A major concern of Machine Learning (ML) models is their opacity. They are deployed in an increasing number of applications where they often operate as black boxes that do not provide explanations for their predictions. Among others, the potential harms associated with the lack of understanding of the models' rationales include privacy violations, adversarial manipulations, and unfair discrimination. As a result, the accountability and transparency of ML models have been posed as critical desiderata by works in policy and law, philosophy, and computer science. In computer science, the decision-making process of ML models has been studied by developing accountability and transparency methods. Accountability methods, such as adversarial attacks and diagnostic datasets, expose vulnerabilities of ML models that could lead to malicious manipulations or systematic faults in their predictions. Transparency methods explain the rationales behind models' predictions gaining the trust of relevant stakeholders and potentially uncovering mistakes and unfairness in models' decisions. To this end, transparency methods have to meet accountability requirements as well, e.g., being robust and faithful to the underlying rationales of a model. This thesis presents my research that expands our collective knowledge in the areas of accountability and transparency of ML models developed for complex reasoning tasks over text.

This paper solves the Watchman Route Problem (WRP) on a general discrete graph with Heuristic Search. Given a graph, a line-of-sight (LOS) function, and a start vertex, the task is to (offline) find a (shortest) path through the graph such that all vertices in the graph will be visually seen by at least one vertex on the path. WRP is reminiscent but different from graph covering and mapping problems, which are done online on an unknown graph. We formalize WRP as a heuristic search problem and solve it optimally with an A*-based algorithm. We develop a series of admissible heuristics with increasing difficulty and accuracy. Our heuristics abstract the underlying graph into a disjoint line-of-sight graph (GDLS) which is based on disjoint clusters of vertices such that vertices within the same cluster have LOS to the same specific vertex. We use solutions for the Minimum Spanning Tree (MST) and the Traveling Salesman Problem (TSP) of GDLS as admissible heuristics for WRP. We theoretically and empirically investigate these heuristics. Then, we show how the optimal methods can be modified (by intelligently pruning away large sub-trees) to obtain various suboptimal solvers with and without bound guarantees. These suboptimal solvers are much faster and expand fewer nodes than the optimal solver with only minor reduction in the quality of the solution.

Conceptual knowledge is fundamental to human cognition and knowledge bases. However, existing knowledge probing works only focus on evaluating factual knowledge of pre-trained language models (PLMs) and ignore conceptual knowledge. Since conceptual knowledge often appears as implicit commonsense behind texts, designing probes for conceptual knowledge is hard. Inspired by knowledge representation schemata, we comprehensively evaluate conceptual knowledge of PLMs by designing three tasks to probe whether PLMs organize entities by conceptual similarities, learn conceptual properties, and conceptualize entities in contexts, respectively. For the tasks, we collect and annotate 24k data instances covering 393 concepts, which is COPEN, a COnceptual knowledge Probing bENchmark. Extensive experiments on different sizes and types of PLMs show that existing PLMs systematically lack conceptual knowledge and suffer from various spurious correlations. We believe this is a critical bottleneck for realizing human-like cognition in PLMs. COPEN and our codes are publicly released at https://github.com/THU-KEG/COPEN.

Human behavior emerges from planning over elaborate decompositions of tasks into goals, subgoals, and low-level actions. How are these decompositions created and used? Here, we propose and evaluate a normative framework for task decomposition based on the simple idea that people decompose tasks to reduce the overall cost of planning while maintaining task performance. Analyzing 11,117 distinct graph-structured planning tasks, we find that our framework justifies several existing heuristics for task decomposition and makes predictions that can be distinguished from two alternative normative accounts. We report a behavioral study of task decomposition ($N=806$) that uses 30 randomly sampled graphs, a larger and more diverse set than that of any previous behavioral study on this topic. We find that human responses are more consistent with our framework for task decomposition than alternative normative accounts and are most consistent with a heuristic -- betweenness centrality -- that is justified by our approach. Taken together, our results provide new theoretical insight into the computational principles underlying the intelligent structuring of goal-directed behavior.

This report presents the evaluation approach developed for the DARPA Big Mechanism program, which aimed at developing computer systems that will read research papers, integrate the information into a computer model of cancer mechanisms, and frame new hypotheses. We employed an iterative, incremental approach to the evaluation of the three phases of the program. In Phase I, we evaluated the ability of system and human teams ability to read-with-a-model to capture mechanistic information from the biomedical literature, integrated with information from expert curated biological databases. In Phase II we evaluated the ability of systems to assemble fragments of information into a mechanistic model. The Phase III evaluation focused on the ability of systems to provide explanations of experimental observations based on models assembled (largely automatically) by the Big Mechanism process. The evaluation for each phase built on earlier evaluations and guided developers towards creating capabilities for the new phase. The report describes our approach, including innovations such as a reference set (a curated data set limited to major findings of each paper) to assess the accuracy of systems in extracting mechanistic findings in the absence of a gold standard, and a method to evaluate model-based explanations of experimental data. Results of the evaluation and supporting materials are included in the appendices.

Many recent publications report on methods for achieving safety in Markov Decision Processes (MDPs), where temporal logic (safety) specifications must be satisfied [1-4]. However, it is typically assumed that 1) the safety specification is given, and 2) that the states in the underlying MDP are unstructured. In this paper, we are interested in 1) learning the safety specification from examples, and 2) working with relational MDPs. More specifically, in our learning setting we assume that there is a domain expert who is presented with a set of system states E, a probability threshold α and a step-bound k (number of action executions). If the expert believes that the system, starting in s E will perform actions that lead to a dangerous temporal situation within k steps with probability at least α, then she will label s as dangerous, else, as safe. Now, given this set E of labeled states, we want to learn a compact temporal logic formula summarizing the expert's advice. There are at least three reasons to infer a property (expressed as a temporal logic formula) from an expert's advice. Firstly, to obtain a concise, human-interpretable expression of some aspects of the domain [5-7], secondly, to verify a system's control behavior (policy) w.r.t. a set of (safety) standards [6, 8] and thirdly, to use the (safety) property to devise strategies for the system or agent to avoid undesirable situations [8-10]. Furthermore, we consider systems that can be modelled as relational MDPs (RMDPs).

This paper introduces and presents a new language named MAIL (Malware Analysis Intermediate Language). MAIL is basically used for building malware analysis and detection tools. MAIL provides an abstract representation of an assembly program and hence the ability of a tool to automate malware analysis and detection. By translating binaries compiled for different platforms to MAIL, a tool can achieve platform independence. Each MAIL statement is annotated with patterns that can be used by a tool to optimize malware analysis and detection.

A growing body of work studies how to answer a question or verify a claim by generating a natural language "proof": a chain of deductive inferences yielding the answer based on a set of premises. However, these methods can only make sound deductions when they follow from evidence that is given. We propose a new system that can handle the underspecified setting where not all premises are stated at the outset; that is, additional assumptions need to be materialized to prove a claim. By using a natural language generation model to abductively infer a premise given another premise and a conclusion, we can impute missing pieces of evidence needed for the conclusion to be true. Our system searches over two fringes in a bidirectional fashion, interleaving deductive (forward-chaining) and abductive (backward-chaining) generation steps. We sample multiple possible outputs for each step to achieve coverage of the search space, at the same time ensuring correctness by filtering low-quality generations with a round-trip validation procedure. Results on a modified version of the EntailmentBank dataset and a new dataset called Everyday Norms: Why Not? show that abductive generation with validation can recover premises across in- and out-of-domain settings.

In September 2016, Stanford's "One Hundred Year Study on Artificial Intelligence" project (AI100) issued the first report of its planned long-term periodic assessment of artificial intelligence (AI) and its impact on society. It was written by a panel of 17 study authors, each of whom is deeply rooted in AI research, chaired by Peter Stone of the University of Texas at Austin. The report, entitled "Artificial Intelligence and Life in 2030," examines eight domains of typical urban settings on which AI is likely to have impact over the coming years: transportation, home and service robots, healthcare, education, public safety and security, low-resource communities, employment and workplace, and entertainment. It aims to provide the general public with a scientifically and technologically accurate portrayal of the current state of AI and its potential and to help guide decisions in industry and governments, as well as to inform research and development in the field. The charge for this report was given to the panel by the AI100 Standing Committee, chaired by Barbara Grosz of Harvard University.

She believed that we can be hopeful that the answer is yes and that it may happen sooner than we might expect. As scientific questions become significantly more complex, our capabilities to do scientific breakthroughs need to be augmented. Compare for instance the challenges of formulating Kepler's laws of planetary motion or the discovery of a cure for Polio with demonstrating the existence of binary stellar-mass black hole systems (Abbott et al., 2016) or the treatment of glioblastoma, a type of brain cancer. While the former were achieved by a single scientist, the latter require large and interdisciplinary teams involving the collaboration of hundreds of scientists from different fields to work together during years to produce results.