ADDITIONAL SHARED DECODER ON SIAMESE MUL TI-VIEW ENCODERS FOR LEARNING ACOUSTIC WORD EMBEDDINGS Myunghun Jung, Hyungjun Lim, Jahyun Goo, Y oungmoon Jung, and Hoirin Kim School of Electrical Engineering, KAIST, Daejeon, Republic of Korea ABSTRACT Acoustic word embeddings -- fixed-dimensional vector representations of arbitrary-length words -- have attracted increasing interest in query-by-example spoken term detection. Recently, on the fact that the orthography of text labels partly reflects the phonetic similarity between the words' pronunciation, a multi-view approach has been introduced that jointly learns acoustic and text embeddings. It showed that it is possible to learn discriminative embeddings by designing the objective which takes text labels as well as word segments. In this paper, we propose a network architecture that expands the multi-view approach by combining the Siamese multi-view encoders with a shared decoder network to maximize the effect of the relationship between acoustic and text em-beddings in embedding space. Discriminatively trained with multi-view triplet loss and decoding loss, our proposed approach achieves better performance on acoustic word discrimination task with the WSJ dataset, resulting in 11.1% relative improvement in average precision. Index T erms -- acoustic word embedding, query-by- example spoken term detection, multi-view learning, Siamese network, encoder-decoder 1. INTRODUCTION Query-by-example spoken term detection (QbE-STD) is the task of retrieving a spoken query from a set of speech utterances. Amazon Echo, Google Home, Apple Siri), the QbE-STD has drawn interest as a technique that can be applied to wake-up or command word detection, search engine, etc.

Exposure bias refers to the train-test discrepancy that seemingly arises when an autoregressive generative model uses only ground-truth contexts at training time but generated ones at test time. We separate the contributions of the model and the learning framework to clarify the debate on consequences and review proposed counter-measures. In this light, we argue that generalization is the underlying property to address and propose unconditional generation as its fundamental benchmark. Finally, we combine latent variable modeling with a recent formulation of exploration in reinforcement learning to obtain a rigorous handling of true and generated contexts. Results on language modeling and variational sentence auto-encoding confirm the model's generalization capability.

Machine learning in context of physical systems merits a re-examination of the learning strategy. In addition to data, one can leverage a vast library of physical prior models (e.g. kinematics, fluid flow, etc) to perform more robust inference. The nascent sub-field of \emph{physics-based learning} (PBL) studies the blending of neural networks with physical priors. While previous PBL algorithms have been applied successfully to specific tasks, it is hard to generalize existing PBL methods to a wide range of physics-based problems. Such generalization would require an architecture that can adapt to variations in the correctness of the physics, or in the quality of training data. No such architecture exists. In this paper, we aim to generalize PBL, by making a first attempt to bring neural architecture search (NAS) to the realm of PBL. We introduce a new method known as physics-based neural architecture search (PhysicsNAS) that is a top-performer across a diverse range of quality in the physical model and the dataset.

Under Review.I NTERPRETATIONS ARE USEFUL: PENALIZING EXPLA - NATIONS TO ALIGN NEURAL NETWORKS WITH PRIOR KNOWLEDGE Laura Rieger DTU Compute DTU 2800 Kgs. A BSTRACT For an explanation of a deep learning model to be effective, it must provide both insight into a model and suggest a corresponding action in order to achieve some objective. Too often, the litany of proposed explainable deep learning methods stop at the first step, providing practitioners with insight into a model, but no way to act on it. In this paper, we propose contextual decomposition explanation penalization (CDEP), a method which enables practitioners to leverage existing explanation methods in order to increase the predictive accuracy of deep learning models. In particular, when shown that a model has incorrectly assigned importance to some features, CDEP enables practitioners to correct these errors by directly regularizing the provided explanations. Using explanations provided by contextual decomposition (CD) (Murdoch et al., 2018), we demonstrate the ability of our method to increase performance on an array of toy and real datasets. However, in order to achieve that accuracy, they sometimes latch onto spurious correlations, leading to undesirable behavior as a result of dataset bias (Winkler et al., 2019), racial and ethnic stereotypes (Garg et al., 2018), or simply overfitting. While recent work into explaining neural network predictions (Murdoch et al., 2019; Doshi-V elez & Kim, 2017) has demonstrated an ability to uncover the relationships learned by a model, it is still unclear how to actually alter the model in order to remove incorrect, or undesirable, relationships. We introduce c ontextual d ecomposition e xplanation p enalization (CDEP), a method which leverages existing explanation techniques for neural networks in order to prevent a model from learning unwanted relationships and ultimately improve predictive accuracy.

In this paper, well-calibrated model uncertainty is obtained by using temperature scaling together with Monte Carlo dropout as approximation to Bayesian inference. The proposed approach can easily be derived from frequentist temperature scaling and yields well-calibrated model uncertainty as well as softmax likelihood.

For effectively interacting with humans in collaborative environments, machines need to be able anticipate future events, in order to execute actions in a timely manner. However, the observation of the human limbs movements may not be sufficient to anticipate their actions in an unambiguous manner. In this work we consider two additional sources of information (i.e. context) over time, gaze movements and object information, and study how these additional contextual cues improve the action anticipation performance. We address action anticipation as a classification task, where the model takes the available information as the input, and predicts the most likely action. We propose to use the uncertainty about each prediction as an online decision-making criterion for action anticipation. Uncertainty is modeled as a stochastic process applied to a time-based neural network architecture, which improves the conventional class-likelihood (i.e. deterministic) criterion. The main contributions of this paper are three-fold: (i) we propose a deep architecture that outperforms previous results in the action anticipation task; (ii) we show that contextual information is important do disambiguate the interpretation of similar actions; (iii) we propose the minimization of uncertainty as a more effective criterion for action anticipation, when compared with the maximization of class probability. Our results on the Acticipate dataset showed the importance of contextual information and the uncertainty criterion for action anticipation. We achieve an average accuracy of 98.75% in the anticipation task using only an average of 25% of observations. In addition, considering that a good anticipation model should also perform well in the action recognition task, we achieve an average accuracy of 100% in action recognition on the Acticipate dataset, when the entire observation set is used.

Invariances to translation, rotation and other spatial transformations are a hallmark of the laws of motion, and have widespread use in the natural sciences to reduce the dimensionality of systems of equations. In supervised learning, such as in image classification tasks, rotation, translation and scale invariances are used to augment training datasets. In this work, we use data augmentation in a similar way, exploiting symmetry in the quadruped domain of the DeepMind control suite (Tassa et al. 2018) to add to the trajectories experienced by the actor in the actor-critic algorithm of Abdolmaleki et al. (2018). In a data-limited regime, the agent using a set of experiences augmented through symmetry is able to learn faster. Our approach can be used to inject knowledge of invariances in the domain and task to augment learning in robots, and more generally, to speed up learning in realistic robotics applications.

Semantic Image Interpretation is the task of extracting a structured semantic description from images. This requires the detection of visual relationships: triples (subject,relation,object) describing a semantic relation between a subject and an object. A pure supervised approach to visual relationship detection requires a complete and balanced training set for all the possible combinations of (subject, relation, object). However, such training sets are not available and would require a prohibitive human effort. This implies the ability of predicting triples which do not appear in the training set. This problem is called zero-shot learning. State-of-the-art approaches to zero-shot learning exploit similarities among relationships in the training set or external linguistic knowledge. In this paper, we perform zero-shot learning by using Logic Tensor Networks, a novel Statistical Relational Learning framework that exploits both the similarities with other seen relationships and background knowledge, expressed with logical constraints between subjects, relations and objects. The experiments on the Visual Relationship Dataset show that the use of logical constraints outperforms the current methods. This implies that background knowledge can be used to alleviate the incompleteness of training sets.

Parametric models, and particularly neural networks, require weight initialization as a starting point for gradient-based optimization. In most current practices, this is accomplished by using some form of random initialization. Instead, recent work shows that a specific initial parameter set can be learned from a population of tasks, i.e., dataset and target variable for supervised learning tasks. Using this initial parameter set leads to faster convergence for new tasks (model-agnostic meta-learning). Currently, methods for learning model initializations are limited to a population of tasks sharing the same schema, i.e., the same number, order, type and semantics of predictor and target variables. In this paper, we address the problem of meta-learning parameter initialization across tasks with different schemas, i.e., if the number of predictors varies across tasks, while they still share some variables. We propose Chameleon, a model that learns to align different predictor schemas to a common representation. We use permutations and masks of the predictors of the training tasks at hand. In experiments on real-life data sets, we show that Chameleon successfully can learn parameter initializations across tasks with different schemas providing a 26% lift on accuracy on average over random initialization and of 5% over a state-of-the-art method for fixed-schema learning model initializations. To the best of our knowledge, our paper is the first work on the problem of learning model initialization across tasks with different schemas.

-- We present a relational graph learning approach for robotic crowd navigation using model-based deep reinforcement learning that plans actions by looking into the future. Our approach reasons about the relations between all agents based on their latent features and uses a Graph Convolutional Network to encode higher-order interactions in each agent's state representation, which is subsequently leveraged for state prediction and value estimation. The ability to predict human motion allows us to perform multi-step lookahead planning, taking into account the temporal evolution of human crowds. We evaluate our approach against a state-of-the-art baseline for crowd navigation and ablations of our model to demonstrate that navigation with our approach is more efficient, results in fewer collisions, and avoids failure cases involving oscillatory and freezing behaviors. I. INTRODUCTION Inferring the underlying relations between components of complex dynamic systems can inform decision making for autonomous agents. One natural system with complex dynamics is crowd navigation (i.e., navigation in the presence of multiple humans). The crowd navigation task is challenging as the agent must predict and plan relative to likely human motions so as to avoid collisions and remain at safe and socially appropriate distances from people. Some prior work predicts human trajectories using handcrafted social interaction models [1] or by modeling the temporal behavior of humans [2]. Although these methods can estimate human trajectories, they do not use the prediction to inform the navigation policy.