We present DocInfer -a novel, end-to-end Document-level Natural Language Inference model that builds a hierarchical document graph enriched through inter-sentence relations (topical, entity-based, concept-based), performs paragraph pruning using the novel SubGraph Pooling layer, followed by optimal evidence selection based on REINFORCE algorithm to identify the most important context sentences for a given hypothesis.Our evidence selection mechanism allows it to transcend the input length limitation of modern BERT-like Transformer models while presenting the entire evidence together for inferential reasoning.We show this is an important property needed to reason on large documents where the evidence may be fragmented and located arbitrarily far from each other.Extensive experiments on popular corpora -DocNLI, ContractNLI, and ConTRoL datasets, and our new proposed dataset called CaseHoldNLI on the task of legal judicial reasoning, demonstrate significant performance gains of 8-12% over SOTA methods.Our ablation studies validate the impact of our model.Performance improvement of ∼ 3 -6% on annotation-scarce downstream tasks of fact verification, multiple-choice QA, and contract clause retrieval demonstrates the usefulness of DocInfer beyond primary NLI tasks.

Counting the number of true instances of a clause is arguably a major bottleneck in relational probabilistic inference and learning. We approximate counts in two steps: (1) transform the fully grounded relational model to a large hypergraph, and partially-instantiated clauses to hypergraph motifs; (2) since the expected counts of the motifs are provably the clause counts, approximate them using summary statistics (in/outdegrees, edge counts, etc). Our experimental results demonstrate the efficiency of these approximations, which can be applied to many complex statistical relational models, and can be significantly faster than state-of-the-art, both for inference and learning, without sacrificing effectiveness.

We consider the problem of actively feature elicitation in which given a few examples with all the features (say the full EHR) and a few examples with some of the features (say demographics), the goal is to identify the set of examples on whom more information (say the lab tests) needs to be collected. The observation is that some set of features may be more expensive, personal or cumbersome to collect. We propose an active learning approach which identifies examples that are dissimilar to the ones with the full set of data and acquire the complete set of features for these examples. Motivated by real clinical tasks, our extensive evaluation on three clinical tasks demonstrate the effectiveness of this approach.

We consider the problem of learning probabilistic models from relational data. One of the key issues with relational data is class imbalance where the number of negative examples far outnumbers the number of positive examples. The common approach for dealing with this problem is the use of sub-sampling of negative examples. We, on the other hand, consider a soft margin approach that explicitly trades off between the false positives and false negatives. We apply this approach to the recently successful formalism of relational functional gradient boosting. Specifically, we modify the objective function of the learning problem to explicitly include the trade-off between false positives and negatives. We show empirically that this approach is more successful in handling the class imbalance problem than the original framework that weighed all the examples equally.

Improvements in medical imaging techniques have provided clinicians the ability to obtain detailed brain images of patients at lower costs. This increased availability of rich data opens up new avenues of research that promise better understanding of common brain ailments such as Alzheimer's Disease and dementia. Improved data mining techniques, however, are required to leverage these new data sets to identify intermediate disease states (e.g., mild cognitive impairment) and perform early diagnosis. We propose a graphical model framework based on conditional random fields (CRFs) to mine MRI brain images. As a proof-of-concept, we apply CRFs to the problem of brain tissue segmentation. Experimental results show robust and accurate performance on tissue segmentation comparable to other state-of-the-art segmentation methods. In addition, results show that our algorithm generalizes well across data sets and is less susceptible to outliers. Our method relies on minimal prior knowledge unlike atlas-based techniques, which assume images map to a normal template. Our results show that CRFs are a promising model for tissue segmentation, as well as other MRI data mining problems such as anatomical segmentation and disease diagnosis where atlas assumptions are unreliable in abnormal brain images.

Inverse Reinforcement Learning (IRL) is an approach for domain-reward discovery from demonstration, where an agent mines the reward function of a Markov decision process by observing an expert acting in the domain. In the standard setting, it is assumed that the expert acts (nearly) optimally, and a large number of trajectories, i.e., training examples are available for reward discovery (and consequently, learning domain behavior). These are not practical assumptions: trajectories are often noisy, and there can be a paucity of examples. Our novel approach incorporates advice-giving into the IRL framework to address these issues. Inspired by preference elicitation, a domain expert provides advice on states and actions (features) by stating preferences over them. We evaluate our approach on several domains and show that with small amounts of targeted preference advice, learning is possible from noisy demonstrations, and requires far fewer trajectories compared to simply learning from trajectories alone.

Some supervised-learning algorithms can make effective use of domain knowledge in addition to the input-output pairs commonly used in machine learning. However, formulating this additional information often requires an in-depth understanding of the specific knowledge representation used by a given learning algorithm. The requirement to use a formal knowledge-representation language means that most domain experts will not be able to articulate their expertise, even when a learning algorithm is capable of exploiting such valuable information. We investigate a method to ease this knowledge acquisition through the use of a graphical, human-computer interface. Our interface allows users to easily provide advice about specific examples, rather than requiring them to provide general rules; we leave the task of properly generalizing such advice to the learning algorithms. We demonstrate the effectiveness of our approach using the Wargus real-time strategy game, comparing learning with no advice to learning with concrete advice provided through our interface, as well as comparing to using generalized advice written by an AI expert. Our results show that our approach of combining a GUI-based advice language with an advice-taking learning algorithm is an effective way to capture domain knowledge.

Knowledge-based support vector machines (KBSVMs) incorporate advice from domain experts, which can improve generalization significantly. A major limitation that has not been fully addressed occurs when the expert advice is imperfect, which can lead to poorer models. We propose a model that extends KBSVMs and is able to not only learn from data and advice, but also simultaneously improve the advice. The proposed approach is particularly effective for knowledge discovery in domains with few labeled examples. The proposed model contains bilinear constraints, and is solved using two iterative approaches: successive linear programming and a constrained concave-convex approach. Experimental results demonstrate that these algorithms yield useful refinements to expert advice, as well as improve the performance of the learning algorithm overall.

Learning the reward function of an agent by observing its behavior is termed inverse reinforcement learning and has applications in learning from demonstration or apprenticeship learning. We introduce the problem of multi-agent inverse reinforcement learning, where reward functions of multiple agents are learned by observing their uncoordinated behavior. A centralized controller then learns to coordinate their behavior by optimizing a weighted sum of reward functions of all the agents. We evaluate our approach on a traffic-routing domain, in which a controller coordinates actions of multiple traffic signals to regulate traffic density. We show that the learner is not only able to match but even significantly outperform the expert.

Languages that combine predicate logic with probabilities are needed to succinctly represent knowledge in many real-world domains. We consider a formalism based on universally quantified conditional influence statements that capture local interactions between object attributes. The effects of different conditional influence statements can be combined using rules such as Noisy-OR. To combine multiple instantiations of the same rule we need other combining rules at a lower level. In this paper we derive and implement algorithms based on gradient-descent and EM for learning the parameters of these multi-level combining rules. We compare our approaches to learning in Markov Logic Networks and show superior performance in multiple domains.

A key step in many statistical learning methods used in machine learning involves solving a convex optimization problem containing one or more hyper-parameters that must be selected by the users. While cross validation is a commonly employed and widely accepted method for selecting these parameters, its implementation by a grid-search procedure in the parameter space effectively limits the desirable number of hyper-parameters in a model, due to the combinatorial explosion of grid points in high dimensions. This paper proposes a novel bilevel optimization approach to cross validation that provides a systematic search of the hyper-parameters. The bilevel approach enables the use of the state-of-the-art optimization methods and their well-supported softwares. After introducing the bilevel programming approach, we discuss computational methods for solving a bilevel cross-validation program, and present numerical results to substantiate the viability of this novel approach as a promising computational tool for model selection in machine learning.