Attention based models have become the new state-of-the-art in natural language understanding tasks such as question-answering and sentence similarity. Recent models, such as BERT and XLNet, score a pair of sentences (A and B) using multiple cross-attention operations - a process in which each word in sentence A attends to all words in sentence B and vice versa. As a result, computing the similarity between a query sentence and a set of candidate sentences, requires the propagation of all query-candidate sentence-pairs throughout a stack of cross-attention layers. This exhaustive process becomes computationally prohibitive when the number of candidate sentences is large. In contrast, sentence embedding techniques learn a sentence-to-vector mapping and compute the similarity between the sentence vectors via simple elementary operations such as dot product or cosine similarity. In this paper, we introduce a sentence embedding method that is based on knowledge distillation from cross-attentive models, focusing on sentence-pair tasks. The outline of the proposed method is as follows: Given a cross-attentive teacher model (e.g. a fine-tuned BERT), we train a sentence embedding based student model to reconstruct the sentence-pair scores obtained by the teacher model. We empirically demonstrate the effectiveness of our distillation method on five GLUE sentence-pair tasks. Our method significantly outperforms several ELMO variants and other sentence embedding methods, while accelerating computation of the query-candidate sentence-pairs similarities by several orders of magnitude, with an average relative degradation of 4.6% compared to BERT.

The initial analysis of any large data set can be divided into two phases: (1) the identification of common trends or patterns and (2) the identification of anomalies or outliers that deviate from those trends. We focus on the goal of detecting observations with novel content, which can alert us to artifacts in the data set or, potentially, the discovery of previously unknown phenomena. To aid in interpreting and diagnosing the novel aspect of these selected observations, we recommend the use of novelty detection methods that generate explanations. In the context of large image data sets, these explanations should highlight what aspect of a given image is new (color, shape, texture, content) in a human-comprehensible form. We propose DEMUD-VIS, the first method for providing visual explanations of novel image content by employing a convolutional neural network (CNN) to extract image features, a method that uses reconstruction error to detect novel content, and an up-convolutional network to convert CNN feature representations back into image space. We demonstrate this approach on diverse images from ImageNet, freshwater streams, and the surface of Mars.

This is especially prevalent in machine learning (Schulman et al., 2015), including variational inference (Ranganath et al., 2014; Rezende et al., 2014) and reinforcement learning (Silver et al., 2014). On the face of it, problems of this nature require solving an intractable integral. Most practical approaches instead use Monte Carlo estimates of expectations and their gradients. These techniques are unbiased but can suffer from high variance when sample size is small--one unlikely sample in the tail of a distribution can heavily skew the final estimate. A simple way to reduce variance is to increase the number of samples; however the computational cost grows quickly.

This paper is about a machine learning approach based on the multilinear projection of an unknown function (or probability distribution) to be estimated towards a linear (or multilinear) dimensional space E'. The proposal transforms the problem of predicting the target of an observation x into a problem of determining a consensus among the k nearest neighbors of x's image within the dimensional space E'. The algorithms that concretize it allow both regression and binary classification. Implementations carried out using Scala/Spark and assessed on a dozen LIBSVM datasets have demonstrated improvements in prediction accuracies in comparison with other prediction algorithms implemented within Spark MLLib such as multilayer perceptrons, logistic regression classifiers and random forests.

In this paper we introduce the ice-start problem, i.e., the challenge of deploying machine learning models when only little or no training data is initially available, and acquiring each feature element of data is associated with costs. This setting is representative for the real-world machine learning applications. For instance, in the health-care domain, when training an AI system for predicting patient metrics from lab tests, obtaining every single measurement comes with a high cost. Active learning, where only the label is associated with a cost does not apply to such problem, because performing all possible lab tests to acquire a new training datum would be costly, as well as unnecessary due to redundancy. We propose Icebreaker, a principled framework to approach the ice-start problem. Icebreaker uses a full Bayesian Deep Latent Gaussian Model (BELGAM) with a novel inference method. Our proposed method combines recent advances in amortized inference and stochastic gradient MCMC to enable fast and accurate posterior inference. By utilizing BELGAM's ability to fully quantify model uncertainty, we also propose two information acquisition functions for imputation and active prediction problems. We demonstrate that BELGAM performs significantly better than the previous VAE (Variational autoencoder) based models, when the data set size is small, using both machine learning benchmarks and real-world recommender systems and health-care applications. Moreover, based on BELGAM, Icebreaker further improves the performance and demonstrate the ability to use minimum amount of the training data to obtain the highest test time performance.

Natural language understanding (NLU) of text is a fundamental challenge in AI, and it has received significant attention throughout the history of NLP research. This primary goal has been studied under different tasks, such as Question Answering (QA) and Textual Entailment (TE). In this thesis, we investigate the NLU problem through the QA task and focus on the aspects that make it a challenge for the current state-of-the-art technology. This thesis is organized into three main parts: In the first part, we explore multiple formalisms to improve existing machine comprehension systems. We propose a formulation for abductive reasoning in natural language and show its effectiveness, especially in domains with limited training data. Additionally, to help reasoning systems cope with irrelevant or redundant information, we create a supervised approach to learn and detect the essential terms in questions. In the second part, we propose two new challenge datasets. In particular, we create two datasets of natural language questions where (i) the first one requires reasoning over multiple sentences; (ii) the second one requires temporal common sense reasoning. We hope that the two proposed datasets will motivate the field to address more complex problems. In the final part, we present the first formal framework for multi-step reasoning algorithms, in the presence of a few important properties of language use, such as incompleteness, ambiguity, etc. We apply this framework to prove fundamental limitations for reasoning algorithms. These theoretical results provide extra intuition into the existing empirical evidence in the field.

MEx: Multi-modal Exercises Dataset is a multi-sensor, multi-modal dataset, implemented to benchmark Human Activity Recognition(HAR) and Multi-modal Fusion algorithms. Collection of this dataset was inspired by the need for recognising and evaluating quality of exercise performance to support patients with Musculoskeletal Disorders(MSD). We select 7 exercises regularly recommended for MSD patients by physiotherapists and collected data with four sensors a pressure mat, a depth camera and two accelerometers. The dataset contains three data modalities; numerical time-series data, video data and pressure sensor data posing interesting research challenges when reasoning for HAR and Exercise Quality Assessment. This paper presents our evaluation of the dataset on number of standard classification algorithms for the HAR task by comparing different feature representation algorithms for each sensor. These results set a reference performance for each individual sensor that expose their strengths and weaknesses for the future tasks. In addition we visualise pressure mat data to explore the potential of the sensor to capture exercise performance quality. With the recent advancement in multi-modal fusion, we also believe MEx is a suitable dataset to benchmark not only HAR algorithms, but also fusion algorithms of heterogeneous data types in multiple application domains.

This short report describes our submission to the ISIC 2018 Challenge in Skin Lesion Analysis Towards Melanoma Detection for Task1 and Task 3. This work has been accomplished by a team of researchers at the University of Dayton Signal and Image Processing Lab. Our proposed approach is computationally efficient are combines information from both deep learning and handcrafted features. For Task3, we form a new type of image features, called hybrid features, which has stronger discrimination ability than single method features. These features are utilized as inputs to a decision-making model that is based on a multiclass Support Vector Machine (SVM) classifier. The proposed technique is evaluated on online validation databases. Our score was 0.841 with SVM classifier on the validation dataset.

Locality preserving projections (LPP) are a classical dimensionality reduction method based on data graph information. However, LPP is still responsive to extreme outliers. LPP aiming for vectorial data may undermine data structural information when it is applied to multidimensional data. Besides, it assumes the dimension of data to be smaller than the number of instances, which is not suitable for high-dimensional data. For high-dimensional data analysis, the tensor-train decomposition is proved to be able to efficiently and effectively capture the spatial relations. Thus, we propose a tensor-train parameterization for ultra dimensionality reduction (TTPUDR) in which the traditional LPP mapping is tensorized in terms of tensor-trains and the LPP objective is replaced with the Frobenius norm to increase the robustness of the model. The manifold optimization technique is utilized to solve the new model. The performance of TTPUDR is assessed on classification problems and TTPUDR significantly outperforms the past methods and the several state-of-the-art methods.

In this work we develop a distributed least squares approximation (DLSA) method, which is able to solve a large family of regression problems (e.g., linear regression, logistic regression, Cox's model) on a distributed system. By approximating the local objective function using a local quadratic form, we are able to obtain a combined estimator by taking a weighted average of local estimators. The resulting estimator is proved to be statistically as efficient as the global estimator. In the meanwhile it requires only one round of communication. We further conduct the shrinkage estimation based on the DLSA estimation by using an adaptive Lasso approach. The solution can be easily obtained by using the LARS algorithm on the master node. It is theoretically shown that the resulting estimator enjoys the oracle property and is selection consistent by using a newly designed distributed Bayesian Information Criterion (DBIC). The finite sample performance as well as the computational efficiency are further illustrated by extensive numerical study and an airline dataset. The airline dataset is 52GB in memory size. The entire methodology has been implemented by Python for a de-facto standard Spark system. By using the proposed DLSA algorithm on the Spark system, it takes 26 minutes to obtain a logistic regression estimator whereas a full likelihood algorithm takes 15 hours to reaches an inferior result.