Self-training is one of the earliest and simplest semi-supervised methods. The key idea is to augment the original labeled dataset with unlabeled data paired with the model's prediction (i.e. pseudo-parallel data). While self-training has been extensively studied on classification problems, in complex sequence generation tasks (e.g. machine translation) it is still unclear how self-training works due to the compositionality of the target space. In this work, we first empirically show that self-training is able to decently improve the supervised baseline on neural sequence generation tasks. Through careful examination of the performance gains, we find that the perturbation on the hidden states (i.e. dropout) is critical for self-training to benefit from the pseudo-parallel data, which acts as a regularizer and forces the model to yield close predictions for similar unlabeled inputs. Such effect helps the model correct some incorrect predictions on unlabeled data. To further encourage this mechanism, we propose to inject noise to the input space, resulting in a "noisy" version of self-training. Empirical study on standard machine translation and text summarization benchmarks shows that noisy self-training is able to effectively utilize unlabeled data and improve the performance of the supervised baseline by a large margin.

Being able to interpret, or explain, the predictions made by a machine learning model is of fundamental importance. This is especially true when there is interest in deploying data-driven models to make high-stakes decisions, e.g. in healthcare. While recent years have seen an increasing interest in interpretable machine learning research, this field is currently lacking an agreed-upon definition of interpretability, and some researchers have called for a more active conversation towards a rigorous approach to interpretability. Joining this conversation, we claim in this paper that the difficulty of interpreting a complex model stems from the existing interactions among features. We argue that by enforcing monotonicity between features and outputs, we are able to reason about the effect of a single feature on an output independently from other features, and consequently better understand the model. We show how to structurally introduce this constraint in deep learning models by adding new simple layers. We validate our model on benchmark datasets, and compare our results with previously proposed interpretable models.

We consider the problem of engineering robust direct perception neural networks with output being regression. Such networks take high dimensional input image data, and they produce affordances such as the curvature of the upcoming road segment or the distance to the front vehicle. Our proposal starts by allowing a neural network prediction to deviate from the label with tolerance $\Delta$. The source of tolerance can be either contractual or from limiting factors where two entities may label the same data with slightly different numerical values. The tolerance motivates the use of a non-standard loss function where the loss is set to $0$ so long as the prediction-to-label distance is less than $\Delta$. We further extend the loss function and define a new provably robust criterion that is parametric to the allowed output tolerance $\Delta$, the layer index $\tilde{l}$ where perturbation is considered, and the maximum perturbation amount $\kappa$. During training, the robust loss is computed by first propagating symbolic errors from the $\tilde{l}$-th layer (with quantity bounded by $\kappa$) to the output layer, followed by computing the overflow between the error bounds and the allowed tolerance. The overall concept is experimented in engineering a direct perception neural network for understanding the central position of the ego-lane in pixel coordinates.

Network representation learning has exploded recently. However, existing studies usually reconstruct networks as sequences or matrices, which may cause information bias or sparsity problem during model training. Inspired by a cognitive model of human memory, we propose a network representation learning scheme. In this scheme, we learn node embeddings by adjusting the proximity of nodes traversing the spreading structure of the network. Our proposed method shows a significant improvement in multiple analysis tasks based on various real-world networks, ranging from semantic networks to protein interaction networks, international trade networks, human behavior networks, etc. In particular, our model can effectively discover the hierarchical structures in networks. The well-organized model training speeds up the convergence to only a small number of iterations, and the training time is linear with respect to the edge numbers.

A BSTRACT In this paper we investigate an artificial agent's ability to perform task-focused tool synthesis via imagination. Our motivation is to explore the richness of information captured by the latent space of an object-centric generative model - and how to exploit it. In particular, our approach employs activation maximisation of a task-based performance predictor to optimise the latent variable of a structured latent-space model in order to generate tool geometries appropriate for the task at hand. We evaluate our model using a novel dataset of synthetic reaching tasks inspired by the cognitive sciences and behavioural ecology. In doing so we examine the model's ability to imagine tools for increasingly complex scenario types, beyond those seen during training. Our experiments demonstrate that the synthesis process modifies emergent, task-relevant object affordances in a targeted and deliberate way: the agents often specifically modify aspects of the tools which relate to meaningful (yet implicitly learned) concepts such as a tool's length, width and configuration. Our results therefore suggest that task relevant object affordances are implicitly encoded as directions in a structured latent space shaped by experience. 1 I NTRODUCTION Deep generative models are gaining in popularity for unsupervised representation learning. In particular, recent models like MONet (Burgess et al., 2019) have been proposed to decompose scenes into object-centric latent representations (cf. The notion of such an object-centric latent representation, trained from examples in an unsupervised way, holds a tantalising prospect: as generative models naturally capture factors of variation, could they also be used to expose these factors such that they can be modified in a task-driven way? We posit that a task-driven traversal of a structured latent space leads to affordances emerging naturally as directions in this space. This is in stark contrast to more common approaches to affordance learning where it is commonly achieved via direct supervision or implicitly via imitation (e.g. Tikhanoff et al., 2013; Myers et al., 2015; Liu et al., 2018; Grabner et al., 2011; Do et al., 2018).

A BSTRACT We show how neural models can be used to realize piece-wise constant functions such as decision trees. Our approach builds on ReLU networks that are piece-wise linear and hence their associated gradients with respect to the inputs are locally constant. We formally establish the equivalence between the classes of locally constant networks and decision trees. Moreover, we highlight several advantageous properties of locally constant networks, including how they realize decision trees with parameter sharing across branching / leaves. Indeed, only M neurons suffice to implicitly model an oblique decision tree with 2 M leaf nodes. The neural representation also enables us to adopt many tools developed for deep networks (e.g., DropConnect (Wan et al., 2013)) while implicitly training decision trees. We demonstrate that our method outperforms alternative techniques for training oblique decision trees in the context of molecular property classification and regression tasks. 1 I NTRODUCTION Decision trees (Breiman et al., 1984) employ a series of simple decision nodes, arranged in a tree, to transparently capture how the predicted outcome is reached. Functionally, such tree-based models, including random forest (Breiman, 2001), realize piece-wise constant functions. Beyond their status as de facto interpretable models, they have also persisted as the state of the art models in some tabular (Sandulescu & Chiru, 2016) and chemical datasets (Wu et al., 2018). Deep neural models, in contrast, are highly flexible and continuous, demonstrably effective in practice, though lack transparency. We merge these two contrasting views by introducing a new family of neural models that implicitly learn and represent oblique decision trees. Prior work has attempted to generalize classic decision trees by extending coordinate-wise cuts to be weighted, linear classifications.

The performance of many network learning applications crucially hinges on the success of network embedding algorithms, which aim to encode rich network information into low-dimensional vertex-based vector representations. This paper considers a novel variational formulation of network embeddings, with special focus on textual networks. Different from most existing methods that optimize a discriminative objective, we introduce Variational Homophilic Embedding (VHE), a fully generative model that learns network embeddings by modeling the semantic (textual) information with a variational autoencoder, while accounting for the structural (topology) information through a novel homophilic prior design. Homophilic vertex embeddings encourage similar embedding vectors for related (connected) vertices. The proposed VHE promises better generalization for downstream tasks, robustness to incomplete observations, and the ability to generalize to unseen vertices. Extensive experiments on real-world networks, for multiple tasks, demonstrate that the proposed method consistently achieves superior performance relative to competing state-of-the-art approaches.

Dynamic Interaction-A ware Scene Understanding for Reinforcement Learning in Autonomous Driving Maria Huegle 1, Gabriel Kalweit 1, Moritz Werling 2 and Joschka Boedecker 1, 3 Abstract -- The common pipeline in autonomous driving systems is highly modular and includes a perception component which extracts lists of surrounding objects and passes these lists to a high-level decision component. In this case, leveraging the benefits of deep reinforcement learning for high-level decision making requires special architectures to deal with multiple variable-length sequences of different object types, such as vehicles, lanes or traffic signs. At the same time, the architecture has to be able to cover interactions between traffic participants in order to find the optimal action to be taken. In this work, we propose the novel Deep Scenes architecture, that can learn complex interaction-aware scene representations based on extensions of either 1) Deep Sets or 2) Graph Convolutional Networks. We present the Graph-Q and DeepScene-Q off-policy reinforcement learning algorithms, both outperforming state-of- the-art methods in evaluations with the publicly available traffic simulator SUMO. I. INTRODUCTION In autonomous driving scenarios, the number of traffic participants and lanes surrounding the agent can vary considerably over time. Common autonomous driving systems use modular pipelines, where a perception component extracts a list of surrounding objects and passes this list to other modules, including localization, mapping, motion planning and high-level decision making components.

In the past few years, off-policy reinforcement learning methods have shown promising results in their application for robot control. Deep Q-learning, however, still suffers from poor data-efficiency which is limiting with regard to real-world applications. We follow the idea of multi-step TD-learning to enhance data-efficiency while remaining off-policy by proposing two novel Temporal-Difference formulations: (1) Truncated Q-functions which represent the return for the first n steps of a policy rollout and (2) Shifted Q-functions, acting as the farsighted return after this truncated rollout. We prove that the combination of these short-and long-term predictions is a representation of the full return, leading to the Composite Q-learning algorithm. We show the efficacy of Composite Q-learning in the tabular case and compare our approach in the function-approximation setting with TD3, Model-based V alue Expansion and TD3(), which we introduce as an off-policy variant of TD(). We show on three simulated robot tasks that Composite TD3 outperforms TD3 as well as state-of-the-art off-policy multi-step approaches in terms of data-efficiency. In recent years, Q-learning (Watkins and Dayan, 1992) has achieved major successes in a broad range of areas by employing deep neural networks (Mnih et al., 2015; Silver et al., 2018; Lillicrap et al., 2016), including environments of higher complexity (Riedmiller et al., 2018) and even in first real world applications (Haarnoja et al., 2019). Due to its off-policy update, Q-learning can leverage transitions collected by any policy which makes it more data-efficient compared to on-policy methods.

Robotic navigation through crowds or herds requires the ability to both predict the future motion of nearby individuals and understand how these predictions might change in response to a robot's future action. State of the art trajectory prediction models using Recurrent Neural Networks (RNNs) do not currently account for a planned future action of a robot, and so cannot predict how an individual will move in response to a robot's planned path. We propose an approach that adapts RNNs to use a robot's next planned action as an input alongside the current position of nearby individuals. This allows the model to learn the response of individuals with regards to a robot's motion from real world observations. By linking a robot's actions to the response of those around it in training, we show that we are able to not only improve prediction accuracy in close range interactions, but also to predict the likely response of surrounding individuals to simulated actions. This allows the use of the model to simulate state transitions, without requiring any assumptions on agent interaction. We apply this model to varied datasets, including crowds of pedestrians interacting with vehicles and bicycles, and livestock interacting with a robotic vehicle.