Pretrained language models such as BERT, GPT have shown great effectiveness in language understanding. The auxiliary predictive tasks in existing pretraining approaches are mostly defined on tokens, thus may not be able to capture sentence-level semantics very well. To address this issue, we propose CERT: Contrastive self-supervised Encoder Representations from Transformers, which pretrains language representation models using contrastive self-supervised learning at the sentence level. CERT creates augmentations of original sentences using back-translation. Then it finetunes a pretrained language encoder (e.g., BERT) by predicting whether two augmented sentences originate from the same sentence. CERT is simple to use and can be flexibly plugged into any pretraining-finetuning NLP pipeline. We evaluate CERT on 11 natural language understanding tasks in the GLUE benchmark where CERT outperforms BERT on 7 tasks, achieves the same performance as BERT on 2 tasks, and performs worse than BERT on 2 tasks. On the averaged score of the 11 tasks, CERT outperforms BERT. The data and code are available at https://github.com/UCSD-AI4H/CERT

State-of-the-art approaches to causal discovery usually assume a fixed underlying causal model. However, it is often the case that causal models vary across domains or subjects, due to possibly omitted factors that affect the quantitative causal effects. As a typical example, causal connectivity in the brain network has been reported to vary across individuals, with significant differences across groups of people, such as autistics and typical controls. In this paper, we develop a unified framework for causal discovery and mechanism-based group identification. In particular, we propose a specific and shared causal model (SSCM), which takes into account the variabilities of causal relations across individuals/groups and leverages their commonalities to achieve statistically reliable estimation.

Adversarial training is a useful approach to promote the learning of transferable representations across the source and target domains, which has been widely applied for domain adaptation (DA) tasks based on deep neural networks. Until very recently, existing adversarial domain adaptation (ADA) methods ignore the useful information from the label space, which is an important factor accountable for the complicated data distributions associated with different semantic classes. Especially, the inter-class semantic relationships have been rarely considered and discussed in the current work of transfer learning. In this paper, we propose a novel relationship-aware adversarial domain adaptation (RADA) algorithm, which first utilizes a single multi-class domain discriminator to enforce the learning of inter-class dependency structure during domain-adversarial training and then aligns this structure with the inter-class dependencies that are characterized from training the label predictor on the source domain. Specifically, we impose a regularization term to penalize the structure discrepancy between the inter-class dependencies respectively estimated from domain discriminator and label predictor. Through this alignment, our proposed method makes the ADA aware of class relationships. Empirical studies show that the incorporation of class relationships significantly improves the performance on benchmark datasets.

Briefness and comprehensiveness are necessary in order to give a lot of information concisely in explaining a black-box decision system. However, existing interpretable machine learning methods fail to consider briefness and comprehensiveness simultaneously, which may lead to redundant explanations. We propose a system-agnostic interpretable method that provides a brief but comprehensive explanation by adopting the inspiring information theoretic principle, information bottleneck principle. Using an information theoretic objective, VIBI selects instance-wise key features that are maximally compressed about an input (briefness), and informative about a decision made by a black-box on that input (comprehensive). The selected key features act as an information bottleneck that serves as a concise explanation for each black-box decision. We show that VIBI outperforms other interpretable machine learning methods in terms of both interpretability and fidelity evaluated by human and quantitative metrics.

Generative Adversarial Nets (GANs) are emerging objects of study in machine learning, computer vision, natural language processing, and many other domains. In machine learning, study of such a framework has led to significant advances in adversarial defenses [28, 24] and machine security [4, 24]. In computer vision and natural language processing, GANs have resulted in improved performance over standard generative models for images and texts [13], such as variational autoencoder [16] and deep Boltzmann machine [22]. A main technique to achieve this goal is to play a minimax two-player game between generator and discriminator under the design that the generator tries to confuse the discriminator with its generated contents and the discriminator tries to distinguish real images/texts from what the generator creates. Despite a large amount of variants of GANs, many fundamental questions remain unresolved. One of the longstanding challenges is designing universal, easy-to-implement architectures that alleviate the instability issue of GANs training. Ideally, GANs are supposed to solve the minimax optimization problem [13], but in practice alternating gradient descent methods do not clearly privilege minimax over maximin or vice versa (page 35, [12]), which may lead to instability in training if there exists a large discrepancy between the minimax and maximin objective values. The focus of this work is on improving the stability of such minimax game in the training process of GANs. 1 Under review as a conference paper at ICLR 2019

This study presents a multimodal machine learning model to predict ICD-10 diagnostic codes. We developed separate machine learning models that can handle data from different modalities, including unstructured text, semi-structured text and structured tabular data. We further employed an ensemble method to integrate all modality-specific models to generate ICD-10 codes. Key evidence was also extracted to make our prediction more convincing and explainable. We used the Medical Information Mart for Intensive Care III (MIMIC -III) dataset to validate our approach. For ICD code prediction, our best-performing model (micro-F1 = 0.7633, micro-AUC = 0.9541) significantly outperforms other baseline models including TF-IDF (micro-F1 = 0.6721, micro-AUC = 0.7879) and Text-CNN model (micro-F1 = 0.6569, micro-AUC = 0.9235). For interpretability, our approach achieves a Jaccard Similarity Coefficient (JSC) of 0.1806 on text data and 0.3105 on tabular data, where well-trained physicians achieve 0.2780 and 0.5002 respectively.

Missing values widely exist in many real-world datasets, which hinders the performing of advanced data analytics. Properly filling these missing values is crucial but challenging, especially when the missing rate is high. Many approaches have been proposed for missing value imputation (MVI), but they are mostly heuristics-based, lacking a principled foundation and do not perform satisfactorily in practice. In this paper, we propose a probabilistic framework based on deep generative models for MVI. Under this framework, imputing the missing entries amounts to seeking a fixed-point solution between two conditional distributions defined on the missing entries and latent variables respectively. These distributions are parameterized by deep neural networks (DNNs) which possess high approximation power and can capture the nonlinear relationships between missing entries and the observed values. The learning of weight parameters of DNNs is performed by maximizing an approximation of the log-likelihood of observed values. We conducted extensive evaluation on 13 datasets and compared with 11 baselines methods, where our methods largely outperforms the baselines.

Distance metric learning (DML), which learns a distance metric from labeled "similar" and "dissimilar" data pairs, is widely utilized. Recently, several works investigate orthogonality-promoting regularization (OPR), which encourages the projection vectors in DML to be close to being orthogonal, to achieve three effects: (1) high balancedness -- achieving comparable performance on both frequent and infrequent classes; (2) high compactness -- using a small number of projection vectors to achieve a "good" metric; (3) good generalizability -- alleviating overfitting to training data. While showing promising results, these approaches suffer three problems. First, they involve solving non-convex optimization problems where achieving the global optimal is NP-hard. Second, it lacks a theoretical understanding why OPR can lead to balancedness. Third, the current generalization error analysis of OPR is not directly on the regularizer. In this paper, we address these three issues by (1) seeking convex relaxations of the original nonconvex problems so that the global optimal is guaranteed to be achievable; (2) providing a formal analysis on OPR's capability of promoting balancedness; (3) providing a theoretical analysis that directly reveals the relationship between OPR and generalization performance. Experiments on various datasets demonstrate that our convex methods are more effective in promoting balancedness, compactness, and generalization, and are computationally more efficient, compared with the nonconvex methods.

In representation learning (RL), how to make the learned representations easy to interpret and less overfitted to training data are two important but challenging issues. To address these problems, we study a new type of regulariza- tion approach that encourages the supports of weight vectors in RL models to have small overlap, by simultaneously promoting near-orthogonality among vectors and sparsity of each vector. We apply the proposed regularizer to two models: neural networks (NNs) and sparse coding (SC), and develop an efficient ADMM-based algorithm for regu- larized SC. Experiments on various datasets demonstrate that weight vectors learned under our regularizer are more interpretable and have better generalization performance.

Kernel methods are powerful tools to capture nonlinear patterns behind data. They implicitly learn high (even infinite) dimensional nonlinear features in the Reproducing Kernel Hilbert Space (RKHS) while making the computation tractable by leveraging the kernel trick. Classic kernel methods learn a single layer of nonlinear features, whose representational power may be limited. Motivated by recent success of deep neural networks (DNNs) that learn multi-layer hierarchical representations, we propose a Stacked Kernel Network (SKN) that learns a hierarchy of RKHS-based nonlinear features. SKN interleaves several layers of nonlinear transformations (from a linear space to a RKHS) and linear transformations (from a RKHS to a linear space). Similar to DNNs, a SKN is composed of multiple layers of hidden units, but each parameterized by a RKHS function rather than a finite-dimensional vector. We propose three ways to represent the RKHS functions in SKN: (1)nonparametric representation, (2)parametric representation and (3)random Fourier feature representation. Furthermore, we expand SKN into CNN architecture called Stacked Kernel Convolutional Network (SKCN). SKCN learning a hierarchy of RKHS-based nonlinear features by convolutional operation with each filter also parameterized by a RKHS function rather than a finite-dimensional matrix in CNN, which is suitable for image inputs. Experiments on various datasets demonstrate the effectiveness of SKN and SKCN, which outperform the competitive methods.