User acceptance of artificial intelligence agents might depend on their ability to explain their reasoning, which requires adding an interpretability layer that fa- cilitates users to understand their behavior. This paper focuses on adding an in- terpretable layer on top of Semantic Textual Similarity (STS), which measures the degree of semantic equivalence between two sentences. The interpretability layer is formalized as the alignment between pairs of segments across the two sentences, where the relation between the segments is labeled with a relation type and a similarity score. We present a publicly available dataset of sentence pairs annotated following the formalization. We then develop a system trained on this dataset which, given a sentence pair, explains what is similar and different, in the form of graded and typed segment alignments. When evaluated on the dataset, the system performs better than an informed baseline, showing that the dataset and task are well-defined and feasible. Most importantly, two user studies show how the system output can be used to automatically produce explanations in natural language. Users performed better when having access to the explanations, pro- viding preliminary evidence that our dataset and method to automatically produce explanations is useful in real applications.
We investigate a human-machine collaborative drawing environment in which an autonomous agent sketches images while optionally allowing a user to directly influence the agent's trajectory. We combine Monte Carlo Tree Search with image classifiers and test both shallow models (e.g. multinomial logistic regression) and deep Convolutional Neural Networks (e.g. LeNet, Inception v3). We found that using the shallow model, the agent produces a limited variety of images, which are noticably recogonisable by humans. However, using the deeper models, the agent produces a more diverse range of images, and while the agent remains very confident (99.99%) in having achieved its objective, to humans they mostly resemble unrecognisable 'random' noise. We relate this to recent research which also discovered that 'deep neural networks are easily fooled' \cite{Nguyen2015} and we discuss possible solutions and future directions for the research.
A prediction market is a useful means of aggregating information about a future event. To function, the market needs a trusted entity who will verify the true outcome in the end. Motivated by the recent introduction of decentralized prediction markets, we introduce a mechanism that allows for the outcome to be determined by the votes of a group of arbiters who may themselves hold stakes in the market. Despite the potential conflict of interest, we derive conditions under which we can incentivize arbiters to vote truthfully by using funds raised from market fees to implement a peer prediction mechanism. Finally, we investigate what parameter values could be used in a real-world implementation of our mechanism.
We demonstrate the possibility of classifying causal systems into kinds that share a common structure without first constructing an explicit dynamical model or using prior knowledge of the system dynamics. The algorithmic ability to determine whether arbitrary systems are governed by causal relations of the same form offers significant practical applications in the development and validation of dynamical models. It is also of theoretical interest as an essential stage in the scientific inference of laws from empirical data. The algorithm presented is based on the dynamical symmetry approach to dynamical kinds. A dynamical symmetry with respect to time is an intervention on one or more variables of a system that commutes with the time evolution of the system. A dynamical kind is a class of systems sharing a set of dynamical symmetries. The algorithm presented classifies deterministic, time-dependent causal systems by directly comparing their exhibited symmetries. Using simulated, noisy data from a variety of nonlinear systems, we show that this algorithm correctly sorts systems into dynamical kinds. It is robust under significant sampling error, is immune to violations of normality in sampling error, and fails gracefully with increasing dynamical similarity. The algorithm we demonstrate is the first to address this aspect of automated scientific discovery.
A good dialogue agent should have the ability to interact with users. In this work, we explore this direction by designing a simulator and a set of synthetic tasks in the movie domain that allow the learner to interact with a teacher by both asking and answering questions. We investigate how a learner can benefit from asking questions in both an offline and online reinforcement learning setting. We demonstrate that the learner improves when asking questions. Our work represents a first step in developing end-to-end learned interactive dialogue agents.
In this paper, we describe the construction of TeKnowbase, a knowledge-base of technical concepts in computer science. Our main information sources are technical websites such as Webopedia and Techtarget as well as Wikipedia and online textbooks. We divide the knowledge-base construction problem into two parts -- the acquisition of entities and the extraction of relationships among these entities. Our knowledge-base consists of approximately 100,000 triples. We conducted an evaluation on a sample of triples and report an accuracy of a little over 90\%. We additionally conducted classification experiments on StackOverflow data with features from TeKnowbase and achieved improved classification accuracy.
Ontohub is a repository engine for managing distributed heterogeneous ontologies. The distributed nature enables communities to share and exchange their contributions easily. The heterogeneous nature makes it possible to integrate ontologies written in various ontology languages. Ontohub supports a wide range of formal logical and ontology languages, as well as various structuring and modularity constructs and inter-theory (concept) mappings, building on the OMG-standardized DOL language. Ontohub repositories are organised as Git repositories, thus inheriting all features of this popular version control system. Moreover, Ontohub is the first repository engine meeting a substantial amount of the requirements formulated in the context of the Open Ontology Repository (OOR) initiative, including an API for federation as well as support for logical inference and axiom selection.
Bayesian inference on structured models typically relies on the ability to infer posterior distributions of underlying hidden variables. However, inference in implicit models or complex posterior distributions is hard. A popular tool for learning implicit models are generative adversarial networks (GANs) which learn parameters of generators by fooling discriminators. Typically, GANs are considered to be models themselves and are not understood in the context of inference. Current techniques rely on inefficient global discrimination of joint distributions to perform learning, or only consider discriminating a single output variable. We overcome these limitations by treating GANs as a basis for likelihood-free inference in generative models and generalize them to Bayesian posterior inference over factor graphs. We propose local learning rules based on message passing minimizing a global divergence criterion involving cooperating local adversaries used to sidestep explicit likelihood evaluations. This allows us to compose models and yields a unified inference and learning framework for adversarial learning. Our framework treats model specification and inference separately and facilitates richly structured models within the family of Directed Acyclic Graphs, including components such as intractable likelihoods, non-differentiable models, simulators and generally cumbersome models. A key result of our treatment is the insight that Bayesian inference on structured models can be performed only with sampling and discrimination when using nonparametric variational families, without access to explicit distributions. As a side-result, we discuss the link to likelihood maximization. These approaches hold promise to be useful in the toolbox of probabilistic modelers and enrich the gamut of current probabilistic programming applications.
In this paper, we propose a framework for solving a single-agent task by using multiple agents, each focusing on different aspects of the task. This approach has two main advantages: 1) it allows for specialized agents for different parts of the task, and 2) it provides a new way to transfer knowledge, by transferring trained agents. Our framework generalizes the traditional hierarchical decomposition, in which, at any moment in time, a single agent has control until it has solved its particular subtask. We illustrate our framework using a number of examples.
Existing models based on artificial neural networks (ANNs) for sentence classification often do not incorporate the context in which sentences appear, and classify sentences individually. However, traditional sentence classification approaches have been shown to greatly benefit from jointly classifying subsequent sentences, such as with conditional random fields. In this work, we present an ANN architecture that combines the effectiveness of typical ANN models to classify sentences in isolation, with the strength of structured prediction. Our model achieves state-of-the-art results on two different datasets for sequential sentence classification in medical abstracts.
In this paper we present an agent-based model (ABM) of scientific inquiry aimed at investigating how different social networks impact the efficiency of scientists in acquiring knowledge. As such, the ABM is a computational tool for tackling issues in the domain of scientific methodology and science policy. In contrast to existing ABMs of science, our model aims to represent the argumentative dynamics that underlies scientific practice. To this end we employ abstract argumentation theory as the core design feature of the model. This helps to avoid a number of problematic idealizations which are present in other ABMs of science and which impede their relevance for actual scientific practice.
Whereas CNNs have demonstrated immense progress in many vision problems, they suffer from a dependence on monumental amounts of labeled training data. On the other hand, dictionary learning does not scale to the size of problems that CNNs can handle, despite being very effective at low-level vision tasks such as denoising and inpainting. Recently, interest has grown in adapting dictionary learning methods for supervised tasks such as classification and inverse problems. We propose two new network layers that are based on dictionary learning: a sparse factorization layer and a convolutional sparse factorization layer, analogous to fully-connected and convolutional layers, respectively. Using our derivations, these layers can be dropped in to existing CNNs, trained together in an end-to-end fashion with back-propagation, and leverage semisupervision in ways classical CNNs cannot. We experimentally compare networks with these two new layers against a baseline CNN. Our results demonstrate that networks with either of the sparse factorization layers are able to outperform classical CNNs when supervised data are few. They also show performance improvements in certain tasks when compared to the CNN with no sparse factorization layers with the same exact number of parameters.
Assumption-Based Argumentation (ABA) is an argumentation framework that has been proposed in the late 20th century. Since then, there was still no solver implemented in a programming language which is easy to setup and no solver have been interfaced to the web, which impedes the interests of the public. This project aims to implement an ABA solver in a modern programming language that performs reasonably well and interface it to the web for easier access by the public. This project has demonstrated the novelty of development of an ABA solver, that computes conflict-free, stable, admissible, grounded, ideal, and complete semantics, in Python programming language which can be used via an easy-to-use web interface for visualization of the argument and dispute trees. Experiments were conducted to determine the project's best configurations and to compare this project with proxdd, a state-of-the-art ABA solver, which has no web interface and computes less number of semantics. From the results of the experiments, this project's best configuration is achieved by utilizing "pickle" technique and tree caching technique. Using this project's best configuration, this project achieved a lower average runtime compared to proxdd. On other aspect, this project encountered more cases with exceptions compared to proxdd, which might be caused by this project computing more semantics and hence requires more resources to do so. Hence, it can be said that this project run comparably well to the state-of-the-art ABA solver proxdd. Future works of this project include computational complexity analysis and efficiency analysis of algorithms implemented, implementation of more semantics in argumentation framework, and usability testing of the web interface.
Recurrent Neural Networks (RNN), particularly Long Short Term Memory (LSTM) RNNs, are a popular and very successful method for learning and generating sequences. However, current generative RNN techniques do not allow real-time interactive control of the sequence generation process, thus aren't well suited for live creative expression. We propose a method of real-time continuous control and 'steering' of sequence generation using an ensemble of RNNs and dynamically altering the mixture weights of the models. We demonstrate the method using character based LSTM networks and a gestural interface allowing users to 'conduct' the generation of text.
We introduce a method for imposing higher-level structure on generated, polyphonic music. A Convolutional Restricted Boltzmann Machine (C-RBM) as a generative model is combined with gradient descent constraint optimization to provide further control over the generation process. Among other things, this allows for the use of a "template" piece, from which some structural properties can be extracted, and transferred as constraints to newly generated material. The sampling process is guided with Simulated Annealing in order to avoid local optima, and find solutions that both satisfy the constraints, and are relatively stable with respect to the C-RBM. Results show that with this approach it is possible to control the higher level self-similarity structure, the meter, as well as tonal properties of the resulting musical piece while preserving its local musical coherence.
Deep models are the defacto standard in visual decision models due to their impressive performance on a wide array of visual tasks. However, they are frequently seen as opaque and are unable to explain their decisions. In contrast, humans can justify their decisions with natural language and point to the evidence in the visual world which led to their decisions. We postulate that deep models can do this as well and propose our Pointing and Justification (PJ-X) model which can justify its decision with a sentence and point to the evidence by introspecting its decision and explanation process using an attention mechanism. Unfortunately there is no dataset available with reference explanations for visual decision making. We thus collect two datasets in two domains where it is interesting and challenging to explain decisions. First, we extend the visual question answering task to not only provide an answer but also a natural language explanation for the answer. Second, we focus on explaining human activities which is traditionally more challenging than object classification. We extensively evaluate our PJ-X model, both on the justification and pointing tasks, by comparing it to prior models and ablations using both automatic and human evaluations.
This paper introduces the probabilistic module interface, which allows encapsulation of complex probabilistic models with latent variables alongside custom stochastic approximate inference machinery, and provides a platform-agnostic abstraction barrier separating the model internals from the host probabilistic inference system. The interface can be seen as a stochastic generalization of a standard simulation and density interface for probabilistic primitives. We show that sound approximate inference algorithms can be constructed for networks of probabilistic modules, and we demonstrate that the interface can be implemented using learned stochastic inference networks and MCMC and SMC approximate inference programs.
In many model-based diagnosis applications it is impossible to provide such a set of observations and/or measurements that allow to identify the real cause of a fault. Therefore, diagnosis systems often return many possible candidates, leaving the burden of selecting the correct diagnosis to a user. Sequential diagnosis techniques solve this problem by automatically generating a sequence of queries to some oracle. The answers to these queries provide additional information necessary to gradually restrict the search space by removing diagnosis candidates inconsistent with the answers.
During query computation, existing sequential diagnosis methods often require the generation of many unnecessary query candidates and strongly rely on expensive logical reasoners. We tackle this issue by devising efficient heuristic query search methods. The proposed methods enable for the first time a completely reasoner-free query generation while at the same time guaranteeing optimality conditions, e.g. minimal cardinality or best understandability, of the returned query that existing methods cannot realize. Hence, the performance of this approach is independent of the (complexity of the) diagnosed system. Experiments conducted using real-world problems show that the new approach is highly scalable and outperforms existing methods by orders of magnitude.
The rapid growth in stored time-oriented data necessitates the development of new methods for handling, processing, and interpreting large amounts of temporal data. One important example of such processing is detecting anomalies in time-oriented data. The Knowledge-Based Temporal Abstraction method was previously proposed for intelligent interpretation of temporal data based on predefined domain knowledge. In this study we propose a framework that integrates the KBTA method with a temporal pattern mining process for anomaly detection. According to the proposed method a temporal pattern mining process is applied on a dataset of basic temporal abstraction database in order to extract patterns representing normal behavior. These patterns are then analyzed in order to identify abnormal time periods characterized by a significantly small number of normal patterns. The proposed approach was demonstrated using a dataset collected from a real server.
Robotic challenges like the Amazon Picking Challenge (APC) or the DARPA Challenges are an established and important way to drive scientific progress. They make research comparable on a well-defined benchmark with equal test conditions for all participants. However, such challenge events occur only occasionally, are limited to a small number of contestants, and the test conditions are very difficult to replicate after the main event. We present a new physical benchmark challenge for robotic picking: the ACRV Picking Benchmark (APB). Designed to be reproducible, it consists of a set of 42 common objects, a widely available shelf, and exact guidelines for object arrangement using stencils. A well-defined evaluation protocol enables the comparison of \emph{complete} robotic systems -- including perception and manipulation -- instead of sub-systems only. Our paper also describes and reports results achieved by an open baseline system based on a Baxter robot.
We propose a novel method of regularization for recurrent neural networks called suprisal-driven zoneout. In this method, states zoneout (maintain their previous value rather than updating), when the suprisal (discrepancy between the last state's prediction and target) is small. Thus regularization is adaptive and input-driven on a per-neuron basis. We demonstrate the effectiveness of this idea by achieving state-of-the-art bits per character of 1.31 on the Hutter Prize Wikipedia dataset, significantly reducing the gap to the best known highly-engineered compression methods.
Recent advances have shown the capability of Fully Convolutional Neural Networks (FCN) to model cost functions for motion planning in the context of learning driving preferences purely based on demonstration data from human drivers. While pure learning from demonstrations in the framework of Inverse Reinforcement Learning (IRL) is a promising approach, we can benefit from well informed human priors and incorporate them into the learning process. Our work achieves this by pretraining a model to regress to a manual cost function and refining it based on Maximum Entropy Deep Inverse Reinforcement Learning. When injecting prior knowledge as pretraining for the network, we achieve higher robustness, more visually distinct obstacle boundaries, and the ability to capture instances of obstacles that elude models that purely learn from demonstration data. Furthermore, by exploiting these human priors, the resulting model can more accurately handle corner cases that are scarcely seen in the demonstration data, such as stairs, slopes, and underpasses.
We propose an online, end-to-end, deep reinforcement learning technique to develop generative conversational agents for open-domain dialogue. We use a unique combination of offline two-phase supervised learning and online reinforcement learning with human users to train our agent. While most existing research proposes hand-crafted and develop-defined reward functions for reinforcement, we devise a novel reward mechanism based on a variant of Beam Search and one-character user-feedback at each step. Experiments show that our model, when trained on a small and shallow Seq2Seq network, successfully promotes the generation of meaningful, diverse and interesting responses, and can be used to train agents with customized personas and conversational styles.
A key drawback of the current generation of artificial decision-makers is that they do not adapt well to changes in unexpected situations. This paper addresses the situation in which an AI for aerial dog fighting, with tunable parameters that govern its behavior, will optimize behavior with respect to an objective function that must be evaluated and learned through simulations. Once this objective function has been modeled, the agent can then choose its desired behavior in different situations. Bayesian optimization with a Gaussian Process surrogate is used as the method for investigating the objective function. One key benefit is that during optimization the Gaussian Process learns a global estimate of the true objective function, with predicted outcomes and a statistical measure of confidence in areas that haven't been investigated yet. However, standard Bayesian optimization does not perform consistently or provide an accurate Gaussian Process surrogate function for highly volatile objective functions. We treat these problems by introducing a novel sampling technique called Hybrid Repeat/Multi-point Sampling. This technique gives the AI ability to learn optimum behaviors in a highly uncertain environment. More importantly, it not only improves the reliability of the optimization, but also creates a better model of the entire objective surface. With this improved model the agent is equipped to better adapt behaviors.
