Hybrid Transformer/CTC Networks for Hardware Efficient Voice Triggering

Recent research in disaster informatics demonstrates a practical and important use case of artificial intelligence to save human lives and suffering during natural disasters based on social media contents (text and images). While notable progress has been made using texts, research on exploiting the images remains relatively under-explored. To advance image-based approaches, we propose MEDIC (https://crisisnlp.qcri.org/medic/index.html), which is the largest social media image classification dataset for humanitarian response consisting of 71,198 images to address four different tasks in a multi-task learning setup. This is the first dataset of its kind: social media images, disaster response, and multi-task learning research. An important property of this dataset is its high potential to facilitate research on multi-task learning, which recently receives much interest from the machine learning community and has shown remarkable results in terms of memory, inference speed, performance, and generalization capability. Therefore, the proposed dataset is an important resource for advancing image-based disaster management and multi-task machine learning research. We experiment with different deep learning architectures and report promising results, which are above the majority baselines for all tasks. Along with the dataset, we also release all relevant scripts (https://github.com/firojalam/medic).

Using multiple treebanks to improve parsing performance has shown positive results. However, to what extent similar, yet competing annotation decisions play in parser behavior is unclear. We investigate this within a multi-task learning (MTL) dependency parser setup on two parallel treebanks, UD and SUD, which, while possessing similar annotation schemes, differ in specific linguistic annotation preferences. We perform a set of experiments with different MTL architectural choices, comparing performance across various input embeddings. We find languages tend to pattern in lose typological associations, but generally the performance within an MTL setting is lower than single model baseline parsers for each annotation scheme. The main contributing factor seems to be the competing syntactic annotation information shared between treebanks in an MTL setting, which is shown in experiments against differently annotated treebanks.This suggests that the impact of how the signal is encoded for annotations and its influence on possible negative transfer is more important than that of the input embeddings in an MTL setting.

In breast cancer screening, binary classification of mammograms is a common task aiming to determine whether a case is malignant or benign. A Computer-Aided Diagnosis (CADx) system based on a trainable classifier requires clean data and labels coming from a confirmed diagnosis. Unfortunately, such labels are not easy to obtain in clinical practice, since the histopathological reports of biopsy may not be available alongside mammograms, while normal cases may not have an explicit follow-up confirmation. Such ambiguities result either in reducing the number of samples eligible for training or in a label uncertainty that may decrease the performances. In this work, we maximize the number of samples for training relying on multi-task learning. We design a deep-neural-network-based classifier yielding multiple outputs in one forward pass. The predicted classes include binary malignancy, cancer probability estimation, breast density, and image laterality. Since few samples have all classes available and confirmed, we propose to introduce the uncertainty related to the classes as a per-sample weight during training. Such weighting prevents updating the network's parameters when training on uncertain or missing labels. We evaluate our approach on the public INBreast and private datasets, showing statistically significant improvements compared to baseline and independent state-of-the-art approaches. Moreover, we use mammograms from Susan G. Komen Tissue Bank for fine-tuning, further demonstrating the ability to improve the performances in our multi-task learning setup from raw clinical data. We achieved the binary classification performance of AUC = 80.46 on our private dataset and AUC = 85.23 on the INBreast dataset.

We propose a novel neural label embedding (NLE) scheme for the domain adaptation of a deep neural network (DNN) acoustic model with unpaired data samples from source and target domains. With NLE method, we distill the knowledge from a powerful source-domain DNN into a dictionary of label embeddings, or l-vectors, one for each senone class. Each l-vector is a representation of the senone-specific output distributions of the source-domain DNN and is learned to minimize the average L <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> , Kullback-Leibler (KL) or symmetric KL distance to the output vectors with the same label through simple averaging or standard back-propagation. During adaptation, the l-vectors serve as the soft targets to train the target-domain model with cross-entropy loss. Without parallel data constraint as in the teacher-student learning, NLE is specially suited for the situation where the paired target-domain data cannot be simulated from the source-domain data. We adapt a 6400 hours multi-conditional US English acoustic model to each of the 9 accented English (80 to 830 hours) and kids' speech (80 hours). NLE achieves up to 14.1% relative word error rate reduction over direct re-training with one-hot labels.

A new hierarchical nonparametric Bayesian framework is proposed for the problem of multi-task learning (MTL) with sequential data. The models for multiple tasks, each characterized by sequential data, are learned jointly, and the intertask relationships are obtained simultaneously. This MTL setting is used to analyze and sort large databases composed of sequential data, such as music clips. Within each data set, we represent the sequential data with an infinite hidden Markov model (iHMM), avoiding the problem of model selection (selecting a number of states). Across the data sets, the multiple iHMMs are learned jointly in a MTL setting, employing a nested Dirichlet process (nDP). The nDP-iHMM MTL method allows simultaneous task-level and data-level clustering, with which the individual iHMMs are enhanced and the between-task similarities are learned. Therefore, in addition to improved learning of each of the models via appropriate data sharing, the learned sharing mechanisms are used to infer interdata relationships of interest for data search. Specifically, the MTL-learned task-level sharing mechanisms are used to define the affinity matrix in a graph-diffusion sorting framework. To speed up the MCMC inference for large databases, the nDP-iHMM is truncated to yield a nested Dirichlet-distribution based HMM representation, which accommodates fast variational Bayesian (VB) analysis for large-scale inference, and the effectiveness of the framework is demonstrated using a database composed of 2500 digital music pieces.

Session-based recommendations aim to predict a user’s next click based on the user’s current and historical sessions, which can be applied to shopping websites and APPs. Existing session-based recommendation methods cannot accurately capture the complex transitions between items. In addition, some approaches compress sessions into a fixed representation vector without taking into account the user’s interest preferences at the current moment, thus limiting the accuracy of recommendations. Considering the diversity of items and users’ interests, a personalized interest attention graph neural network (PIA-GNN) is proposed for session-based recommendation. This approach utilizes personalized graph convolutional networks (PGNN) to capture complex transitions between items, invoking an interest-aware mechanism to activate users’ interest in different items adaptively. In addition, a self-attention layer is used to capture long-term dependencies between items when capturing users’ long-term preferences. In this paper, the cross-entropy loss is used as the objective function to train our model. We conduct rich experiments on two real datasets, and the results show that PIA-GNN outperforms existing personalized session-aware recommendation methods.

Glassy materials are prevalent in many modern-day applications, which place ever new demands on the materials' properties. However, because of their amorphous nature, the design of glasses with specific properties is inherently difficult and necessarily data-driven. Due to the large size of the compositional space and the cost of the manufacturing process, glass data is only sparsely available. It is thus important to design and train models which maximize the benefit of each individual training sample on the models' resulting prediction quality. In this paper, we study a multitask learning approach to meet this challenge. We analyze its effect on the sample efficiency, that is, the dependence of the test error on the training set size of a neural network model for predicting properties of oxide glasses. In contrast to existing encoder-based multitask learning models in computational glass science, we propose a decoder-based model, consisting of task embeddings in the input layer and a subsequent shared network architecture. This allows us to apply well-established single-task models in a multitask learning setting. We demonstrate in a series of numerical experiments, predicting the Young's modulus and the log-viscosity, respectively, that this approach notably improves the model's sample efficiency. To deal with parametric target properties, we complement the task embeddings by a parameter embedding and show that this preserves the advantageous multitask learning effect. In all experiments, we observe an algebraic scaling law for the decay of the test error with respect to the number of training samples.

Speech emotion recognition is a challenging task in natural language processing. It relies heavily on the effectiveness of speech features and acoustic models. However, existing acoustic models may not handle speech emotion recognition efficiently for their built-in limitations. In this work, a novel deep-learning acoustic model called attention-based skip convolution bi-directional long short-term memory, abbreviated as SCBAMM, is proposed to recognize speech emotion. It has eight hidden layers, namely, two dense layers, convolutional layer, skip layer, mask layer, Bi-LSTM layer, attention layer, and pooling layer. SCBAMM makes better use of spatiotemporal information and captures emotion-related features more effectively. In addition, it solves the problems of gradient exploding and gradient vanishing in deep learning to some extent. On the databases EMO-DB and CASIA, the proposed model SCBAMM achieves an accuracy rate of 94.58% and 72.50%, respectively. As far as we know, compared with peer models, this is the best accuracy rate.

Recently, data-driven decision-making has attracted great interest; this requires high-quality datasets. However, real-world datasets often feature missing values for unknown or intentional reasons, rendering data-driven decision-making inaccurate. If a machine learning model is trained using incomplete datasets with missing values, the inferred results may be biased. In this case, a commonly used technique is the missing value imputation (MVI), which fills missing data with possible values estimated based on observed values. Various data imputation methods using machine learning, statistical inference, and relational database theories have been developed. Among them, conventional machine learning based imputation methods that handle tabular data can deal with only numerical columns or are time-consuming and cumbersome because they create an individualized predictive model for each column. Therefore, we have developed a novel imputational neural network that we term the Denoising Self-Attention Network (DSAN). Our proposed DSAN can deal with tabular datasets containing both numerical and categorical columns; it considers discretized numerical values as categorical values for embedding and self-attention layers. Furthermore, the DSAN learns robust feature expression vectors by combining self-attention and denoising techniques, and can predict multiple, appropriate substituted values simultaneously (via multi-task learning). To verify the validity of the method, we performed data imputation experiments after arbitrarily generating missing values for several real-world tabular datasets. We evaluated both imputational and downstream task performances, and we have seen that the DSAN outperformed the other models, especially in terms of category variable imputation.

Anomaly detection in video is a challenging computer vision problem. Due to the lack of anomalous events at training time, anomaly detection requires the design of learning methods without full supervision. In this paper, we approach anomalous event detection in video through self-supervised and multi-task learning at the object level. We first utilize a pre-trained detector to detect objects. Then, we train a 3D convolutional neural network to produce discriminative anomaly-specific information by jointly learning multiple proxy tasks: three self-supervised and one based on knowledge distillation. The self-supervised tasks are: (i) discrimination of forward/backward moving objects (arrow of time), (ii) discrimination of objects in consecutive/intermittent frames (motion irregularity) and (iii) reconstruction of object-specific appearance information. The knowledge distillation task takes into account both classification and detection information, generating large prediction discrepancies between teacher and student models when anomalies occur. To the best of our knowledge, we are the first to approach anomalous event detection in video as a multi-task learning problem, integrating multiple self-supervised and knowledge distillation proxy tasks in a single architecture. Our lightweight architecture outperforms the state-of-the-art methods on three benchmarks: Avenue, ShanghaiTech and UCSD Ped2. Additionally, we perform an ablation study demonstrating the importance of integrating self-supervised learning and normality-specific distillation in a multi-task learning setting.

We propose a multi-task deep-learning approach for estimating the check-worthiness of claims in political debates. Given a political debate, such as the 2016 US Presidential and Vice-Presidential ones, the task is to predict which statements in the debate should be prioritized for fact-checking. While different fact-checking organizations would naturally make different choices when analyzing the same debate, we show that it pays to learn from multiple sources simultaneously (PolitiFact, FactCheck, ABC, CNN, NPR, NYT, Chicago Tribune, The Guardian, and Washington Post) in a multi-task learning setup, even when a particular source is chosen as a target to imitate. Our evaluation shows state-of-the-art results on a standard dataset for the task of check-worthiness prediction.

Entity recommendation, providing search users with an improved experience by assisting them in finding related entities for a given query, has become an indispensable feature of today's Web search engine. Existing studies typically only consider the query issued at the current time step while ignoring the in-session preceding queries. Thus, they typically fail to handle the ambiguous queries such as "apple" because the model could not understand which apple (company or fruit) is talked about. In this work, we believe that the in-session contexts convey valuable evidences that could facilitate the semantic modeling of queries, and take that into consideration for entity recommendation. Furthermore, in order to better model the semantics of queries, we learn the model in a multi-task learning setting where the query representation is shared across entity recommendation and context-aware ranking. We evaluate our approach using large-scale, real-world search logs of a widely used commercial Web search engine. The experimental results show that incorporating context information significantly improves entity recommendation, and learning the model in a multi-task learning setting could bring further improvements.

With abundant labeled data, deep convolutional neural networks have shown great success in various image recognition tasks. However, these models are often less powerful when applied to novel datasets due to a phenomenon known as domain shift. Unsupervised domain adaptation methods aim to address this problem, allowing deep models trained on the labeled source domain to be used on a different target domain (without labels). In this presentation, we investigate whether the generalization ability of an unsupervised domain adaptation method can be improved through multi-task learning, with learned features required to be both domain invariant and discriminative for multiple different but relevant tasks. Experiments evaluating two fundamental recognition tasks---image recognition and segmentation---show that the generalization ability empowered by multi-task learning may not benefit recognition when the model is directly applied on the target domain, but the multi-task learning setting can boost the performance of state-of-the-art unsupervised domain adaptation methods by a non-negligible margin.

We present the Cross-stitched Multi-task Unified Dual Recursive Network (CMUDRN) model targeting the task of unified deraining and desnowing in a multi-task learning setting. This unified model borrows from the basic Dual Recursive Network (DRN) architecture developed by Cai et al. The proposed model makes use of cross-stitch units that enable multi-task learning across two separate DRN models, each tasked for single image deraining and desnowing, respectively. By fixing cross-stitch units at several layers of basic task-specific DRN networks, we perform multi - task learning over the two separate DRN models. To enable blind image restoration, on top of these structures we employ a simple neural fusion scheme which merges the output of each DRN. The separate task-specific DRN models and the fusion scheme are simultaneously trained by enforcing local and global supervision. Local supervision is applied on the two DRN submodules, and global supervision is applied on the data fusion submodule of the proposed model. Consequently, we both enable feature sharing across task-specific DRN models and control the image restoration behavior of the DRN submodules. An ablation study shows the strength of the hypothesized CMUDRN model, and experiments indicate that its performance is comparable or better than baseline DRN models on the single image deraining and desnowing tasks. Moreover, CMUDRN enables blind image restoration for the two underlying image restoration tasks, by unifying task-specific image restoration pipelines via a naive parametric fusion scheme. The CMUDRN implementation is available at https://github.comNCL3D/CMUDRN.

Artificial intelligence (AI) has advanced rapidly, but it has limited impact on biomedical text understanding due to a lack of annotated datasets (a.k.a. few-shot learning). Multi-task learning, which uses data from multiple datasets and tasks with related syntax and semantics, has potential to address this issue. However, the effectiveness of this approach heavily relies on the quality of the available data and its transferability between tasks. In this paper, we propose a framework, built upon a state-of-the-art multi-task method (i.e. MT-DNN), that leverages different publicly available biomedical datasets to enhance relation extraction performance. Our model employs a transformer-based architecture with shared encoding layers across multiple tasks, and task-specific classification layers to generate task-specific representations. To further improve performance, we utilize a knowledge distillation technique. In our experiments, we assess the impact of incorporating biomedical datasets in a multi-task learning setting and demonstrate that it consistently outperforms state-of-the-art few-shot learning methods in cases of limited data. This results in significant improvement across most datasets and few-shot scenarios, particularly in terms of recall scores.