Unsupervised Story Comprehension with Hierarchical Encoder-Decoder

Image super resolution is one of the most popular topics in the field of image processing. However, most of the existing super resolution algorithms are designed for the situation where sufficient training data is available. This paper proposes a new image super resolution approach that is able to handle the situation with sparse training data, using the recently developed ANFIS (Adaptive Network based Fuzzy Inference System) interpolation technique. In particular, the training image data set is divided into different subsets. For subsets with sufficient training data, the ANFIS models are trained using standard ANFIS learning procedure, while for those with insufficient data, the models are obtained through ANFIS interpolation. In the literature, little work exists for image super resolution on sparse data. Therefore, in the experimental evaluations of this paper, the proposed approach is compared with existing super resolution methods with full data, demonstrating that this work is able to produce highly promising results.

Machine learning (ML) techniques are increasingly applied to decision-making and control problems in Cyber-Physical Systems among which many are safety-critical, e.g., chemical plants, robotics, autonomous vehicles. Despite the significant benefits brought by ML techniques, they also raise additional safety issues because 1) most expressive and powerful ML models are not transparent and behave as a black box and 2) the training data which plays a crucial role in ML safety is usually incomplete. An important technique to achieve safety for ML models is "Safe Fail", i.e., a model selects a reject option and applies the backup solution, a traditional controller or a human operator for example, when it has low confidence in a prediction. Data-driven models produced by ML algorithms learn from training data, and hence they are only as good as the examples they have learnt. As pointed in [17], ML models work well in the "training space" (i.e., feature space with sufficient training data), but they could not extrapolate beyond the training space. As observed in many previous studies, a feature space that lacks training data generally has a much higher error rate than the one that contains sufficient training samples [31]. Therefore, it is essential to identify the training space and avoid extrapolating beyond the training space. In this paper, we propose an efficient Feature Space Partitioning Tree (FSPT) to address this problem. Using experiments, we also show that, a strong relationship exists between model performance and FSPT score.

In traditional transductive learning, all queries are used in learning to rank in order to generate pseudo-labels when sufficient training data are not available. However, low quality queries may affect retrieval performance in transductive learning. We thus think that it is important to improve the quality of queries in transductive learning to train an effective ranking model. By using a small number of reliable samples and data close to the boundaries of classification, we propose building a query quality estimator by establishing a relationship between the benefits of good retrieval performance and features of the normalized query commitment that influence query quality. In our proposed transduction model, all queries available are filtered by the proposed query quality estimator and only high quality queries that enhance the effectiveness of retrieval such that they yield performance-related benefits, are used to generate pseudo-labels for learning to rank. Queries that can degrade performance benefits are discarded while creating the pseudo-labels. Pseudo-labels aggregated by high quality queries in transductive learning are then leveraged in learning to rank scenarios without sufficient training data. The results of extensive experiments on the standard LETOR 4.0 dataset showed that our proposed method can outperform strong baselines and the average normalized discounted cumulative gain is enhanced up to 7.77% in some case.

SUMMARY Precise real time estimates of earthquake magnitude and location are essential for early warning and rapid response. While recently multiple deep learning approaches for fast assessment of earthquakes have been proposed, they usually rely on either seismic records from a single station or from a fixed set of seismic stations. Here we introduce a new model for real-time magnitude and location estimation using the attention based transformer networks. Our approach incorporates waveforms from a dynamically varying set of stations and outperforms deep learning baselines in both magnitude and location estimation performance. Furthermore, it outperforms a classical magnitude estimation algorithm considerably and shows promising performance in comparison to a classical localization algorithm. Our model is applicable to real-time prediction and provides realistic uncertainty estimates based on probabilistic inference. In this work, we furthermore conduct a comprehensive study of the requirements on training data, the training procedures and the typical failure modes. Using three diverse and large scale data sets, we conduct targeted experiments and a qualitative error analysis. Our analysis gives several key insights. First, we can precisely pinpoint the effect of large training data; for example, a four times larger training set reduces average errors for both magnitude and location prediction by more than half, and reduces the required time for real time assessment by a factor of four. Secondly, the basic model systematically underestimates large magnitude events. This issue can be mitigated, and in some cases completely resolved, by incorporating events from other regions into the training through transfer learning. Thirdly, location estimation is highly precise in areas with sufficient training data, but is strongly degraded for events outside the training distribution, sometimes producing massive outliers. Our analysis suggests that these characteristics are not only present for our model, but for most deep learning models for fast assessment published so far. They result from the black box modeling and their mitigation will likely require imposing physics derived constraints on the neural network. These characteristics need to be taken into consideration for practical applications.

MACHINE LEARNING PREDICTION OF IN-HOSPITAL DISEASE PROGRESSION IN COMMUNITY-ACQUIRED PNEUMONIA: DERIVATION AND VALIDATION OF CLINICAL DETERIORATION IN PNEUMONIA (CDIP)

Traditional machine learning (ML) techniques are often employed to perform complex pattern recognition tasks for remote sensing images, such as land-use classification. In order to obtain acceptable classification results, these techniques require there to be sufficient training data available for every particular image. Obtaining training samples is challenging, particularly for near real-time applications. Therefore, past knowledge must be utilized to overcome the lack of training data in the current regime. This challenge is known as domain adaptation (DA), and one of the common approaches to this problem is based on finding invariant representations for both the training and test data, which are often assumed to come from different “domains.” In this study, we consider two deep learning techniques for learning domain-invariant representations: Denoising autoencoders (DAE) and domain-adversarial neural networks (DANN). While the DAE is a typical two-stage DA technique (unsupervised invariant representation learning followed by supervised classification), DANN is an end-to-end approach where invariant representation learning and classification are considered jointly during training. The proposed techniques are applied to both hyperspectral and multispectral images under different DA scenarios. Results obtained show that the proposed techniques outperform traditional approaches, such as principal component analysis (PCA) and kernel PCA, and can also compete with a fully supervised model in the multispatial scenario.

In recent years, Convolutional Neural Networks (CNNs) have been widely successful for numerous computer vision tasks using 2D color and 3D point data, such as object detection and classification, and 3D pose estimation. Despite these successes, it comes with a price --- CNNs require a vast amount of labeled training data to be successful. Many training datasets are now available, however, they are mostly limited to general objects in the public domain such as people, cars, buildings, fruit, and more. Thus, not all fields have sufficient training data. Even if an adequate number of data samples can be built into a dataset, manually collecting and labeling the training data remains a laborious task. Several approaches to reduce or eradicate the need for training data have been proposed by researchers. The two most popular approaches are training with synthetic data and training on generalized object features. Training with synthetic data refers to the use of computer graphics to generate training data. While this approach has demonstrated success, it comes with its own challenges. In contrast, training a CNN on generalized object features identifies specific local 3D features or control points on objects. Local 3D features and their descriptors have proven in the past to be universally applicable and effective for general object detection, and thus, this method allows for a generalized CNN. A generalized CNN refers to a network architecture that, after initial training, can be specialized for a new task with a minimum amount or no new training data. This dissertation investigates an encoder-decoder architecture and determines the capabilities of it to generalize the CNN architecture for object detection tasks in point cloud data. A CNN is trained to recognize 3D features and demonstrates, without retraining, its successful applicability for point cloud object detection using disparate data. Additionally, the geometric consistency of the descriptors is evaluated to obtain further insight. The results demonstrate that the proposed 3D feature descriptors, and their increased geometric consistency contribute to the increased encoder-decoder architecture performance. In summary, this dissertation contributes a state-of-the-art, generalized, feature descriptor-based CNN architecture that can be transferred to different objects without retraining. Furthermore, it provides insight explaining the increased performance of the CNN compared to the state-of-the-art.

SummaryAccurate prediction of the solubility of chemical substances in solvents remains a challenge. The sparsity of high-quality solubility data is recognized as the biggest hurdle in the development of robust data-driven methods for practical use. Nonetheless, the effects of the quality and quantity of data on aqueous solubility predictions have not yet been scrutinized. In this study, the roles of the size and the quality of data sets on the performances of the solubility prediction models are unraveled, and the concepts of actual and observed performances are introduced. In an effort to curtail the gap between actual and observed performances, a quality-oriented data selection method, which evaluates the quality of data and extracts the most accurate part of it through statistical validation, is designed. Applying this method on the largest publicly available solubility database and using a consensus machine learning approach, a top-performing solubility prediction model is achieved.

Deep learning, through the use of neural networks, has demonstrated remarkable ability to automate many routine tasks when presented with sufficient data for training. The neural network architecture (e.g. number of layers, types of layers, connections between layers, etc.) plays a critical role in determining what, if anything, the neural network is able to learn from the training data. The trend for neural network architectures, especially those trained on ImageNet, has been to grow ever deeper and more complex. The result has been ever increasing accuracy on benchmark datasets with the cost of increased computational demands. In this paper we demonstrate that neural network architectures can be automatically generated, tailored for a specific application, with dual objectives: accuracy of prediction and speed of prediction. Using MENNDL--an HPC-enabled software stack for neural architecture search--we generate a neural network with comparable accuracy to state-of-the-art networks on a cancer pathology dataset that is also $16\times$ faster at inference. The speedup in inference is necessary because of the volume and velocity of cancer pathology data; specifically, the previous state-of-the-art networks are too slow for individual researchers without access to HPC systems to keep pace with the rate of data generation. Our new model enables researchers with modest computational resources to analyze newly generated data faster than it is collected.

Generative adversarial network for road damage detection

With the increasing presence of NLU tools such as automatic dialogue systems, achieving high accuracy of intent recognition in chatbots becomes an especially important problem to tackle. This issue cannot be solved without sufficient training data, but the scarcity of labelled training data often poses a major challenge to the development of real-life chatbots. Therefore, methods utilizing unlabelled data resources have been recently gaining interest. One of most notable approaches is the use of pre-trained encoders based on language models. Trained for general purposes, they benefit from further domain adjustments. In our work we offer an approach that can increase the model’s accuracy for text classification, which can serve as an alternative for standard methods of domain adaptation. Our approach consists of a combination of methods: a clustering approach, similar to intent induction; an encoder domain adaptation on a cluster classification task, similar to intent recognition using unlabelled data; and model fine-tuning on labelled datasets. In this approach unlabelled data becomes complementary to labelled data, reducing the time needed for corpus building. We evaluate our approach on: 1) the public WebApp dataset and 2) a demanding real-life banking domain dataset, achieving 0.97 and 0.93 accuracy respectively. This approach, called Conversation Clustering Adaptation (CCA), when applied to an encoder, increases the accuracy of intent recognition up by to 12.4pp and exceeds current state-of-the-art methods while benefiting from the use of additional training data. We share our code at https://github.com/michal-lew/cca.

In the current era of information technology, blockchain is widely used in various fields, and the monitoring of the security and status of the blockchain system is of great concern. Online anomaly detection for the real-time stream data plays vital role in monitoring strategy to find abnormal events and status of blockchain system. However, as the high requirements of real-time and online scenario, online anomaly detection faces many problems such as limited training data, distribution drift, and limited update frequency. In this paper, we propose an adaptive stream outlier detection method (ASOD) to overcome the limitations. It first designs a K-nearest neighbor Gaussian mixture model (KNN-GMM) and utilizes online learning strategy. So, it is suitable for online scenarios and does not rely on large training data. The K-nearest neighbor optimization limits the influence of new data locally rather than globally, thus improving the stability. Then, ASOD applies the mechanism of dynamic maintenance of Gaussian components and the strategy of dynamic context control to achieve self-adaptation to the distribution drift. And finally, ASOD adopts a dimensionless distance metric based on Mahalanobis distance and proposes an automatic threshold method to accomplish anomaly detection. In addition, the KNN-GMM provides the life cycle and the anomaly index for continuous tracking and analysis, which facilities the cause analysis and further interpretation and traceability. From the experimental results, it can be seen that ASOD achieves near-optimal F1 and recall on the NAB dataset with an improvement of 6% and 20.3% over the average, compared to baselines with sufficient training data. ASOD has the lowest F1 variance among the five best methods, indicating that it is effective and stable for online anomaly detection on stream data.

Sufficient training data normally is required to train deeply learned models. However, due to the expensive manual process for labelling large number of images (i.e., annotation), the amount of available training data (i.e., real data) is always limited. To produce more data for training a deep network, Generative Adversarial Network (GAN) can be used to generate artificial sample data (i.e., generated data). However, the generated data usually does not have annotation labels. To solve this problem, in this paper, we propose a virtual label called Multi-pseudo Regularized Label (MpRL) and assign it to the generated data. With MpRL, the generated data will be used as the supplementary of real training data to train a deep neural network in a semi-supervised learning fashion. To build the corresponding relationship between the real data and generated data, MpRL assigns each generated data a proper virtual label which reflects the likelihood of the affiliation of the generated data to predefined training classes in the real data domain. Unlike the traditional label which usually is a single integral number, the virtual label proposed in this work is a set of weight-based values each individual of which is a number in (0,1] called multi-pseudo label and reflects the degree of relation between each generated data to every pre-defined class of real data. A comprehensive evaluation is carried out by adopting two state-of-the-art convolutional neural networks (CNNs) in our experiments to verify the effectiveness of MpRL. Experiments demonstrate that by assigning MpRL to generated data, we can further improve the person re-ID performance on five re-ID datasets, i.e., Market-1501, DukeMTMC-reID, CUHK03, VIPeR, and CUHK01. The proposed method obtains +6.29%, +6.30%, +5.58%, +5.84%, and +3.48% improvements in rank-1 accuracy over a strong CNN baseline on the five datasets respectively, and outperforms state-of-the-art methods.

State-of-the-art decision tree methods apply heuristics recursively to create each split in isolation, which may not capture well the underlying characteristics of the dataset. The optimal decision tree problem attempts to resolve this by creating the entire decision tree at once to achieve global optimality. In the last 25?years, algorithmic advances in integer optimization coupled with hardware improvements have resulted in an astonishing 800 billion factor speedup in mixed-integer optimization (MIO). Motivated by this speedup, we present optimal classification trees, a novel formulation of the decision tree problem using modern MIO techniques that yields the optimal decision tree for axes-aligned splits. We also show the richness of this MIO formulation by adapting it to give optimal classification trees with hyperplanes that generates optimal decision trees with multivariate splits. Synthetic tests demonstrate that these methods recover the true decision tree more closely than heuristics, refuting the notion that optimal methods overfit the training data. We comprehensively benchmark these methods on a sample of 53 datasets from the UCI machine learning repository. We establish that these MIO methods are practically solvable on real-world datasets with sizes in the 1000s, and give average absolute improvements in out-of-sample accuracy over CART of 1---2 and 3---5% for the univariate and multivariate cases, respectively. Furthermore, we identify that optimal classification trees are likely to outperform CART by 1.2---1.3% in situations where the CART accuracy is high and we have sufficient training data, while the multivariate version outperforms CART by 4---7% when the CART accuracy or dimension of the dataset is low.

Theory guided Lagrange programming neural network for subsurface flow problems