Computer Science

Medical toxicology is the clinical specialty that treats the toxic effects of substances, be it an overdose, a medication error, or a scorpion sting. The volume of toxicological knowledge and research has, as with other medical specialties, outstripped the ability of the individual clinician to entirely master and stay current with it. The application of machine learning techniques to medical toxicology is challenging because initial treatment decisions are often based on a few pieces of textual data and rely heavily on prior knowledge. ML techniques often do not represent knowledge in a way that is transparent for the physician, raising barriers to usability. Rule-based systems and decision tree learning are more transparent approaches, but often generalize poorly and require expert curation to implement and maintain. Here, we construct a probabilistic logic network to represent a portion of the knowledge base of a medical toxicologist. Our approach transparently mimics the knowledge representation and clinical decision-making of practicing clinicians. The software, dubbed Tak, performs comparably to humans on straightforward cases and intermediate difficulty cases, but is outperformed by humans on challenging clinical cases. Tak outperforms a decision tree classifier at all levels of difficulty. Probabilistic logic provides one form of explainable artificial intelligence that may be more acceptable for use in healthcare, if it can achieve acceptable levels of performance.

The integration of reasoning, learning, and decision-making is key to build more general AI systems. As a step in this direction, we propose a novel neural-logic architecture that can solve both inductive logic programming (ILP) and deep reinforcement learning (RL) problems. Our architecture defines a restricted but expressive continuous space of first-order logic programs by assigning weights to predicates instead of rules. Therefore, it is fully differentiable and can be efficiently trained with gradient descent. Besides, in the deep RL setting with actor-critic algorithms, we propose a novel efficient critic architecture. Compared to state-of-the-art methods on both ILP and RL problems, our proposition achieves excellent performance, while being able to provide a fully interpretable solution and scaling much better, especially during the testing phase.

Commonsense knowledge is essential for many AI applications, including those in natural language processing, visual processing, and planning. Consequently, many sources that include commonsense knowledge have been designed and constructed over the past decades. Recently, the focus has been on large text-based sources, which facilitate easier integration with neural (language) models and application on textual tasks, typically at the expense of the semantics of the sources. Such practice prevents the harmonization of these sources, understanding their coverage and gaps, and may hinder the semantic alignment of their knowledge with downstream tasks. Efforts to consolidate commonsense knowledge have yielded partial success, but provide no clear path towards a comprehensive consolidation of existing commonsense knowledge. The ambition of this paper is to organize these sources around a common set of dimensions of commonsense knowledge. For this purpose, we survey a wide range of popular commonsense sources with a special focus on their relations. We consolidate these relations into 13 knowledge dimensions, each abstracting over more specific relations found in sources. This consolidation allows us to unify the separate sources and to compute indications of their coverage, overlap, and gaps with respect to the knowledge dimensions. Moreover, we analyze the impact of each dimension on downstream reasoning tasks that require commonsense knowledge, observing that the temporal and desire/goal dimensions are very beneficial for reasoning on current downstream tasks, while distinctness and lexical knowledge have little impact. These results reveal focus towards some dimensions in current evaluation, and potential neglect of others.

A generative model is a statistical model that is able to generate new data instances from previously observed ones. In the context of business processes, a generative model creates new execution traces from a set of historical traces, also known as an event log. Two families of generative process simulation models have been developed in previous work: data-driven simulation models and deep learning models. Until now, these two approaches have evolved independently and their relative performance has not been studied. This paper fills this gap by empirically comparing a data-driven simulation technique with multiple deep learning techniques, which construct models are capable of generating execution traces with timestamped events. The study sheds light into the relative strengths of both approaches and raises the prospect of developing hybrid approaches that combine these strengths.

We study the problem of how to construct a set of policies that can be composed together to solve a collection of reinforcement learning tasks. Each task is a different reward function defined as a linear combination of known features. We consider a specific class of policy compositions which we call set improving policies (SIPs): given a set of policies and a set of tasks, a SIP is any composition of the former whose performance is at least as good as that of its constituents across all the tasks. We focus on the most conservative instantiation of SIPs, set-max policies (SMPs), so our analysis extends to any SIP. This includes known policy-composition operators like generalized policy improvement. Our main contribution is a policy iteration algorithm that builds a set of policies in order to maximize the worst-case performance of the resulting SMP on the set of tasks. The algorithm works by successively adding new policies to the set. We show that the worst-case performance of the resulting SMP strictly improves at each iteration, and the algorithm only stops when there does not exist a policy that leads to improved performance. We empirically evaluate our algorithm on a grid world and also on a set of domains from the DeepMind control suite. We confirm our theoretical results regarding the monotonically improving performance of our algorithm. Interestingly, we also show empirically that the sets of policies computed by the algorithm are diverse, leading to different trajectories in the grid world and very distinct locomotion skills in the control suite.

Complex machine learning models are deployed in several critical domains including healthcare and autonomous vehicles nowadays, albeit as functional black boxes. Consequently, there has been a recent surge in interpreting decisions of such complex models in order to explain their actions to humans. Models that correspond to human interpretation of a task are more desirable in certain contexts and can help attribute liability, build trust, expose biases and in turn build better models. It is, therefore, crucial to understand how and which models conform to human understanding of tasks. In this paper, we present a large-scale crowdsourcing study that reveals and quantifies the dissonance between human and machine understanding, through the lens of an image classification task. In particular, we seek to answer the following questions: Which (well-performing) complex ML models are closer to humans in their use of features to make accurate predictions? How does task difficulty affect the feature selection capability of machines in comparison to humans? Are humans consistently better at selecting features that make image recognition more accurate? Our findings have important implications on human-machine collaboration, considering that a long term goal in the field of artificial intelligence is to make machines capable of learning and reasoning like humans.

Knowledge Graph Embedding (KGE) is a popular method for KG reasoning and usually a higher dimensional one ensures better reasoning capability. However, high-dimensional KGEs pose huge challenges to storage and computing resources and are not suitable for resource-limited or time-constrained applications, for which faster and cheaper reasoning is necessary. To address this problem, we propose DistilE, a knowledge distillation method to build low-dimensional student KGE from pre-trained high-dimensional teacher KGE. We take the original KGE loss as hard label loss and design specific soft label loss for different KGEs in DistilE. We also propose a two-stage distillation approach to make the student and teacher adapt to each other and further improve the reasoning capability of the student. Our DistilE is general enough to be applied to various KGEs. Experimental results of link prediction show that our method successfully distills a good student which performs better than a same dimensional one directly trained, and sometimes even better than the teacher, and it can achieve 2 times - 8 times embedding compression rate and more than 10 times faster inference speed than the teacher with a small performance loss. We also experimentally prove the effectiveness of our two-stage training proposal via ablation study.

Traditionally Guideline(GL)based Decision Support Systems (DSSs) use a centralized infrastructure to generate recommendations to care providers. However, managing patients at home is preferable, reducing costs and empowering patients. We aimed to design, implement, and demonstrate the feasibility of a new architecture for a distributed DSS that provides patients with personalized, context-sensitive, evidence based guidance through their mobile device, and increases the robustness of the distributed application of the GL, while maintaining access to the patient longitudinal record and to an up to date evidence based GL repository. We have designed and implemented a novel projection and callback (PCB) model, in which small portions of the evidence based GL procedural knowledge, adapted to the patient preferences and to their current context, are projected from a central DSS server, to a local DSS on the patient mobile device that applies that knowledge. When appropriate, as defined by a temporal pattern within the projected plan, the local DSS calls back the central DSS, requesting further assistance, possibly another projection. Thus, the GL specification includes two levels: one for the central DSS, one for the local DSS. We successfully evaluated the PCB model within the MobiGuide EU project by managing Gestational Diabetes Mellitus patients in Spain, and Atrial Fibrillation patients in Italy. Significant differences exist between the two GL representations, suggesting additional ways to characterize GLs. Mean time between the central and local interactions was quite different for the two GLs: 3.95 days for gestational diabetes, 23.80 days for atrial fibrillation. Most interactions, 83%, were due to projections to the mDSS. Others were data notifications, mostly to change context. Robustness was demonstrated through successful recovery from multiple local DSS crashes.

Multiagent reinforcement learning (MARL) has achieved a remarkable amount of success in solving various types of video games. A cornerstone of this success is the auto-curriculum framework, which shapes the learning process by continually creating new challenging tasks for agents to adapt to, thereby facilitating the acquisition of new skills. In order to extend MARL methods to real-world domains outside of video games, we envision in this blue sky paper that maintaining a diversity-aware auto-curriculum is critical for successful MARL applications. Specifically, we argue that \emph{behavioural diversity} is a pivotal, yet under-explored, component for real-world multiagent learning systems, and that significant work remains in understanding how to design a diversity-aware auto-curriculum. We list four open challenges for auto-curriculum techniques, which we believe deserve more attention from this community. Towards validating our vision, we recommend modelling realistic interactive behaviours in autonomous driving as an important test bed, and recommend the SMARTS/ULTRA benchmark.

Statistical samples, in order to be representative, have to be drawn from a population in a random and unbiased way. Nevertheless, it is common practice in the field of model-based diagnosis to make estimations from (biased) best-first samples. One example is the computation of a few most probable possible fault explanations for a defective system and the use of these to assess which aspect of the system, if measured, would bring the highest information gain. In this work, we scrutinize whether these statistically not well-founded conventions, that both diagnosis researchers and practitioners have adhered to for decades, are indeed reasonable. To this end, we empirically analyze various sampling methods that generate fault explanations. We study the representativeness of the produced samples in terms of their estimations about fault explanations and how well they guide diagnostic decisions, and we investigate the impact of sample size, the optimal trade-off between sampling efficiency and effectivity, and how approximate sampling techniques compare to exact ones.