A generalist robot must be able to complete a variety of tasks in its environment. One appealing way to specify each task is in terms of a goal observation. However, learning goal-reaching policies with reinforcement learning remains a challenging problem, particularly when hand-engineered reward functions are not available. Learned dynamics models are a promising approach for learning about the environment without rewards or task-directed data, but planning to reach goals with such a model requires a notion of functional similarity between observations and goal states. We present a self-supervised method for model-based visual goal reaching, which uses both a visual dynamics model as well as a dynamical distance function learned using model-free reinforcement learning. Our approach learns entirely using offline, unlabeled data, making it practical to scale to large and diverse datasets. In our experiments, we find that our method can successfully learn models that perform a variety of tasks at test-time, moving objects amid distractors with a simulated robotic arm and even learning to open and close a drawer using a real-world robot. In comparisons, we find that this approach substantially outperforms both model-free and model-based prior methods. Videos and visualizations are available here: https://sites.google.com/berkeley.edu/mbold. Designing general-purpose robots that can perform a wide range of tasks remains an open problem in AI and robotics. Reinforcement learning (RL) represents a particularly promising tool for learning robotic behaviors when skills can be learned one at a time from user-defined reward functions. However, general-purpose robots will likely require large and diverse repertoires of skills, and learning individual tasks one at a time from manually-specified rewards is onerous and time-consuming. How can we design learning systems that can autonomously acquire general-purpose knowledge that allows them to solve many different downstream tasks? To address this problem, we must resolve three questions.

Training a classifier under fairness constraints has gotten increasing attention in the machine learning community thanks to moral, legal, and business reasons. However, several recent works addressing algorithmic fairness have only focused on simple models such as logistic regression or support vector machines due to non-convex and non-differentiable fairness criteria across protected groups, such as race or gender. Neural networks, the most widely used models for classification nowadays, are precluded and lack theoretical guarantees. This paper aims to fill this missing but crucial part of the literature of algorithmic fairness for neural networks. In particular, we show that overparametrized neural networks could meet the fairness constraints. The key ingredient of building a fair neural network classifier is establishing no-regret analysis for neural networks in the overparameterization regime, which may be of independent interest in the online learning of neural networks and related applications.

Deep Reinforcement Learning is a topic that has gained a lot of attention recently, due to the unprecedented achievements and remarkable performance of such algorithms in various benchmark tests and environmental setups. The power of such methods comes from the combination of an already established and strong field of Deep Learning, with the unique nature of Reinforcement Learning methods. It is, however, deemed necessary to provide a compact, accurate and comparable view of these methods and their results for the means of gaining valuable technical and practical insights. In this work we gather the essential methods related to Deep Reinforcement Learning, extracting common property structures for three complementary core categories: a) Model-Free, b) Model-Based and c) Modular algorithms. For each category, we present, analyze and compare state-of-the-art Deep Reinforcement Learning algorithms that achieve high performance in various environments and tackle challenging problems in complex and demanding tasks. In order to give a compact and practical overview of their differences, we present comprehensive comparison figures and tables, produced by reported performances of the algorithms under two popular simulation platforms: the Atari Learning Environment and the MuJoCo physics simulation platform. We discuss the key differences of the various kinds of algorithms, indicate their potential and limitations, as well as provide insights to researchers regarding future directions of the field.

Pre-trained Language Models (PLMs) have shown strong performance in various downstream Natural Language Processing (NLP) tasks. However, PLMs still cannot well capture the factual knowledge in the text, which is crucial for understanding the whole text, especially for document-level language understanding tasks. To address this issue, we propose a novel contrastive learning framework named ERICA in pre-training phase to obtain a deeper understanding of the entities and their relations in text. Specifically, (1) to better understand entities, we propose an entity discrimination task that distinguishes which tail entity can be inferred by the given head entity and relation. (2) Besides, to better understand relations, we employ a relation discrimination task which distinguishes whether two entity pairs are close or not in relational semantics. Experimental results demonstrate that our proposed ERICA framework achieves consistent improvements on several document-level language understanding tasks, including relation extraction and reading comprehension, especially under low resource setting. Meanwhile, ERICA achieves comparable or better performance on sentence-level tasks. We will release the datasets, source codes and pre-trained language models for further research explorations.

Online class imbalance learning deals with data streams that are affected by both concept drift and class imbalance. Online learning tries to find a trade-off between exploiting previously learned information and incorporating new information into the model. This requires both the incremental update of the model and the ability to unlearn outdated information. The improper use of unlearning, however, can lead to the retroactive interference problem, a phenomenon that occurs when newly learned information interferes with the old information and impedes the recall of previously learned information. The problem becomes more severe when the classes are not equally represented, resulting in the removal of minority information from the model. In this work, we propose the Drift-Aware Multi-Memory Model (DAM3), which addresses the class imbalance problem in online learning for memory-based models. DAM3 mitigates class imbalance by incorporating an imbalance-sensitive drift detector, preserving a balanced representation of classes in the model, and resolving retroactive interference using a working memory that prevents the forgetting of old information. We show through experiments on real-world and synthetic datasets that the proposed method mitigates class imbalance and outperforms the state-of-the-art methods.

The multi-armed bandit (MAB) is a simple instance of online learning problem with partial feedback (see Bubeck and Cesa-Bianchi [13]; Lattimore and Szepesvári [34]; Slivkins [44], and references therein). In the stochastic bandit framework, a learning agent sequentially pulls arms and obtains noisy rewards. The goal of the agent is to maximize her expected cumulative reward, or equivalently, to minimize her regret, which is the difference between the cumulative reward of an oracle (that knows the mean rewards of arms) and the one of the agent. There is thus a clear trade-off, called "exploration vs exploitation", that arises between gathering information on uncertain arms by pulling them (exploration) and leveraging the information obtained so far by greedily pulling the arm with the highest estimated reward (exploitation). MAB has received a large attention recently due to its wide applications range and the theoretical guarantees associated with learning algorithms (see also Bouneffouf and Rish [9], and references therein for a survey on practical applications of MAB).

Manifold learning methods play a prominent role in nonlinear dimensionality reduction and other tasks involving high-dimensional data sets with low intrinsic dimensionality. Many of these methods are graph-based: they associate a vertex with each data point and a weighted edge between each pair of close points. Existing theory shows, under certain conditions, that the Laplacian matrix of the constructed graph converges to the Laplace-Beltrami operator of the data manifold. However, this result assumes the Euclidean norm is used for measuring distances. In this paper, we determine the limiting differential operator for graph Laplacians constructed using $\textit{any}$ norm. The proof involves a subtle interplay between the second fundamental form of the underlying manifold and the convex geometry of the norm's unit ball. To motivate the use of non-Euclidean norms, we show in a numerical simulation that manifold learning based on Earthmover's distances outperforms the standard Euclidean variant for learning molecular shape spaces, in terms of both sample complexity and computational complexity.

A short review of the literature on measurement and detection of artificial general intelligence is made. Proposed benchmarks and tests for artificial general intelligence are critically evaluated against multiple criteria. Based on the findings, the most promising approaches are identified and some useful directions for future work are proposed.

Dialogue policy learning based on reinforcement learning is difficult to be applied to real users to train dialogue agents from scratch because of the high cost. User simulators, which choose random user goals for the dialogue agent to train on, have been considered as an affordable substitute for real users. However, this random sampling method ignores the law of human learning, making the learned dialogue policy inefficient and unstable. We propose a novel framework, Automatic Curriculum Learning-based Deep Q-Network (ACL-DQN), which replaces the traditional random sampling method with a teacher policy model to realize the dialogue policy for automatic curriculum learning. The teacher model arranges a meaningful ordered curriculum and automatically adjusts it by monitoring the learning progress of the dialogue agent and the over-repetition penalty without any requirement of prior knowledge. The learning progress of the dialogue agent reflects the relationship between the dialogue agent's ability and the sampled goals' difficulty for sample efficiency. The over-repetition penalty guarantees the sampled diversity. Experiments show that the ACL-DQN significantly improves the effectiveness and stability of dialogue tasks with a statistically significant margin. Furthermore, the framework can be further improved by equipping with different curriculum schedules, which demonstrates that the framework has strong generalizability.

Solving algebra story problems remains a challenging task in artificial intelligence, which requires a detailed understanding of real-world situations and a strong mathematical reasoning capability. Previous neural solvers of math word problems directly translate problem texts into equations, lacking an explicit interpretation of the situations, and often fail to handle more sophisticated situations. To address such limits of neural solvers, we introduce the concept of a \emph{situation model}, which originates from psychology studies to represent the mental states of humans in problem-solving, and propose \emph{SMART}, which adopts attributed grammar as the representation of situation models for algebra story problems. Specifically, we first train an information extraction module to extract nodes, attributes, and relations from problem texts and then generate a parse graph based on a pre-defined attributed grammar. An iterative learning strategy is also proposed to improve the performance of SMART further. To rigorously study this task, we carefully curate a new dataset named \emph{ASP6.6k}. Experimental results on ASP6.6k show that the proposed model outperforms all previous neural solvers by a large margin while preserving much better interpretability. To test these models' generalization capability, we also design an out-of-distribution (OOD) evaluation, in which problems are more complex than those in the training set. Our model exceeds state-of-the-art models by 17\% in the OOD evaluation, demonstrating its superior generalization ability.