Ebook: Artificial Intelligence Research and Development

The International Conference of the Catalan Association for Artificial Intelligence (CCIA) is an international event (now in its 22th edition) that serves as a meeting point not only for researchers in Artificial Intelligence based in the area of the Catalan speaking territories (south France, Catalonia, Valencia, Balearic Islands and Alghero in Italy), but also for researchers around the world.

The volume you have in your hands is a compilation of the articles that were presented in the 2019 edition that took place at the “Colònia de Sant Jordi” in Mallorca (Balearic Islands). Previous editions of the CCIA were in Tarragona (1998), Girona (1999), Vilanova i la Geltrú (2000), Barcelona (2001, 2004, 2014, 2016), Castelló de la Plana (2002), Mallorca (2003), L’Alguer (Italy) (2005), Perpinyà (France) (2006), Andorra (2007), Sant Martí d’Empúries (2008), Cardona (2009), L’Espluga de Francolí (2010), Lleida (2011), Alacant (2012), Vic (2013), València (2015), Deltebre (2017) and Roses (2018).

The CCIA embraces all areas of Artificial Intelligence including logics for AI, machine learning, autonomous agents and multi-agent systems, computer vision, image processing or recommender systems among others. All the articles in this volume have gone through a fully peer-review process performed by an international programme committee that certifies their quality.

We want to express our sincere gratitude to the Catalan Association for Artificial Intelligence (ACIA), the “Universitat de les Illes Balears (UIB)”, the “Departament de Ciències Matemàtiques i Informàtica (UIB)”, the IIIA-CSIC, the SCOPIA Research Group, the “Govern Illes Balears (Conselleria de medi ambient, agricultura i pesca)” and the “Generalitat de Catalunya (Departament de Polítiques Digitals i Administració Pública)” for their support.

Colònia de Sant Jordi (Mallorca), October 2019

Jordi Sabater-Mir (IIIA-CSIC)

Vicenç Torra (Hamilton Institute, Maynooth University)

Isabel Aguiló (UIB)

Manuel González-Hidalgo (UIB)

In the context of Digital Democracy, online participation platforms have emerged as innovative tools that enable citizens to participate in the decision making of their nation, region, or local government. Users can issue proposals and arguments in favour or against them and they can also support other people’s arguments. This paper proposes two alternative support aggregation methods and applies them into debates conducted in the Decidim platform.

In order to be successfully integrated in our society, artificial moral agents need to know not only how to act in a moral scenario, but also how to identify the scenario first as being morally-relevant. This work looks at certain complex video games as simulations of artificial societies and studies the way in which morally-qualifiable actions are identified and assessed in them. Then, this analysis is used to distill a general formal model for moral actions aimed to be used as a first step towards identifying morally-qualifiable actions in the field of artificial morality. After discussing which elements are represented in this model, and how they are enhanced with respect to those already existing in the analyzed games, this work points out to some caveats that those games fail to address, and which would need to be tackled properly by artificial moral systems.

Artificial intelligent (AI) systems making autonomous decisions are present in many areas of our everyday lives. Ideally, and in order to facilitate the integration of these systems into our society, citizens should avoid thinking neither that AI resembles those rogue systems often found in fictional works, nor that it has an intrinsic understanding of human well-being, as both cases would draw a biased picture of what AI actually is. This position paper argues, through some examples, how the terminology used in certain articles aimed for the general public often depict AI as one of those two aforementioned images, which attribute intentions and moral values that those systems do not have and which could, in turn, have a detrimental effect on the way the general public understand (and are willing to accept) those systems into our everyday lives.

Finding tumour genetic markers is essential to biomedicine due to their relevance for cancer detection and therapy development. In this paper, we explore a recently released dataset of chromosome rearrangements in 2,586 cancer patients, where different sorts of alterations have been detected. Using a Random Forest classifier, we evaluate the relevance of several features (some directly available in the original data, some engineered by us) related to chromosome rearrangements. This evaluation results in a set of potential tumour genetic markers, some of which are validated in the bibliography, while others are potentially novel.

With the widespread use of mobile phones, the number of malware targeting smart devices has increased exponentially. In particular, the number of malware targeting Android devices, as it is the most popular operative system among smartphones. This paper proposes a novel framework for android malware detection based on the function call graph representation of an application. Our method generates an embedding of the function call graph using random walks and then, a convolutional neural network extracts features from their embedded matrix representation and labels a given application as benign or malicious considering the learned features. The method has been evaluated on a dataset of 3871 APKs and compared against DREBIN, a baseline benchmark. Experiments show that the method achieves competitive results without relying on the manual extraction of features.

Stochastic Gradient Descent (SGD) is the workhorse beneath the deep learning revolution. However, SGD is known to reduce its convergence speed due to the plateau phenomenon. Stochastic Natural Gradient Descent (SNGD) was proposed by Amari to resolve that problem by taking benefit of the geometry of the space. Nevertheless, the convergence of SNGD is not guaranteed.

The aim of this article is to modify SNGD to obtain a convergent variant, that we name Convergent SNGD (CSNGD), and test it in a specific toy optimization problem. In particular, we concentrate on the problem of learning a discrete probability distribution.

Based on variable metric convergence results presented by Sunehag et al. [13], we prove the convergence of CSNGD. Furthermore, we provide experimental results showing that it significantly improves over SGD. We claim that the approach developed in this paper could be extensible to more complex optimization problems making it a promising research line.

The purpose of feature extraction on convolutional neural networks is to reuse deep representations learnt for a pre-trained model to solve a new, potentially unrelated problem. However, raw feature extraction from all layers is unfeasible given the massive size of these networks. Recently, a supervised method using complexity reduction was proposed, resulting in significant improvements in performance for transfer learning tasks. This approach first computes the discriminative power of features, and then discretises them using thresholds computed for the task. In this paper, we analyse the behaviour of these thresholds, with the purpose of finding a methodology for their estimation. After a comprehensive study, we find a very strong correlation between problem size and threshold value, with coefficient of determination above 90%. These results allow us to propose a unified model for threshold estimation, with potential application to transfer learning tasks.

Laser Ablation – Inductively Coupled Plasma Mass Spectrometry (LA-ICPMS) is a surface-based technique used to quantify the chemical composition of a solid to its elemental and isotopic level. The output signal for each LA-shot corresponds to a set of time series, in intensities (counts-per-second, cps), that provides information on the quantity of each isotope. LA-ICPMS is widely used in biological sciences. For instance in fish ecology, it is used to analyze fish otoliths (ear stones) to obtain information on the fish’s life history (i.e., origin, migrations or exposure to contaminants). The experimental protocol for translating the actual output from LA-ICPMS into isotope concentration is long and complex. The first step is specially time consuming: the intensities obtained from each shot have to be reviewed one by one by an expert to eliminate procedural spikes and define the intervals that optimally represent (1) the background noise (blank) and (2) the background noise plus the signal (plateau). Here we propose a method to facilitate this first step using a trained neural network. The ELM was trained using cases previously processed to emulate the decisions of the expert. Our results showed that in comparison to the manual treatment the quality of the assessment with ELM was optimal for an automatic processing.

When playing video-games we immediately detect which entity we control and we center the attention towards it to focus the learning and reduce its dimensionality. Reinforcement Learning (RL) has been able to deal with big state spaces, including states derived from pixel images in Atari games, but the learning is slow, depends on the brute force mapping from the global state to the actions values (Q-function), thus its performance is severely affected by the dimensionallity of the state and cannot be transferred to other games or other parts of the same game. We propose different transformations of the input state that combine attention and agency detection mechanisms which both have been addressed separately in RL but not together to our knowledge. We propose and benchmark different architectures including both global and local agency centered versions of the state and also including summaries of the surroundings. Results suggest that even a redundant global local state network can learn faster than the global alone. Summarized versions of the state look promising to achieve input-size independence learning.

Importance sampling is a Monte Carlo method that samples from an alternative distribution, the proposal distribution. It focuses the sampling process in the interesting parts of space reducing the variance. The efficiency of importance sampling grows as the proposal distribution resembles the original probability distribution (with thicker tails). A reasonable idea for selecting an appropriate proposal distribution is to use a variational approach. Variational inference selects, from a given family, the distribution which minimizes the divergence to the distribution of interest. In this study, we identify the Rényi projection of order 2 as the one that leads to minimum variance estimators. However, its high computational cost pushes us to consider the standard variational approach (I-projection, which considers the reverse Kullback-Leibler divergence), with limited performance in the general case. We finally show that softening the I-projection is an interesting practical work-around to take the best from both projections. The whole discussion is supported by an empirical study in a simple domain of discrete multivariate distributions.

The paper presented exposes a novel approach to feed data to a Convolutional Neural Network (CNN) while training. Normally, neural networks are fed with shuffled data without any control of what type of examples contains a mini-batch. For situations where data are abundant and there does not exist an unbalancing between classes, shuffling the training data is enough to ensure a balanced mini-batch. On the contrary, most real-world problems end up with databases where some classes are predominant vs others, ill-conditioning the training network to learn those classes forgetting the others.

For those conditioned cases, most common methods simply discard a certain number of samples until the data is balanced, but this paper proposes an ordered method of feeding data while preserving randomness in the mini-batch composition and using all available samples. This method has proven to solve the problem with unbalanced data-sets while competing with other methods. Moreover, the paper will focus its attention to a well know CNN network structure, named Deep Residual Networks.

The transition from conventional vehicles to autonomous vehicles is regulated thorough ADAS (Advanced Driver Assistance Systems) functionalities. The combination of different ADAS functions allows vehicles navigate on a highway autonomously, but at the same time, following the traffic rules and regulations requirements, and also guaranteeing safety on the road. The practical objective in this article is to implement a Reinforcement Learning method whose actions are based in these regulated functions for autonomous vehicles navigation. With this aim, a study of the state-of-the-art of autonomous vehicles simulators has been completed. Hence, the algorithm will be tested using a five-lane highway simulator, previously selected. Results and performance of the model through experimentation will be presented and evaluated using the simulator for different network architectures.

In many every day examples trained machine learning models are rendered obsolete by an inability to adapt to an ever changing environment. This may happen either because their performance decreases in time or because external agents impose new constraints, for example, in the form of regulations. This situation is particularly worrying in company productions environments where model accuracy needs to be preserved. In such situations, model-agnostic copies have been proposed as a viable method to adapt pre-existing models to the new requirements. In this article we study how the use of copies can be extended to endow classifiers trained in batch with online learning. We propose two online algorithms and validate their performance in a series of well-known problems.

In this paper, we explore a Convolutional Neural Network (CNN) based architecture that learns the audio cues to predict the Big Five personality traits score of a speaker. Our model takes advantage from a pre-trained model on a large database for audio event recognition (AudioSet). The pre-trained model has been fine-tuned on the First Impression Dataset to obtain an audio representation for personality trait recognition. In addition, we interpret our model and generate the visual correlation between the model parameters and learned representations by exploring the Class Activation Maps (CAM). Our results show that our interpretable CNN architecture slightly outperforms, in terms of accuracy, previous methods based on hand-crafted features.We also explore a CNN model trained from scratch which takes as input the raw audio data in the frequency domain, finding some discriminative frequency patterns for each personality trait. The interpretability part reveals the inter-mechanism of the model, showing that some frequency bands are more discriminative for personality trait recognition than others.

Wearable cameras are become more popular in recent years for capturing the unscripted moments of the first-person that help to analyze the users lifestyle. In this work, we aim to recognize the places related to food in egocentric images during a day to identify the daily food patterns of the first-person. Thus, this system can assist to improve their eating behavior to protect users against food-related diseases. In this paper, we use Siamese Neural Networks to learn the similarity between images from corresponding inputs for one-shot food places classification. We tested our proposed method with ‘MiniEgoFoodPlaces’ with 15 food related places. The proposed Siamese Neural Networks model with MobileNet achieved an overall classification accuracy of 76.74% and 77.53% on the validation and test sets of the “MiniEgoFoodPlaces” dataset, respectively outperforming with the base models, such as ResNet50, InceptionV3, and InceptionResNetV2.

Meteorology studies the behaviour of atmospheric phenomena in a specific place and time, it is of enormous dynamism and complexity. This paper presents an approach based on a deep semantic neural network that accurately performs a multi-class pixel-wise classification of diverse cloud types, providing information of great value to complement traditional weather assessment techniques. The obtained results suggest that simple web-cams can be of great help for observations that complement traditional data for short-term weather predictions and generations of warnings.

This paper proposes the use of a Case-Based Reasoning (CBR) system for the control and the supervision of a real wastewater treatment plant (WWTP). A WWTP is a critical system which aims to ensure the quality of the water discharged to the receiving bodies, stablished by applicable regulations. At the current stage the proposed methodology has been tested off-line on a real system for the control of the aeration process in the biological treatment of a WWTP within the ambit ofConsorci Besòs Tordera (CBT), a local water administration in the area of Barcelona. For this purpose, data mining methods are considered to extract the available knowledge from historical data to find a useful case base to be able to generate set-points for the local controllers in the WWTP. The results presented in this work are evaluated taking into account the performance of the CBR method e.g. case base size, CBR cycle time or number of cases resolved satisfactorily (forthcoming steps will include on-line tests). For this purpose, some Key Performance Indicators (KPI) are designed together with the plant manager and process experts, in order to monitor key parameters of the WWTP which are representative of the performance of the control and supervision system. Hence, these KPI are related with water quality regulations —e.g. ammonia concentration in the WWTP effluent— and the economic cost efficiency —e.g. electrical consumption of the installation. In order to evaluate the results, different flat-based memory organizations (i.e. cases are stored sequentially in a list) for the case base are considered. First, a unique case base is used. At the current stage and for the results shown in this work, this case base is divided in multiple libraries depending on a case classification. Finally, the combination of this approach with Rule-Based Reasoning (RBR) methods is proposed for the next stages of the work.

In the Play&Sing project, we are developing an AI platform to support home-based self-training interventions for chronic stroke patients. A large percentage of patients suffering from this disease show motor deficits that clearly hinder their daily activities and diminish their quality of life. In this project we are proposing and testing a new Music Supported Therapy (MST) to induce upper limb motor recovery.With the help of a tablet-based application and a small musical keyboard, we are developing an AI platform to support home-based MST. Specifically, the role of AI algorithms is to support therapists and to boost user engagement by personalizing the interventions according to patient needs and preferences. AI algorithms will provide the therapists with hindsight and foresight tools. In the proposed MST, patients are performing 30 training sessions of 45 minutes with a frequency of 3 sessions per week. In this paper we present our platform and preliminary experiments conducted at a pilot phase.