-source Python project, integrates different neuroimaging software packages uniformly to improve the reproducibility of neuroimaging analyses. Nipype has the advantage over traditional software packages (e.g., FSL, ANFI, SPM, etc.) by (1) providing comprehensive software development frameworks and usage information, (2) improving computational efficiency, (3) facilitating reproducibility through sufficient details, and (4) easing the steep learning curve. Despite the rich tutorials it has provided, the Nipype community lacks a standard three-level GLM tutorial for FSL. Using the classical Flanker task dataset, we first precisely reproduce a three-level GLM analysis with FSL via Nipype. Next, we point out some undocumented discrepancies between Nipype and FSL functions that led to substantial differences in results. Finally, we provide revised Nipype code in re-executable notebooks that assure result invariability between FSL and Nipype. Our analyses, notebooks, and operating software specifications (e.g., docker build files) are available on the Open Science Framework platform.

A Photovoltaic Power Predicting Model Using the Differential Evolution Algorithm and Multi-Task Learning

The rapid growth of photovoltaic installed capacity exacerbates the power management challenges faced by photovoltaic power stations, emphasizing the importance of accurate and stable photovoltaic generation forecasting. As a result, researchers conducted research and developed several photovoltaic power prediction models. However, many prediction models focus exclusively on the algorithm structure in order to improve model accuracy, oblivious to how the dataset is constructed and divided for the prediction model. This paper proposes a comprehensive model to address this gap. To be more precise, the differential evolution algorithm is constantly looking for optimal values between different populations and determining the best way to construct datasets for prediction tasks. Multi-task learning enables the transfer of knowledge between related tasks via parameter sharing layers, referring to the accuracy and stability of prediction models. Overall, the proposed model achieves high prediction accuracy and stability. The prediction error of the proposed model is less than 450W in RMSE, NRMSE is less than 2.5%, and R-Square is greater than 99% in multiple prediction tasks. Additionally, when compared to other single-task prediction models with an R-Square greater than 96%, the proposed model further reduces the root mean squared error by an average of 28% and the standard deviation of root mean squared error by 54%.

CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks

General-purpose robots coexisting with humans in their environment must learn to relate human language to their perceptions and actions to be useful in a range of daily tasks. Moreover, they need to acquire a diverse repertoire of general-purpose skills that allow composing long-horizon tasks by following unconstrained language instructions. In this letter, we present Composing Actions from Language and Vision (CALVIN) ( <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</u> omposing <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</u> ctions from <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">L</u> anguage and <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Vi</u> sio <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">n</u> ), an open-source simulated benchmark to learn long-horizon language-conditioned tasks. Our aim is to make it possible to develop agents that can solve many robotic manipulation tasks over a long horizon, from onboard sensors, and specified only via human language. CALVIN tasks are more complex in terms of sequence length, action space, and language than existing vision-and-language task datasets and supports flexible specification of sensor suites. We evaluate the agents in zero-shot to novel language instructions and to novel environments. We show that a baseline model based on multi-context imitation learning performs poorly on CALVIN, suggesting that there is significant room for developing innovative agents that learn to relate human language to their world models with this benchmark.

Hierarchical Decision Granules Optimization through the Principle of Justifiable Granularity

Interpretable Machine Learning (IML) aims to establish more transparent decision processes where the human can understand the reason behind the models’ decisions. In this work a methodology to create intrinsically interpretable models based on fuzzy rules is proposed. There is a selection to identify the rule structure by extracting the most significant elements from a decision tree by the principle of justifiable granularity. There are defined hierarchical decision granules and their quality metrics. The proposal is evaluated with ten publicly available datasets for classification tasks. It is shown that through the principle of justified granularity, rule-based models can be greatly compressed through their fuzzy representation, not only without significantly losing performance but even with compression of 40% it manages to exceed the performance of the initial model.

Unsupervised Sentence Representation via Contrastive Learning with Mixing Negatives

Unsupervised sentence representation learning is a fundamental problem in natural language processing. Recently, contrastive learning has made great success on this task. Existing constrastive learning based models usually apply random sampling to select negative examples for training. Previous work in computer vision has shown that hard negative examples help contrastive learning to achieve faster convergency and better optimization for representation learning. However, the importance of hard negatives in contrastive learning for sentence representation is yet to be explored. In this study, we prove that hard negatives are essential for maintaining strong gradient signals in the training process while random sampling negative examples is ineffective for sentence representation. Accordingly, we present a contrastive model, MixCSE, that extends the current state-of-the-art SimCSE by continually constructing hard negatives via mixing both positive and negative features. The superior performance of the proposed approach is demonstrated via empirical studies on Semantic Textual Similarity datasets and Transfer task datasets.

Polarity and Subjectivity Detection with Multitask Learning and BERT Embedding

In recent years, deep learning-based sentiment analysis has received attention mainly because of the rise of social media and e-commerce. In this paper, we showcase the fact that the polarity detection and subjectivity detection subtasks of sentiment analysis are inter-related. To this end, we propose a knowledge-sharing-based multitask learning framework. To ensure high-quality knowledge sharing between the tasks, we use the Neural Tensor Network, which consists of a bilinear tensor layer that links the two entity vectors. We show that BERT-based embedding with our MTL framework outperforms the baselines and achieves a new state-of-the-art status in multitask learning. Our framework shows that the information across datasets for related tasks can be helpful for understanding task-specific features.

Occupational sex composition and the relative pay of managerial work

ABSTRACT This article examines a pathway whereby occupational sex composition widens gender gap in pay, asking how differently gender tokens—men and women in the numeric minority in opposite-sex occupations—derive monetary rewards from engaging in managerial tasks. Using a dataset of job tasks from the Princeton Data Improvement Initiative, I find that a large portion of the cross-occupation variation in wage returns to managerial work is attributed to the sex composition of occupations. Female tokens derive lower returns to managerial tasks compared to their non-token counterparts, whereas male tokens benefit from managerial work to a greater extent than when they are in mixed-sex or male-dominated occupations. The study concludes by discussing the gendered outcomes of engaging in managerial tasks, typically sex-typed as being “men’s work,” in gender-atypical work environments.

CDANet: Common-and-Differential Attention Network for Object Detection and Instance Segmentation

• A novel Common-and-Differential Attention Network (CDANet) is proposed to enhance feature representations. • Common-and-Differential operations are introduced to restrain redundant and confusing background interference. • The effectiveness of CDANet is validated on object detection and instance segmentation tasks . In this paper, we propose a simple and effective C ommon-and- D ifferential A ttention N etwork ( CDANet ) for object detection and instance segmentation. For an input intermediate feature map, CDANet infers parallelly attention modules along channel and spatial dimensions respectively, then both attention modules are multiplied with the input feature map for the refined features. Specially, since redundant and confusing background may misdirect the localization at object boundary, our attention network applies Common-and-Differential operations to weaken useless background interference and focus on meaningful object features. The proposed CDANet is verified performance through comprehensive experiments on PASCAL VOC2007 and MS COCO2017 datasets for object detection and instance segmentation tasks. CDANet obtains consistently improved results on various detectors with different backbones, indicating the significant effectiveness and applicability of CDANet.

Adversarial training for supervised relation extraction

Most supervised methods for relation extraction (RE) involve time-consuming human annotation. Distant supervision for RE is an efficient method to obtain large corpora that contains thousands of instances and various relations. However, the existing approaches rely heavily on knowledge bases (e.g., Freebase), thereby introducing data noise. Various relations and noisy labeling instances make the issue difficult to solve. In this study, we propose a model based on a piecewise convolution neural network with adversarial training. Inspired by generative adversarial networks, we adopt a heuristic algorithm to identify noisy datasets and apply adversarial training to RE. Experiments on the extended dataset of SemEval-2010 Task 8 show that our model can obtain more accurate training data for RE and significantly outperforms several competitive baseline models. Our model has an Fi score of 89.61%.

VDM-DA: Virtual Domain Modeling for Source Data-Free Domain Adaptation

Domain adaptation aims to leverage a label-rich domain (the source domain) to help model learning in a label-scarce domain (the target domain). Most domain adaptation methods require the co-existence of source and target domain samples to reduce the distribution mismatch. However, access to the source domain samples may not always be feasible in real-world applications due to different problems ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e.g.</i> , storage, transmission, and privacy issues). In this work, we deal with the source data-free unsupervised domain adaptation problem and propose a novel approach referred to as Virtual Domain Modeling for Domain Adaptation (VDM-DA), in which the virtual domain acts as a bridge between the source and target domains. Specifically, based on the pre-trained source model, we generate the virtual domain samples by using an approximated Gaussian Mixture Model (GMM) in the feature space, such that the virtual domain maintains a similar distribution with the source domain without access to the original source data. Moreover, we also design an effective distribution alignment method to reduce the distribution divergence between the virtual domain and the target domain by gradually improving the compactness of the target domain distribution through model learning. In this way, we successfully achieve the goal of distribution alignment between the source and target domains when training deep networks without access to the source domain data. We conduct extensive experiments on four benchmark datasets for both 2D image-based and 3D point cloud-based cross-domain object recognition tasks, where the proposed method referred to as Virtual Domain Modeling for Domain Adaptation (VDM-DA) achieves the promising performance on all datasets.

Computational Models of Linguistic Alignment for Clustering Group Participants and Predicting Task Outcomes

This work examines the relationships between several measures of linguistic alignment in task-based group conversation, and assesses how useful these measures are for predicting task performance and participant affect. The study is carried out using human-human conversational data, with potential implications for human-AI conversations where an artificial agent can decide if and how to align itself linguistically with human subjects. We implement several alignment measures including long-term measures that assess the level of convergence over the course of the conversation, as well as short-term coordination measures that have been related in previous research to power dynamics. The study uses two publicly available English-language survival task datasets. After analyzing correlations between the various linguistic alignment measures, we perform clustering in order to unveil the main types of alignment patterns that are prevalent in the data. Finally, we use the alignment measures as machine learning features to predict participant task performance and participant affect.

Progressive Training Technique with Weak-Label Boosting for Fine-Grained Classification on Unbalanced Training Data

In practical classification tasks, the sample distribution of the dataset is often unbalanced; for example, this is the case in a dataset that contains a massive quantity of samples with weak labels and for which concrete identification is unavailable. Even in samples with exact labels, the number of samples corresponding to many labels is small, resulting in difficulties in learning the concepts through a small number of labeled samples. In addition, there is always a small interclass variance and a large intraclass variance among categories. Weak labels, few-shot problems, and fine-grained analysis are the key challenges affecting the performance of the classification model. In this paper, we develop a progressive training technique to address the few-shot challenge, along with a weak-label boosting method, by considering all of the weak IDs as negative samples of every predefined ID in order to take full advantage of the more numerous weak-label data. We introduce an instance-aware hard ID mining strategy in the classification loss and then further develop the global and local feature-mapping loss to expand the decision margin. We entered the proposed method into the Kaggle competition, which aims to build an algorithm to identify individual humpback whales in images. With a few other common training tricks, the proposed approach won first place in the competition. All three problems (weak labels, few-shot problems, and fine-grained analysis) exist in the dataset used in the competition. Additionally, we applied our method to CUB-2011 and Cars-196, which are the most widely-used datasets for fine-grained visual categorization tasks, and achieved respective accuracies of 90.1% and 94.9%. This experiment shows that the proposed method achieves the optimal effect compared with other common baselines, and verifies the effectiveness of our method. Our solution has been made available as an open source project.

NEAR: Named entity and attribute recognition of clinical concepts

Named Entity Recognition (NER) or the extraction of concepts from clinical text is the task of identifying entities in text and slotting them into categories such as problems, treatments, tests, clinical departments, occurrences (such as admission and discharge) and others. NER forms a critical component of processing and leveraging unstructured data from Electronic Health Records (EHR). While identifying the spans and categories of concepts is itself a challenging task, these entities could also have attributes such as negation that pivot their meanings implied to the consumers of the named entities. There has been little research dedicated to identifying the entities and their qualifying attributes together. This research hopes to contribute to the area of detecting entities and their corresponding attributes by modelling the NER task as a supervised, multi-label tagging problem with each of the attributes assigned tagging sequence labels. In this paper, we propose 3 architectures to achieve this multi-label entity tagging: BiLSTM n-CRF, BiLSTM-CRF-Smax-TF and BiLSTM n-CRF-TF. We evaluate these methods on the 2010 i2b2/VA and the i2b2 2012 shared task datasets. Our different models obtain best NER scores of 0.903 and 0.808 on the i2b2 2010/VA and i2b2 2012 respectively. The highest span based micro-averaged F1 polarity scores obtained were 0.832 and 0.836 on the i2b2 2010/VA and i2b2 2012 datasets respectively, and the highest macro-averaged F1 polarity scores obtained were 0.924 and 0.888 respectively. The modality studies conducted on i2b2 2012 dataset revealed high scores of 0.818 and 0.501 for span based micro-averaged F1 and macro-averaged F1 respectively.

Challenges of Vehicle Classification Using Acoustics

Automated acoustic classification of vehicles is a challenging problem with many variables. Vehicles produce complex sounds from many sources, sound signatures vary between similar vehicles, and background noise has a large impact on audio data. Past research studied different vehicle classification techniques but often relied on datasets with little variation in vehicle model, environmental conditions, or microphones. High-accuracy results on these datasets suggest issues of overfitting. This paper highlights the challenges of creating robust datasets for difficult acoustic classification tasks. Challenges include collecting on multiple instances of a target class, recording with varied environmental noise, and collecting across multiple types of microphones. The work presented also evaluates classification performance on combinations of acoustic feature sets and common machine learning algorithms. Classification F-score performance drops 38.67% when the test set has background noise that differs from the classifier’s training set. Performance also significantly drops when the classifier is tested on a vehicle of a model year not included in the training set. Lastly, the vehicle classifier heavily overfits on the signatures of the microphones it is trained on.

SIMCD: SIMulated crowd data for anomaly detection and prediction

Smart Crowd management (SCM) solutions can mitigate overcrowding disasters by implementing efficient crowd learning models that can anticipate critical crowd conditions and potential catastrophes. Developing an SCM solution involves monitoring crowds and modelling their dynamics. Crowd monitoring produces vast amounts of data, with features such as densities and speeds, which are essential for training and evaluating crowd learning models. By and large, crowd datasets can be classified as real (e.g., real monitoring of crowds) or synthetic (e.g., simulation of crowds). Using real crowd datasets can produce effective and reliable crowd learning models. However, acquiring real crowd data faces several challenges, including the expensive installation of a sensory infrastructure, the data pre-processing costs and the lack of real datasets that cover particular crowd scenarios. Consequently, crowd management literature has adopted simulation tools for generating synthetic datasets to overcome the challenges associated with their real counterparts. The majority of existing datasets, whether real or synthetic, can be used for crowd counting applications or analysing the activities of individuals rather than collective crowd behaviour. Accordingly, this paper demonstrates the process of generating bespoke synthetic crowd datasets that can be used for crowd anomaly detection and prediction, using the MassMotion crowd simulator. The developed datasets present two types of crowd anomalies; namely, high densities and contra-flow walking direction. These datasets are: SIMulated Crowd Data (SIMCD)-Single Anomaly and SIMCD-Multiple Anomalies for anomaly detection tasks, besides two SIMCD-Prediction datasets for crowd prediction tasks. Furthermore, the paper demonstrates the data preparation (pre-processing) process by aggregating the data and proposing new essential features, such as the level of crowdedness and the crowd severity level, that are useful for developing crowd prediction and anomaly detection models.

3DGT-DDI: 3D graph and text based neural network for drug-drug interaction prediction.

Drug-drug interactions (DDIs) occur during the combination of drugs. Identifying potential DDI helps us to study the mechanism behind the combination medication or adverse reactions so as to avoid the side effects. Although many artificial intelligence methods predict and mine potential DDI, they ignore the 3D structure information of drug molecules and do not fully consider the contribution of molecular substructure in DDI. We proposed a new deep learning architecture, 3DGT-DDI, a model composed of a 3D graph neural network and pre-trained text attention mechanism. We used 3D molecular graph structure and position information to enhance the prediction ability of the model for DDI, which enabled us to deeply explore the effect of drug substructure on DDI relationship. The results showed that 3DGT-DDI outperforms other state-of-the-art baselines. It achieved an 84.48% macro F1 score in the DDIExtraction 2013 shared task dataset. Also, our 3D graph model proves its performance and explainability through weight visualization on the DrugBank dataset. 3DGT-DDI can help us better understand and identify potential DDI, thereby helping to avoid the side effects of drug mixing. The source code and data are available at https://github.com/hehh77/3DGT-DDI.