loading subjects...
Ebook: Artificial Intelligence Research and Development
loading...
It is almost impossible today to find an economic sector or aspect of society which does not involve AI techniques in some way. This pervasive technology has become indispensible in a multitude of ways, from supporting decision making to managing digital devices such as smart sensors, mechanical arms or artificial eyes. The ability of AI to emulate intelligence in the resolution of challenging problems has placed it at the centre of problem solving in all areas of our society.
This book presents contributions from CCIA 2018, the 21st International Conference of the Catalan Association for Artificial Intelligence which took place in Alt Empordà, Catalonia, Spain, on 8-10th October 2018. The book aims to provide a picture of what is being achieved and what is under development in AI today. As such, its contents represent the diversity of approaches and applications currently being researched, but it also presents invited contributions which deal with some of the challenges that will have to be faced in the decade to come. The contributions included in this book are organized under the following headings: logic, satisfiability and fuzzy sets; classifiers, networks and machine learning; data science, recommender systems and case-based reasoning; natural language and sound processing; cognitive systems and agents; and computer vision and robotics. The book also covers a number of current AI challenges and new trends like big data, spatial problem solving, ethics and AI, and how blockchain impacts AI.
Providing an up-to-the-minute overview of current AI technology and research, this book will be of value to all those with an interest in the subject.
Nowadays, AI techniques are pervading all sectors of economy and society for supporting decision making, for identification of customer segments, for diagnosis generation, for finding relationships between causes and consequences, for preventing undesired situations (e.g. breakdowns, illnesses, etc.), for managing digital devices such as smart sensors, mechanical arms or artificial eyes, and so on. The transversality of AI and its ability to emulate intelligence in the resolution of challenging problems has currently placed AI at the focus of problem solving in all areas in our society.
This book aims at providing a current image of what is being achieved and what is under development at the moment in AI. As such, its contents are representative of the diversity of approaches and applications which are currently being researched. But it also presents invited contributions which deal with some of the challenges that must be explored in the next decade. Specifically, the research works included in this book can be categorized into the following areas: logic, satisfiability and fuzzy sets; classifiers, networks and machine learning; data science, recommender systems, case-based reasoning; natural language and sound processing; cognitive systems and agents; and computer vision and robotics. It also presents current AI challenges and new trends like big data, spatial problem solving, ethics and AI, and how blockchain impacts AI.
The section on Logic, satisfiability and fuzzy sets presents works on: clause branching for improving MaxSAT and MinSAT solvers; using decimation to improve solvers in large distributed constraint optimization problems; applying horn clauses for classifying and generating explanations; and defining a lattice structure for the set of interval-valued fuzzy qualitative labels.
The section Classifiers, networks and machine learning includes applications on: predicting people's next location; finding the optimal sensor placement in big-size water distribution systems; forecasting wind time series; sorting hams; classifying visitors in a tourist attraction; quantifying similarities between medical drugs; developing a universal encoder for time series; determining modern versus classic style in fashion; capturing equivariant features (related to illumination and rotation) in a new feature representation for deep learning; emulating brain prediction as a set of neural networks that perform Bayesian inference; and detecting shock fronts in computational fluid mechanics to find structures in computational astrophysics plasma simulations.
The following section on Data science, recommender systems, and case-based reasoning presents advances on: obesity prediction; including the context in playlist music recommendations; discovering the most probable sequences of web domains visited by people as input to marketing campaign design; supporting responsible repetitive research and innovation in health science using a case-based reasoning tool; developing personal menu planners depending on people's allergies, intolerances, preferences or diets; integrating semantic criteria into a recommender system; labeling feature-trees for case-base maintenance; and on analysing the utility and risk of masked data applied in big data algorithms.
Next section on Natural language and sound processing gathers studies on: visualizing and analysing sentiments in tweets using manifold dimensionality reduction; detecting irony in Twitter using convolutional neural networks; finding out if Twitter messages impact in stock liquidity; analysing agreements in Reddit-database comments using an argumentation approach; enhancing speech using multivariate empirical mode decomposition; classifying and generalizing parameter combinations in sound design; computing a distance for WordNet concepts based on visual features learnt from ImageNet; and on detecting and recognizing text in images using a language model and visual context information.
The section Cognitive systems and agents joins works which research on: using wearable sensors to detect workload on driving simulated scenarios; introducing the notion of trust in the form of mutual agreements between agents that might enhance stability in the formation of conventions such as turn-taking; developing a framework for organizations to become cognitive systems using a knowledge management process; determining the factors that can accelerate the emergence of collaborative behaviours among independent selfish learning agents (i.e. loss aversion); developing and evaluating an architecture for sentient embodied conversational agents with proactive and sensitive behaviours in educational applications; and on how intelligence can be defined as a generic statistical mechanics theory.
The section on Computer vision and robotics presents research works on: segmenting brain magnetic resonance images using multiphase active contours; estimating vehicle 3D pose from single monocular RGB images using a new architecture, G-Net, based on convolutional neural networks; classifying food attributes (cuisine origin and flavor) using a multi-scale convolutional network; improving retinal image segmentation based on conditional generative adversarial networks to obtain the optic disc; detecting foreground in multi-target fish tracking videos using fully connected convolutional neural networks; improving breast density segmentation using conditional generative adversarial networks; adapting robotic manipulation for industrial environments where teams are built by humans and robots; dealing with the anchoring problem (relating sensed data to symbolic information) in robotic container/bin picking challenges; and analysing human walking and gait behaviour from the interaction between the i-Walker automated rollator and its users.
The invited contribution by Ricardo Baeza-Yates (CTO of NTENT, California, USA, and UPF, Catalonia) on big data technologies outlines that the challenges to face by the companies are many (transparency, explainability, ethics, privacy, etc.) and questions the real need to indiscriminately deal with bigdata, proposing a more critical approach where identifying the right data to analyze is more relevant that the datasize itself. The invited contribution by Carlos Mérida (Accenture-Financial Services, Barcelona, Catalonia, Spain) on blockchain presents this emerging technology as a multi-agent system for data and process managing in a decentralised and secured manner, while analysing its opportunities in AI. The invited contribution by Christian Freksa (University of Bremen, Bremen Spatial Cognition Centre, Germany) on spatial problem solving explains that spatial problems (i.e. untangling cables, finding routes, solving puzzles, etc.) can be solved by cognitive agents either: (i) directly in space by means of perception and manipulation or (ii) they can be transformed into abstract representations to be solved by computational reasoning and then being transformed back into physical space; and it proposes “mild abstraction” as a way of combining best features of both solving methods. And the invited contribution by Luc Steels (ICREA/Institut de Biologia Evolutiva, UPF-CSIC, Catalonia) on the ethical use of AI provides ideas to foster positive uses of AI (i.e. improve healthcare, aid environmental monitoring, support community formation, make culture more widely accessible, etc.) and guard us against negative ones (i.e. perpetuation of racial and gender bias in recruitment, manipulation of public opinion, enforcing senseless bureaucracy, cyberattacks, autonomous weapons, etc.) building further on the “Barcelona Declaration for the Proper Development and Use of AI in Europe”.
Barcelona Declaration: http://www.iiia.csic.es/barcelonadeclaration/
This book is the outcome of the 21st edition of the International Conference of the Catalan Association for Artificial Intelligence (CCIA 2018
CCIA'18: https://ccia2018.upc.edu/en
This year, the Catalan Association for Artificial Intelligence (ACIA
ACIA: https://www.acia.cat/ EurAI: https://www.eurai.org/
The chairs of 21st edition of CCIA would like to express our sincere gratitude to the authors for their contributions to this book, to the invited authors for their enlightening contributions, and to all members of the Program and Organizing Committees who have worked hard to make CCIA'18 a success. We also thank the support of a non-neglectable number of scientific program members and authors who were formed in the Catalan AI community and are now developing their scientific activities in international research centers all over the world: Germany, Italy, Venezuela, Brazil, UK, Australia, France, Switzerland, Czech Republic, Sweden and Andorra. Last but not least, we would like to thank Josep Pujol and Aïda Valls, ACIA president and vicepresident, respectively, and all ACIA main board members for their kind support organizing this CCIA'18 were we will also commemorate the 25th anniversary of ACIA.
We wish all participants a successful and inspiring conference and a pleasant stay in Roses.
Zoe Falomir, University of Bremen, Germany (U. Bremen)
Karina Gibert, Universitat Politècnica de Catalunya – BarcelonaTech (UPC)
Enric Plaza, IIIA-CSIC
Big data is trendy, but there are many possible interpretations of its real impact, as well as the opportunities, risks and technological challenges. We will start with two key questions: can a company use big data? If so, should it? The opportunities are clear, while the challenges are many, including scalability, bias, and privacy on the problem side, as well as transparency, explainability, and ethics on the machine learning side. So we perform an analysis that includes all the data pipeline process. At the end we will conclude that what is important is the right data, not big data. In fact, the real challenge today, is machine learning for small data.
Everyday spatial problems such as tying shoelaces, untangling cables, locating objects, opening doors, matching shapes, finding routes, solving puzzles, … are embedded in 3-dimensional physical space and their solutions need to be manifested in physical space. Such problems can be solved either directly in space by means of perception and manipulation or they can be transformed into abstract representations in order to be solved by reasoning or computation and subsequently transformed back into physical space. In my contribution I will compare the two approaches and I will discuss how they can be combined. I will introduce the notion of mild abstraction as a way of combining the best features of both worlds, present a variety of forms of mild abstraction, demonstrate their uses for spatial problem solving, and propose that mild abstraction can be exploited for non-spatial domains, as well.
Blockchain is a technology to manage data and processes in a decentralised and secured manner. Its characteristics make it an interesting choice for developing disintermediated systems of assets transfers or data sharing. In the last four years, this technology has experienced a big impulse, with thousands of startups being created, open source frameworks developed and significant investments made by big companies.
All the blockchain developments are happening at the same time that artificial intelligence is becoming mainstream in the industry. Despite Blockchain is at a different point in the adoption path than some artificial intelligence applications, it is possible to see confluence points between them from two main perspectives: blockchain can be seen as a multi-agent system in itself and from a different perspective, it can be the necessary infrastructure for other autonomous systems to be audited.
Is Blockchain a hype? Or is it mature? What are the industry expectations about this technology? How AI relates to it? These questions will be addressed in the talk where several real use cases and industry feedback will be shared.
It is now clear to even the greatest doubters that AI is a powerful technology with considerable practical usage. But it has also become clear that this usage can be both positive and negative. AI can improve healthcare, aid environmental monitoring, support community formation, make culture more widely accessible, and do many other good things – always in combination of course with other digital technologies. But AI can also play a role in deceit and manipulation of public opinion, stoking societal divisions, enforcing senseless bureaucracy, cyberattacks, autonomous weapons, job loss, etc. Moreover the application and promises of AI appear to be premature in the face of autonomous cars going through red traffic lights, unacceptable decisions on granting parole, perpetuation of racial and gender bias in recruitment, etc. This discussion paper provides some ideas to foster positive uses of AI and guard us against negative ones, building further on the ‘Barcelona Declaration for the Proper Development and Use of AI in Europe’, which you can read and sign here: http://www.iiia.csic.es/barcelonadeclaration/.
We describe a new exact procedure, inspired on recent works on tableaux, for simultaneously solving MaxSAT and MinSAT. The main novelty of our procedure is that it implements a clause branching rule that preserves both the minimum and the maximum number of clauses than can be unsatisfied. Moreover, we prove the correctness of the procedure, and provide a review of the different types of branchings which are available for MaxSAT and MinSAT. We also give some preliminary experimental results that show that the new branching rule, when implemented into an existing branch-and-bound MaxSAT solver, produces substantial gains on industrial instances of the MaxSAT Evaluation.
When solving large distributed constraint optimization problems (DCOP), belief-propagation and incomplete inference algorithms are candidates of choice. However, these methods perform poorly when the factor graph is very loopy (i.e. cyclic) because of non-convergence and high communication costs. As to improving performances of the Max-Sum inference algorithm when solving loopy constraint optimization problems, we propose to take inspiration from the belief-propagation-guided decimation used to solve sparse random graphs (k-satisfiability). We introduce the DECIMAXSUM method, parameterized in terms of policies to decide when to trigger decimation, which variables to decimate, and which values to assign to decimated variables. Our empirical analysis on a classical BP benchmark indicates that some of these combinations of policies outperform state-of-the-art competitors.
This paper presents an style painting classifier, based on Horn clauses in rational Pavelka logic and qualitative colour descriptors, that provides explanations (SHE). A fuzzy representation of colour traits for Baroque, Impressionism and Post-Impressionism art styles is introduced. The SHE-algorithm has been implemented in Swi-Prolog and tested on the 90 paintings in QArt-Dataset. the Baroque style was classified in 90% of accuracy. The general accuracy for all the art styles was 73.3%. SHE provided explanations of right classifications, and also of outliers by giving a second option depending on the membership degree of the painting to a certain style.
Order-of-magnitude models in qualitative reasoning are basic theoretical tools to model real-world problems in which only incomplete qualitative knowledge is available. In this paper, we introduce interval-valued fuzzy sets into order-of-magnitude qualitative reasoning, generalizing the order-of-magnitude spaces of both qualitative labels and fuzzy-qualitative labels. We present a new lattice structure of linguistic descriptions in an interval-valued fuzzy framework, and introduce the interval-valued fuzzy-qualitative labels as interval-valued fuzzy sets with certain requirements. We define two operations related to the inclusion in the set of interval-valued fuzzy-qualitative labels. The set of interval-valued fuzzy-qualitative labels is structured as a lattice with these two operations.
In this paper, we assess the suitability of a number of different machine learning (ML) methods for detecting shock fronts in Computational Fluid Mechanics (CFD) simulations. Detection and handling of shock fronts is important in a wide variety of fluid mechanics problems. We focus on computational astrophysics, where a successful algorithm must be able to classify the simulated fluid elements as belonging to a no-shock, ahead-of-the-shock-front and behind-the-shock-front class. We implement and test several supervised multi-class classification ML methods for highly imbalanced classes. The training data set is generated by an exact solver for a Riemann Problem [1] (one of the most straightforward non-trivial CFD tests). The most suited algorithm(s) are chosen according to their accuracy, speed, and ease of training. Our preliminary results show that the random forest algorithm (with class balancing) is the best method for classification.
Predicting users' next location becomes an important requirement in various Location-based applications and services. This prediction provides the most probable next location in order to make proactive offerings or services based on those predicted locations. However, prediction models in the literature achieve satisfactory results, but ignore an important fact that values and representations of some variables can be much more relevant to the final location prediction than the rest of variables. In this paper, we study the impact of Space-Time representation learning in location prediction model through evaluating different architectural configurations. First, we evaluate the impact of many different data inputs on the model final prediction performance. Based on that, different prediction models are proposed that vary in terms of the number and type of input features. Second, we investigate the impact of input representation techniques on the prediction performance using both embedding representation learning and one-hot vector representation (i.e. static vectors). We conduct thorough experiments in all our previous models on two real-world datasets, GeoLife and Gowalla.
This paper proposes a multiresolution methodology to visualise and analyse big complex networks. The approach is useful for sensor placement in water distribution systems. Traditional approaches such as CPLEX for facility location in networked structures and eigenvector centrality measures benefit from being addressed at various hierarchical levels of coarseness.
One of the ways of reducing the effects of Climate Change is to rely on renewable energy sources. Their intermittent nature makes necessary to obtain a mid-long term accurate forecasting. Wind Energy prediction is based on the ability to forecast wind speed. This has been a problem approached using different methods based on the statistical properties of the wind time series.
Wind Time series are non-linear and non-stationary, making their forecasting very challenging. Deep neural networks have shown their success recently for problems involving sequences with non-linear behavior. In this work, we perform experiments comparing the capability of different neural network architectures for multi-step forecasting obtaining a 12 hours ahead prediction using data from the National Renewable Energy Laboratory's WIND dataset
We thank the US National Renewable Energy Laboratory (NREL) for the use of their wind energy datasets.
This preliminary work
This work is partially supported by the Joint Study Agreement no. W156463 under the IBM/BSC Deep Learning Center agreement, by the Spanish Government through Programa Severo Ochoa (SEV-2015-0493), by the Spanish Ministry of Science and Technology through TIN2015-65316-P project, and by the Generalitat de Catalunya (contracts 2014-SGR-1051).
Classification of pig carcasses according to its characteristics is a determining factor in slaughterhouses, since allows the optimization of production and improves the performance in cutting plants and in other subsequent processes. Usual criteria in carcasses classification are the weight and the fat content of the ham, especially in regions such as Spain where a significant proportion of hams are used for curing to produce Jamón Serrano. The objective of this study is to compare different models based on Decision Trees and using intrinsic data of pigs to classify hams in four groups according to their predicted weight. The model presented is based on Bagged Decision Trees and use as input: the weight and the lean meat percentage (LMP) of the carcass, the sex and the breed of the pig and the manual classification of the ham according to the thickness of the subcutaneous fat. The results show a success rate of 81.7%, improving by 4.4% the results obtained with a more straightforward decision tree based only on the weight and the LMP.
The main aim of this work is to present a Quaternion Phase Convolutional Neural Network. We encode 3 quaternion phases and its magnitude as an input. Our approach is bio-inspired and is expressed in one mathematical framework. The main result is to obtain a new space feature representation for deep learning which can capture non-trivial equivariant features, related to illumination, and rotation.
We propose in this paper a method which segments demographic and contextual attributes of tourists when visiting an attraction. Clustering patterns based on categorical attributes can be challenging as it is difficult to define a distance between two categorical attributes where a natural order does not exist. Our proposed measure, based on a fuzzy aggregation operator, can be easily implemented in a hierarchical agglomerative clustering algorithm. The method has been implemented in a particular tourist attraction example with 2937 visitors.
We face in this paper the problem of computing distance measures between medical entities. Specifically we deal with the most challenging type of medical entity: drugs. Three different similarity measures between drugs are presented, based each one on specific dimensions of drugs description, namely textual, taxonomic and molecular information. All the information has been extracted from the same resource, the DrugBank database.
Is the brain a predictive machine? Is prediction its only function or one among many? Predictive coding theories argue that all the brain is doing is inference, and the way it is doing it is through Bayesian inference (i.e. predicting what will be the input to the brain in a Bayes-optimal way) and active inference (i.e. the brain acts through a body with the intention of ensuring the predictability of the environment or the same, reduce the free energy [2]). Nonetheless, Bayesian inference is a computationally expensive process for which there's no explanation on how a brain built with neurons and neuromodulators could perform such process. Previous attempts at defining a neuronal mechanism for performing Bayesian inference failed at defining how a model of the world was originally formed [1]. We reformulate the previous attempts to provide a neuronal model that performs Bayesian inference without previous assumptions about the world except for normality. This model becomes a biologically plausible learning mechanism that uses unsupervised local learning rules to learn the statistics of an input and stop learning when the prediction error can't be further reduced.
We study the use of a time series encoder to learn representations that are useful on data set types with which it has not been trained on. The encoder is formed of a convolutional neural network whose temporal output is summarized by a convolutional attention mechanism. This way, we obtain a compact, fixed-length representation from longer, variable-length time series. We evaluate the performance of the proposed approach on a well-known time series classification benchmark, considering full adaptation, partial adaptation, and no adaptation of the encoder to the new data type. Results show that such strategies are competitive with the state-of-the-art, often outperforming conceptually-matching approaches. Besides accuracy scores, the facility of adaptation and the efficiency of pre-trained encoders make them an appealing option for the processing of scarcely- or non-labeled time series.
The role of the coolhunter (trendsetter) has gained importance in the fashion world. This role is in charge of detecting when new trends are occurring. Social media, with its blogs, forums, social platforms, etc., has made the information about clothes and trends more accessible. The task of filtering the relevant from the irrelevant is very difficult for the coolhunter, due to the avalanche of data and information produced in these media. Therefore, there is a need for an automatic system for detecting trends. Trends, or clothing style recognition, is a challenging problem due to wide variation in clothing item appearance, and even the subjectivity associate with the style concept. In fact, clothes that people wear contain latent fashion concepts capturing styles. Fashion theorists have proposed that these concepts are shaped by design elements such as color, material, and silhouette. While a dress may be “classic” because of its color, material, or some combination thereof, it is not always clear how low-level elements translate to high-level styles. This paper presents the firsts steps towards a system that helps in the task of trend determination. The paper presents a tool which has learnt to classify a style in a picture. Specifically, it is able to determine whether the clothes in an image represent a classic outfit or a modern outfit. This system is a cognitive system in the sense that it has been improved with data from a test designed to capture what people think it is classic or modern. We also provide the results of some experiments done using supervised machine learning techniques.
Feature selection (FS) is essential for the analysis of genomic datasets with millions of features. In such context, Big Data tools are paramount, but the use of standard machine learning models is limited for data with such low instances to features ratios. Apache Spark is a distributed in-memory big data system with the potential to overcome this bottleneck. This study analyzes genomic data related to prediction of human obesity. Since Apache Spark is unable to cope with our dataset containing ≈ 0.74 million features, we propose here a pipeline to solve this problem using partitioning strategies, both vertical, by dividing the data based on gender, and horizontal, by splitting each chromosome into 5,000-instances subsets. For each subset, Minimum Redundancy and Maximum Relevance FS was used to find rankings of the most relevant features. The challenge, thus, is making accurate obesity predictions with parsimonious subsets of features selected from millions of them. We tackle it by defining a 2-phase pipeline: first learning with individual chromosomes and then learning with joined 22 chromosomes from selected features.
