Traditionally, the way one evaluates the performance of an Artificial Intelligence (AI) system is via a comparison to human performance in specific tasks, treating humans as a reference for high-level cognition. However, these comparisons leave out important features of human intelligence: the capability to transfer knowledge and make complex decisions based on emotional and rational reasoning. These decisions are influenced by current inferences as well as prior experiences, making the decision process strongly subjective and apparently biased. In this context, a definition of compositional intelligence is necessary to incorporate these features in future AI tests. Here, a concrete implementation of this will be suggested, using recent developments in quantum cognition, natural language and compositional meaning of sentences, thanks to categorical compositional models of meaning.

The ability to infer the intentions of others, predict their goals, and deduce their plans are critical features for intelligent agents. For a long time, several approaches investigated the use of symbolic representations and inferences with limited success, principally because it is difficult to capture the cognitive knowledge behind human decisions explicitly. The trend, nowadays, is increasingly focusing on learning to infer intentions directly from data, using deep learning in particular. We are now observing interesting applications of intent classification in natural language processing, visual activity recognition, and emerging approaches in other domains. This paper discusses a novel approach combining few-shot and transfer learning with cross-domain features, to learn to infer the intent of an agent navigating in physical environments, executing arbitrary long sequences of actions to achieve their goals. Experiments in synthetic environments demonstrate improved performance in terms of learning from few samples and generalizing to unseen configurations, compared to a deep-learning baseline approach.

In many settings, as for example wind farms, multiple machines are instantiated to perform the same task, which is called a fleet. The recent advances with respect to the Internet of Things allow control devices and/or machines to connect through cloud-based architectures in order to share information about their status and environment. Such an infrastructure allows seamless data sharing between fleet members, which could greatly improve the sample-efficiency of reinforcement learning techniques. However in practice, these machines, while almost identical in design, have small discrepancies due to production errors or degradation, preventing control algorithms to simply aggregate and employ all fleet data. We propose a novel reinforcement learning method that learns to transfer knowledge between similar fleet members and creates member-specific dynamics models for control. Our algorithm uses Gaussian processes to establish cross-member covariances. This is significantly different from standard transfer learning methods, as the focus is not on sharing information over tasks, but rather over system specifications. We demonstrate our approach on two benchmarks and a realistic wind farm setting. Our method significantly outperforms two baseline approaches, namely individual learning and joint learning where all samples are aggregated, in terms of the median and variance of the results.

Artificial Intelligence (AI) has become an integral part of domains such as security, finance, healthcare, medicine, and criminal justice. Explaining the decisions of AI systems in human terms is a key challenge--due to the high complexity of the model, as well as the potential implications on human interests, rights, and lives . While Explainable AI is an emerging field of research, there is no consensus on the definition, quantification, and formalization of explainability. In fact, the quantification of explainability is an open challenge. In our previous work, we incorporated domain knowledge for better explainability, however, we were unable to quantify the extent of explainability. In this work, we (1) briefly analyze the definitions of explainability from the perspective of different disciplines (e.g., psychology, social science), properties of explanation, explanation methods, and human-friendly explanations; and (2) propose and formulate an approach to quantify the extent of explainability. Our experimental result suggests a reasonable and model-agnostic way to quantify explainability

The process of identifying and understanding art styles to discover artistic influences is essential to the study of art history. Traditionally, trained experts review fine details of the works and compare them to other known works. To automate and scale this task, we use several state-of-the-art CNN architectures to explore how a machine may help perceive and quantify art styles. This study explores: (1) How accurately can a machine classify art styles? (2) What may be the underlying relationships among different styles and artists? To help answer the first question, our best-performing model using Inception V3 achieves a 9-class classification accuracy of 88.35%, which outperforms the model in Elgammal et al.'s study by more than 20 percent. Visualizations using Grad-CAM heat maps confirm that the model correctly focuses on the characteristic parts of paintings. To help address the second question, we conduct network analysis on the influences among styles and artists by extracting 512 features from the best-performing classification model. Through 2D and 3D T-SNE visualizations, we observe clear chronological patterns of development and separation among the art styles. The network analysis also appears to show anticipated artist level connections from an art historical perspective. This technique appears to help identify some previously unknown linkages that may shed light upon new directions for further exploration by art historians. We hope that humans and machines working in concert may bring new opportunities to the field.

The ability to observe the effects of actions performed by others and to infer their intent, most likely goals, or course of action, is known as a plan or intention recognition cognitive capability and has long been one of the fundamental research challenges in AI. Deep learning has recently been making significant inroads on various pattern recognition problems, except for intention recognition. While extensively explored since the seventies, the problem remains unsolved for most interesting cases in various areas, ranging from natural language understanding to human behavior understanding based on video feeds. This paper compares symbolic inverse planning, one of the most investigated approaches to goal recognition, to deep learning using CNN and LTSM neural network architectures, on five synthetic benchmarks often used in the literature. The results show that the deep learning approach achieves better goal-prediction accuracy and timeliness than the symbolic cost-based plan recognizer in these domains. Although preliminary, these results point to interesting future research avenues.

Human decision-makers often receive assistance from data-driven algorithmic systems that provide a score for evaluating objects, including individuals. The scores are generated by a function (mechanism) that takes a set of features as input and generates a score.The scoring functions are either machine-learned or human-designed and can be used for different decision purposes such as ranking or classification. Given the potential impact of these scoring mechanisms on individuals' lives and on society, it is important to make sure these scores are computed responsibly. Hence we need tools for responsible scoring mechanism design. In this paper, focusing on linear scoring functions, we highlight the importance of unbiased function sampling and perturbation in the function space for devising such tools. We provide unbiased samplers for the entire function space, as well as a $\theta$-vicinity around a given function. We then illustrate the value of these samplers for designing effective algorithms in three diverse problem scenarios in the context of ranking. Finally, as a fundamental method for designing responsible scoring mechanisms, we propose a novel approach for approximating the construction of the arrangement of hyperplanes. Despite the exponential complexity of an arrangement in the number of dimensions, using function sampling, our algorithm is linear in the number of samples and hyperplanes, and independent of the number of dimensions.

Anaphora resolution is a challenging task which has been the interest of NLP researchers for a long time. Traditional resolution techniques like eliminative constraints and weighted preferences were successful in many languages. However, they are ineffective in free word order languages like most SouthAsian languages.Heuristic and rule-based techniques were typical in these languages, which are constrained to context and domain.In this paper, we venture a new strategy us-ing neural networks for resolving anaphora in human-human dialogues. The architecture chiefly consists of three components, a shallow parser for extracting features, a feature vector generator which produces the word embed-dings, and a neural network model which will predict the antecedent mention of an anaphora.The system has been trained and tested on Telugu conversation corpus we generated. Given the advantage of the semantic information in word embeddings and appending actor, gender, number, person and part of plural features the model has reached an F1-score of 86.

Recent studies have revealed that neural network-based policies can be easily fooled by adversarial examples. However, while most prior works analyze the effects of perturbing every pixel of every frame assuming white-box policy access, in this paper we take a more restrictive view towards adversary generation - with the goal of unveiling the limits of a model's vulnerability. In particular, we explore minimalistic attacks by defining three key settings: (1) black-box policy access: where the attacker only has access to the input (state) and output (action probability) of an RL policy; (2) fractional-state adversary: where only several pixels are perturbed, with the extreme case being a single-pixel adversary; and (3) tactically-chanced attack: where only significant frames are tactically chosen to be attacked. We formulate the adversarial attack by accommodating the three key settings and explore their potency on six Atari games by examining four fully trained state-of-the-art policies. In Breakout, for example, we surprisingly find that: (i) all policies showcase significant performance degradation by merely modifying 0.01% of the input state, and (ii) the policy trained by DQN is totally deceived by perturbation to only 1% frames.

Most machine learning theory and practice is concerned with learning a single task. In this thesis it is argued that in general there is insufficient information in a single task for a learner to generalise well and that what is required for good generalisation is information about many similar learning tasks. Similar learning tasks form a body of prior information that can be used to constrain the learner and make it generalise better. Examples of learning scenarios in which there are many similar tasks are handwritten character recognition and spoken word recognition. The concept of the environment of a learner is introduced as a probability measure over the set of learning problems the learner might be expected to learn. It is shown how a sample from the environment may be used to learn a representation, or recoding of the input space that is appropriate for the environment. Learning a representation can equivalently be thought of as learning the appropriate features of the environment. Bounds are derived on the sample size required to ensure good generalisation from a representation learning process. These bounds show that under certain circumstances learning a representation appropriate for $n$ tasks reduces the number of examples required of each task by a factor of $n$. Once a representation is learnt it can be used to learn novel tasks from the same environment, with the result that far fewer examples are required of the new tasks to ensure good generalisation. Bounds are given on the number of tasks and the number of samples from each task required to ensure that a representation will be a good one for learning novel tasks. The results on representation learning are generalised to cover any form of automated hypothesis space bias.