As autonomous vehicles (AVs) need to interact with other road users, it is of importance to comprehensively understand the dynamic traffic environment, especially the future possible trajectories of surrounding vehicles. This paper presents an algorithm for long-horizon trajectory prediction of surrounding vehicles using a dual long short term memory (LSTM) network, which is capable of effectively improving prediction accuracy in strongly interactive driving environments. In contrast to traditional approaches which require trajectory matching and manual feature selection, this method can automatically learn high-level spatial-temporal features of driver behaviors from naturalistic driving data through sequence learning. By employing two blocks of LSTMs, the proposed method feeds the sequential trajectory to the first LSTM for driver intention recognition as an intermediate indicator, which is immediately followed by a second LSTM for future trajectory prediction. Test results from real-world highway driving data show that the proposed method can, in comparison to state-of-art methods, output more accurate and reasonable estimate of different future trajectories over 5s time horizon with root mean square error (RMSE) for longitudinal and lateral prediction less than 5.77m and 0.49m, respectively.

Face occlusions, covering either the majority or discriminative parts of the face, can break facial perception and produce a drastic loss of information. Biometric systems such as recent deep face recognition models are not immune to obstructions or other objects covering parts of the face. While most of the current face recognition methods are not optimized to handle occlusions, there have been a few attempts to improve robustness directly in the training stage. Unlike those, we propose to study the effect of generative face completion on the recognition. We offer a face completion encoder-decoder, based on a convolutional operator with a gating mechanism, trained with an ample set of face occlusions. To systematically evaluate the impact of realistic occlusions on recognition, we propose to play the occlusion game: we render 3D objects onto different face parts, providing precious knowledge of what the impact is of effectively removing those occlusions. Extensive experiments on the Labeled Faces in the Wild (LFW), and its more difficult variant LFW-BLUFR, testify that face completion is able to partially restore face perception in machine vision systems for improved recognition.

We introduce a pooling method for sets of feature vectors based on sorting features across elements of the set. This allows a deep neural network for sets to learn more flexible representations. We also demonstrate how FSPool can be used to construct a permutation-equivariant auto-encoder. On a toy dataset of polygons and a set version of MNIST, we show that such an auto-encoder produces considerably better reconstructions. Used in set classification, FSPool significantly improves accuracy and convergence speed on the set versions of MNIST and CLEVR.

Although neural conversation models are effective in learning how to produce fluent responses, their primary challenge lies in knowing what to say to make the conversation contentful and non-vacuous. We present a new end-to-end approach to contentful neural conversation that jointly models response generation and on-demand machine reading. The key idea is to provide the conversation model with relevant long-form text on the fly as a source of external knowledge. The model performs QA-style reading comprehension on this text in response to each conversational turn, thereby allowing for more focused integration of external knowledge than has been possible in prior approaches. To support further research on knowledge-grounded conversation, we introduce a new large-scale conversation dataset grounded in external web pages (2.8M turns, 7.4M sentences of grounding). Both human evaluation and automated metrics show that our approach results in more contentful responses compared to a variety of previous methods, improving both the informativeness and diversity of generated output.

An attribution method was developed to interpret a recurrent neural network (RNN) trained to predict a child's risk of ICU mortality using multi-modal, time series data in the Electronic Medical Records. By learning a sparse, binary mask that highlights salient features of the input data, critical features determining an individual patient's severity of illness could be identified. The method, called Learned Binary Masks (LBM), demonstrated that the RNN used different feature sets specific to each patient's illness; and further, the features highlighted aligned with clinical intuition of the patient's disease trajectories. LBM was also used to identify the most salient features across the model, analogous to "feature importance" computed in the Random Forest. This measure of the RNN's feature importance was further used to select the 25% most used features for training a second RNN model. Interestingly, but not surprisingly, the second model maintained similar performance to the model trained on all features. LBM is data-agnostic and can be used to interpret the predictions of any differentiable model.

Deep approaches to anomaly detection have recently shown promising results over shallow approaches on high-dimensional data. Typically anomaly detection is treated as an unsupervised learning problem. In practice however, one may have---in addition to a large set of unlabeled samples---access to a small pool of labeled samples, e.g. a subset verified by some domain expert as being normal or anomalous. Semi-supervised approaches to anomaly detection make use of such labeled data to improve detection performance. Few deep semi-supervised approaches to anomaly detection have been proposed so far and those that exist are domain-specific. In this work, we present Deep SAD, an end-to-end methodology for deep semi-supervised anomaly detection. Using an information-theoretic perspective on anomaly detection, we derive a loss motivated by the idea that the entropy for the latent distribution of normal data should be lower than the entropy of the anomalous distribution. We demonstrate in extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 along with other anomaly detection benchmark datasets that our approach is on par or outperforms shallow, hybrid, and deep competitors, even when provided with only few labeled training data.

Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive {\em uncertainty}. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model's output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian-and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous large-scale empirical comparison of these methods under dataset shift. We present a large-scale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks.

Bayesian methods promise to fix many shortcomings of deep learning, but they are impractical and rarely match the performance of standard methods, let alone improve them. In this paper, we demonstrate practical training of deep networks with natural-gradient variational inference. By applying techniques such as batch normalisation, data augmentation, and distributed training, we achieve similar performance in about the same number of epochs as the Adam optimiser, even on large datasets such as ImageNet. Importantly, the benefits of Bayesian principles are preserved: predictive probabilities are well-calibrated and uncertainties on out-of-distribution data are improved. This work enables practical deep learning while preserving benefits of Bayesian principles. A PyTorch implementation will be available as a plug-and-play optimiser.

The state-of-the-art solutions for Aspect-Level Sentiment Analysis (ALSA) are built on a variety of deep neural networks (DNN), whose efficacy depends on large amounts of accurately labeled training data. Unfortunately, high-quality labeled training data usually require expensive manual work, and are thus not readily available in many real scenarios. In this paper, we aim to enable effective machine labeling for ALSA without the requirement for manual labeling effort. Towards this aim, we present a novel solution based on the recently proposed paradigm of gradual machine learning. It begins with some easy instances in an ALSA task, which can be automatically labeled by the machine with high accuracy, and then gradually labels the more challenging instances by iterative factor graph inference. In the process of gradual machine learning, the hard instances are gradually labeled in small stages based on the estimated evidential certainty provided by the labeled easier instances. Our extensive experiments on the benchmark datasets have shown that the performance of the proposed approach is considerably better than its unsupervised alternatives, and also highly competitive compared to the state-of-the-art supervised DNN techniques.

The vulnerability to slight input perturbations is a worrying yet intriguing property of deep neural networks (DNNs). Despite many previous works studying the reason behind such adversarial behavior, the relationship between the generalization performance and adversarial behavior of DNNs is still little understood. In this work, we reveal such relation by introducing a metric characterizing the generalization performance of a DNN. The metric can be disentangled into an information-theoretic non-robust component, responsible for adversarial behavior, and a robust component. Then, we show by experiments that current DNNs rely heavily on optimizing the non-robust component in achieving decent performance. We also demonstrate that current state-of-the-art adversarial training algorithms indeed try to robustify the DNNs by preventing them from using the non-robust component to distinguish samples from different categories. Also, based on our findings, we take a step forward and point out the possible direction for achieving decent standard performance and adversarial robustness simultaneously. We believe that our theory could further inspire the community to make more interesting discoveries about the relationship between standard generalization and adversarial generalization of deep learning models.