Recently, barrier function-based safe reinforcement learning (RL) with the actor-critic structure for continuous control tasks has received increasing attention. It is still challenging to learn a near-optimal control policy with safety and convergence guarantees. Also, few works have addressed the safe RL algorithm design under time-varying safety constraints. This paper proposes a model-based safe RL algorithm for optimal control of nonlinear systems with time-varying state and control constraints. In the proposed approach, we construct a novel barrier-based control policy structure that can guarantee control safety. A multi-step policy evaluation mechanism is proposed to predict the policy's safety risk under time-varying safety constraints and guide the policy to update safely. Theoretical results on stability and robustness are proven. Also, the convergence of the actor-critic learning algorithm is analyzed. The performance of the proposed algorithm outperforms several state-of-the-art RL algorithms in the simulated Safety Gym environment. Furthermore, the approach is applied to the integrated path following and collision avoidance problem for two real-world intelligent vehicles. A differential-drive vehicle and an Ackermann-drive one are used to verify the offline deployment performance and the online learning performance, respectively. Our approach shows an impressive sim-to-real transfer capability and a satisfactory online control performance in the experiment.

Abstract--Machine learning researchers have long noted a trade-off between interpretability and prediction performance. On the one hand, traditional models are often interpretable to humans but they cannot achieve high prediction performances. At the opposite end of the spectrum, deep models can achieve state-of-the-art performances in many tasks. However, deep models' predictions are known to be uninterpretable to humans. In this paper we present a framework that shortens the gap between the two aforementioned groups of methods. Given an artificial neural network (ANN), our method finds a Gaussian process (GP) whose predictions almost match those of the ANN. As GPs are highly interpretable, we use the trained GP to explain the ANN's decisions. We use our method to explain ANNs' decisions on may datasets. The explanations provide intriguing insights about the ANNs' decisions. With the best of our knowledge, our inference formulation for GPs is the first one in which an ANN and a similarly behaving Gaussian process naturally appear. Furthermore, we examine some of the known theoretical conditions under which an ANN is interpretable by GPs. Some of those theoretical conditions are too restrictive for modern architectures. However, we hypothesize that only a subset of those theoretical conditions are sufficient. Finally, we implement our framework as a publicly available tool called GPEX. Given any pytorch feed-forward module, GPEX allows users to interpret any ANN subcomponent of the module effortlessly and without having to be involved in the inference algorithm.

Abstract--Training with the true labels of a dataset as opposed to randomized labels leads to faster optimization and better generalization. This difference is attributed to a notion of alignment between inputs and labels in natural datasets. We find that training neural networks with different architectures and optimizers on random or true labels enforces the same relationship between the hidden representations and the training labels, elucidating why neural network representations have been so successful for transfer. We first highlight why aligned features promote transfer and show in a classic synthetic transfer problem that alignment is the determining factor for positive and negative transfer to similar and dissimilar tasks. We then investigate a variety of neural network architectures and find that (a) alignment emerges across a variety of different architectures and optimizers, with more alignment arising from depth (b) alignment increases for layers closer to the output and (c) existing high-performance deep CNNs exhibit high levels of alignment. One direction has been to analyze what abstractions the of that network and finally train a subsequent model on network has learned, agnostic to exactly how it is represented. The Shwartz et al. [12] studied neural networks through the lens premise is that neural networks adapt their intermediate of information theory and found that, during training, the representations--hidden representations--to the source task network preserves the information necessary for predicting and, due to the commonalities between the two tasks, these the output while throwing away unnecessary information learned representations help training on the target task [1]. Using representational Availability of large datasets like ImageNet [2] and the News similarity matrices, [13] found that on synthetic datasets Dataset for Word2Vec [3] provides suitable source tasks that where task-relevance of features can be controlled, learned facilitate using neural networks for feature construction for hidden representations suppress task-irrelevant features and Computer Vision and Natural Language Processing (NLP) enhance task-relevant features.

Goal-achieving problems are puzzles that set up a specific situation with a clear objective. An example that is well-studied is the category of life-and-death (L&D) problems for Go, which helps players hone their skill of identifying region safety. Many previous methods like lambda search try null moves first, then derive so-called relevance zones (RZs), outside of which the opponent does not need to search. This paper first proposes a novel RZ-based approach, called the RZ-Based Search (RZS), to solving L&D problems for Go. RZS tries moves before determining whether they are null moves post-hoc. This means we do not need to rely on null move heuristics, resulting in a more elegant algorithm, so that it can also be seamlessly incorporated into AlphaZero's super-human level play in our solver. To repurpose AlphaZero for solving, we also propose a new training method called Faster to Life (FTL), which modifies AlphaZero to entice it to win more quickly. We use RZS and FTL to solve L&D problems on Go, namely solving 68 among 106 problems from a professional L&D book while a previous program solves 11 only. Finally, we discuss that the approach is generic in the sense that RZS is applicable to solving many other goal-achieving problems for board games.

A major component for developing intelligent and autonomous robots is a suitable knowledge representation, from which a robot can acquire knowledge about its actions or world. However, unlike humans, robots cannot creatively adapt to novel scenarios, as their knowledge and environment are rigidly defined. To address the problem of producing novel and flexible task plans called task trees, we explore how we can derive plans with concepts not originally in the robot's knowledge base. Existing knowledge in the form of a knowledge graph is used as a base of reference to create task trees that are modified with new object or state combinations. To demonstrate the flexibility of our method, we randomly selected recipes from the Recipe1M+ dataset and generated their task trees. The task trees were then thoroughly checked with a visualization tool that portrays how each ingredient changes with each action to produce the desired meal. Our results indicate that the proposed method can produce task plans with high accuracy even for never-before-seen ingredient combinations.

In this research note we show that a simple justification system can be used to teach humans non-trivial strategies of the Olympic sport of curling. This is achieved by justifying the decisions of Kernel Regression UCT (KR-UCT), a tree search algorithm that derives curling strategies by playing the game with itself. Given an action returned by KR-UCT and the expected outcome of that action, we use a decision tree to produce a counterfactual justification of KR-UCT's decision. The system samples other possible outcomes and selects for presentation the outcomes that are most similar to the expected outcome in terms of visual features and most different in terms of expected end-game value. A user study with 122 people shows that the participants who had access to the justifications produced by our system achieved much higher scores in a curling test than those who only observed the decision made by KR-UCT and those with access to the justifications of a baseline system. This is, to the best of our knowledge, the first work showing that a justification system is able to teach humans non-trivial strategies learned by an algorithm operating in self play.

In many recent efforts [1, 2, 3, 4], 3D human pose estimation has been decomposed into a two-stage process: first, the 2D keypoints that correspond to the body joints are detected from the 2D image, after which the detected joints are lifted to obtain 3D pose. This type of solution is elegant in terms of the simplicity of problem formulation, unfortunately it suffers from inherent ambiguities caused by projection: different 3D poses can share the same 2D pose projection given a specific viewpoint; that is, the mapping between the 2D joints detection and 3D pose is not bijective. To resolve this ambiguity of 3D pose estimation from a monocular image, video-based pose estimation is also investigated in the literature [5, 6]. Existing video-based pose estimation methods, however, either need to observe a relatively long history (243 frames [5] or can only handle a short video sequence (4-6 frames [6]) to achieve their best results.

Temporal difference (TD) learning and its variants, such as multistage TD (MS-TD) learning and temporal coherence (TC) learning, have been successfully applied to 2048. These methods rely on the stochasticity of the environment of 2048 for exploration. In this paper, we propose to employ optimistic initialization (OI) to encourage exploration for 2048, and empirically show that the learning quality is significantly improved. This approach optimistically initializes the feature weights to very large values. Since weights tend to be reduced once the states are visited, agents tend to explore those states which are unvisited or visited few times. Our experiments show that both TD and TC learning with OI significantly improve the performance. As a result, the network size required to achieve the same performance is significantly reduced. With additional tunings such as expectimax search, multistage learning, and tile-downgrading technique, our design achieves the state-of-the-art performance, namely an average score of 625 377 and a rate of 72% reaching 32768 tiles. In addition, for sufficiently large tests, 65536 tiles are reached at a rate of 0.02%.

Adversarial patch attack aims to fool a machine learning model by arbitrarily modifying pixels within a restricted region of an input image. Such attacks are a major threat to models deployed in the physical world, as they can be easily realized by presenting a customized object in the camera view. Defending against such attacks is challenging due to the arbitrariness of patches, and existing provable defenses suffer from poor certified accuracy. In this paper, we propose PatchVeto, a zero-shot certified defense against adversarial patches based on Vision Transformer (ViT) models. Rather than training a robust model to resist adversarial patches which may inevitably sacrifice accuracy, PatchVeto reuses a pretrained ViT model without any additional training, which can achieve high accuracy on clean inputs while detecting adversarial patched inputs by simply manipulating the attention map of ViT. Specifically, each input is tested by voting over multiple inferences with different attention masks, where at least one inference is guaranteed to exclude the adversarial patch. The prediction is certifiably robust if all masked inferences reach consensus, which ensures that any adversarial patch would be detected with no false negative. Extensive experiments have shown that PatchVeto is able to achieve high certified accuracy (e.g. 67.1% on ImageNet for 2%-pixel adversarial patches), significantly outperforming state-of-the-art methods. The clean accuracy is the same as vanilla ViT models (81.8% on ImageNet) since the model parameters are directly reused. Meanwhile, our method can flexibly handle different adversarial patch sizes by simply changing the masking strategy.

Cross-domain recommendation can help alleviate the data sparsity issue in traditional sequential recommender systems. In this paper, we propose the RecGURU algorithm framework to generate a Generalized User Representation (GUR) incorporating user information across domains in sequential recommendation, even when there is minimum or no common users in the two domains. We propose a self-attentive autoencoder to derive latent user representations, and a domain discriminator, which aims to predict the origin domain of a generated latent representation. We propose a novel adversarial learning method to train the two modules to unify user embeddings generated from different domains into a single global GUR for each user. The learned GUR captures the overall preferences and characteristics of a user and thus can be used to augment the behavior data and improve recommendations in any single domain in which the user is involved. Extensive experiments have been conducted on two public cross-domain recommendation datasets as well as a large dataset collected from real-world applications. The results demonstrate that RecGURU boosts performance and outperforms various state-of-the-art sequential recommendation and cross-domain recommendation methods. The collected data will be released to facilitate future research.