DaSCI Readings
Are short talks by a DaSCI PhD Student who presents recents results on the different DaSCI research lines. Two presentations per day. Each presentation will be approximately 30 minutes long, followed by 15 minutes for questions
DaSCI Readings 2023
The Surprising Properties of Deep Learning Model Weights and their Relationship to Differential Privacy
Lecturer: Daniel Jiménez López
Date: 17/10/2023
Abstract: The weights of Deep Learning (DL) models hide a surprising amount of information. For example, the weights of a trained model allow reconstructing some training data or inferring the presence of some individual in the training set. Fortunately, privacy defense techniques such as differential privacy maintain the privacy of the training data, preventing the extraction of private information. Interestingly, DL model weights hide information that goes beyond privacy – they can be used to infer model performance, without using datasets!. In this talk, let’s explore whether differential privacy leaves a footprint on model weights and marvel at how we can almost always do this, regardless of training data and model architecture.
Explaining the outputs of agent-based models using subgroup discovery
Lecturer: Elena Romero Contreras
Date: 17/10/2023
Abstract: Agent-based models are a widely used tool for simulating real complex systems. During simulation, a large volume of data is generated about agent actions that is difficult to analyze and understand. To address this problem, we propose the use of explainable machine learning techniques, in particular, subgroup discovery algorithms. These algorithms extract relationships between features that explain the behavior of statistically interesting subgroups of agents with respect to a property of interest, generating a set of rules. We demonstrate the effectiveness of our approach with a case study in the marketing area. The results show that our methodology provides a deeper understanding of consumer behavior and market dynamics, providing potential areas of opportunity for marketing experts.
DL-based computer vision models for problem solving in the steel industry
Lecturer: Marta Fernández Moreno
Date: 09/05/2023
Abstract: One of the biggest challenges in real use cases in the industrial world is the generation of labeled data to create optimal solutions. The generation of such data sets is complex and time-consuming for labeling experts, who have very limited time for such tasks. This problem is further intensified in the case of image segmentation, where detailed and accurate labels are required. In this context, the use of computer vision techniques based on Deep Learning is investigated in order to solve this problem and reduce the time and effort required for the labeling task.
Decision support in craniofacial superimposition using Likelihood Ratio metrics
Lecturer: Práxedes Martínez Moreno
Date: 09/05/2023
Abstract: he evaluation of evidence for presentation in court has been the subject of debate in recent decades. Many score-based biometric systems, such as automatic voice, DNA or fingerprint recognition systems, have been proposed in the literature as a source of information for the evaluation of evidence in forensic cases. However, they have been shown to involve a subjective, non-robust and less informative decision-making process. In 2016, the European Network of Forensic Science Institutes (ENFSI) published a guide for evaluating, interpreting and presenting scientific evidence, which supports the use of a probabilistic model based on Bayes’ Theorem and the Likelihood Ratio (LR) metric for the evaluation and interpretation of different types of evidence. For this reason and due to the lack of proposals based on this metric for the identification of skeletal remains, a methodology is proposed to apply this framework in craniofacial superposition (CFS).
Estimation of biological profile in living and dead using artificial intelligence
Lecturer: Javier Venema Rodríguez
Date: 21/03/2023
Abstract: The estimation of the biological profile (BP), consisting of age, sex, population origin and stature, is a task of great importance and interest in the field of forensic anthropology. In the dead, the analysis of skeletal remains to estimate the BP of an individual is used as a first step to restrict, quickly and economically, the number of individuals to which these remains may belong. In vivo, age estimation is a problem of great interest in the migratory field. However, traditional methods of PB estimation are subjective and manual, which is problematic in the medico-legal field. As an alternative, the use of Artificial Intelligence (AI) techniques could allow automating many of these processes, eliminating this subjectivity and shortening identification times. The seminar will deal with this last point: the creation of an AI-based system that allows the automatic estimation of PB in a reliable, fast, objective and explainable way using different sources of information (radiological images of different types and dimensions, information from different anatomical regions, bone images and metadata). We will delve into why PB estimation is relevant, what we have done so far and what we are working on, what results we have obtained so far, what difficulties we are encountering and what our objectives are. All this involves the use of machine learning and computer vision techniques to solve classification and regression problems using 3D images and models of bones, 2D (radiographs) and 3D (CT scans) medical images, and tabular data relating to the individuals to which the images belong.
Application of Neurosymbolic AI for Sequential Decision Processes
Lecturer: Carlos Núñez Molina
Date: 21/03/2023
Abstract: Throughout the history of AI, two main approaches have emerged: symbolic AI and subsymbolic/connectionist AI. The symbolic approach, predominant until the late 1980s, holds that symbol manipulation is a fundamental part of intelligence and thus proposes to create AI systems based on knowledge and reasoning. In contrast, the subsymbolic approach, predominant from the 1980s to the present day, proposes to build AI systems based on learning from data, as is the case of Deep Learning.
His thesis deals with sequential decision problems, in which an agent must make a series of decisions to achieve long-term goals. In this field, there is also a division between symbolic and subsymbolic methods, epitomized by the fields of Automatic Planning (AP) and Reinforcement Learning (RL), respectively. Thus, the main goal of my thesis is to design neuro-symbolic methods that combine the “symbolic” part of AP, through an explicit representation of knowledge and methods of reasoning about it, with the “neuro” part of RL, through the use of neural networks that allow efficient learning from data.
DaSCI Readings 2022
Deep reinforcement learning in agent-based simulations for media planning
Lecturer: Víctor Alejandro Vargas Pérez
Date: 26/04/2022
Abstract: Agent-based models are an appropriate simulation technique for recreating real complex systems, such as those used in marketing. Reinforcement learning consists of learning a behavioral policy that maximizes a long-term reward signal. In this work we develop a deep reinforcement learning agent that represents a brand in an agent-based model for marketing. The goal of this intelligent agent is to derive an advertising investment strategy that enhances the recognition of that brand. Unlike conventional investment strategies, the learned strategy is dynamic, so the agent makes its online investment decisions based on its current observation of the environment. We chose the Double Deep Q-Network algorithm to train the agent on a set of model instances, each with different dynamics in brand recognition. We first fit a subset of the Double Deep Q-Network hyperparameters over two of the model instances and subsequently use the best configuration to train the agent over all instances. The brand agent learns a dynamic policy that achieves higher levels of brand recognition than a static benchmark policy. We perform an analysis of the obtained policy, where we observe that the branding agent tends to increase investment in the media with the highest impact, but also invest in other media depending on the situation and the characteristics of the instance. These results show the benefits of using an on-line dynamic learning environment in a decision support system for media planning in marketing.
EvoPruneDeepTL: An Evolutionary Pruning Approach for Transfer Learning based Deep Neural Networks
Lecturer: Javier Poyatos Amador
Date: 26/04/2022
Abstract: Finding the best Deep Learning architecture for the problem to be solved is difficult. In recent years, the evolution of neural networks has made it possible to know the appropriate topology for certain problems. On the other hand, the pruning technique aims primarily at reducing the number of trainable parameters of the network, but at the cost of lower network performance. In addition, the fact of eliminating parameters that affect the entire network leads to a complete re-training of the network, which is solved thanks to Transfer Learning, which allows us to set good weights for the convolutional part of the network. EvoPruneDeepTL is an evolutionary pruning model based on Transfer Learning that combines sparse networks with an evolutionary algorithm, so that neurons are removed to adapt the network to the problem to be solved. The evolution of these pruned networks has improved the accuracy of the network and, at the same time, reduced the complexity of the dense part of the network.
Explicability in Artificial Intelligence: how to generate and compare explanations
Lecturer: Iván Sevillano García
Date: 01/03/2022
Abstract: Explainable artificial intelligence is proposed in the literature to provide explanations for reasoning carried out by a machine. However, there is no consensus on how to assess the quality of these explanations. Specifically, for the widely known Local Linear Explanations, there are qualitative proposals for the evaluation of explanations, although they suffer from theoretical inconsistencies. In this article we review the definition of these metrics at a theoretical level and propose a series of more theoretically robust metrics. We also use our new proposal to carry out an analysis of the state of the art of the black box, LIME and SHAP methods, extracting the information that these new metrics offer us on 4 image data sets.
CI-dataset and DetDSCI methodology for detecting too small and too large critical infrastructures in satellite images: Airports and electrical substations as case study
Lecturer: Francisco Pérez Hernández
Date: 01/03/2022
Abstract: This paper presents (1) a specialized Critical Infrastructure dataset, named CI-dataset, organized into two scales, small-scale and large-scale critical infrastructures, and (2) a resolution-independent critical infrastructure detection methodology in two levels (DetDSCI) which first determines the spatial resolution of the input image using a classification model, and then analyzes the image using the appropriate detector for that spatial resolution. This study focuses on two representative classes, airports and electrical substations. Our experiments show that the DetDSCI methodology achieves an improvement in F1 of up to 37.53% with respect to a base model.
DaSCI Readings 2021
Descriptive analysis of breast cancer using data mining
Lecturer: Manuel Trasierras Fresco
Date: 28/06/2021
Abstract: This work presents an approach based on emerging pattern mining to analyse cancer through genomic data. Unlike existing approaches, mainly focused on predictive purposes, this proposal aims to improve the understanding of cancer descriptively, not requiring either any prior knowledge or hypothesis to be validated. Additionally, it enables to consider high-order relationships, so both direct and indirect gene relationships related to different functional pathways in the disease can be retrieved. The prime hypothesis is that splitting genomic cancer data into two subsets, that is, cases and controls, will allow us to determine which genes, and their expressions, are associated with the disease. The possibilities of the proposal are demonstrated by analysing a set of paired breast cancer samples in RNA-Seq format. Some of the extracted insights were already described in the related literature as good cancer biomarkers, while others could describe new functional relationships between different genes.
PAF-ND: addressing multi-class imbalance learning with Nested Dichotomies
Lecturer: José Alberto Fernández Sánchez
Date: 28/06/2021
Abstract: Multi-class classification tasks add additional difficulties to the binary classification problem from several sides. Among them, the possibility of obtaining a homogeneous distribution of the classes involved is often one of the most recurring issues in real world problems. This issue leads to what are known as imbalanced learning scenarios. In this work, we explore a method that improves the predictive ability of models when using a type of decomposition strategy known as Nested Dichotomies. Nested Dichotomies is a solution that hierarchically decomposes the classes of the problem and uses an inference method based on probabilities. The method presented here attempts to modify the probability estimates of these models within the hierarchy towards a more equitable classification of the classes by means of Bézier curves.
Reducing Data Complexity using Autoencoders with Class-informed Loss Functions
Lecturer: David Charte
Date: 15/03/2021
Abstract: The data we currently use for knowledge extraction can show different kinds of complexity: class overlap, complex boundaries, dimensionality, etc. This work proposes and evaluates three autoencoder-based models which help reduce complexity by learning from class labels. We also check which complexity measures are better predictors of classification performance.
Multi-step Histogram Based Outlier Scores for Unsupervised Anomaly Detection: ArcelorMittal Engineering Dataset Case of Study
Lecturer: Ignacio Aguilera
Date: 15/03/2021
Abstract: Multi-step Histogram Based Outlier Scores for Unsupervised Anomaly Detection: ArcelorMittal Engineering Dataset Case of Study.
Abstract: Anomaly detection is the task of detecting samples that behave differently from the rest of the data or that include abnormal values. Unsupervised anomaly detection is the most extended scenario, which means that the algorithms cannot train with a labeled input and do not know the anomaly behavior beforehand. Histogram-based methods are one of the most popular and widely used approaches, remarking a good performance and a low runtime. Despite the good performance, histogram-based anomaly detectors are not capable of processing data flows while updating their knowledge and deal with a high amount of samples. In this paper we propose a new histogram-based approach for addressing the aforementioned problems introducing the ability of updating the information inside a histogram. We have applied these strategies to design a new algorithm called Multi-step Histogram Based Outlier Scores (MHBOS), including five new histogram update mechanisms. MHBOS has been validated using the ODDS Library as a general case of use. A real engineering problem provided by the multinational company ArcelorMittal has been used to further validate the performance in a real scenario. The results have shown the performance and validity of MHBOS as well as the proposed strategies in terms of performance and computing times.
StyleGAN: Background and evolution
Lecturer: Guillermo Gómez Trenado
Date: 22/02/2021
Abstract: The work developed by Tero Karras and his team at Nvidia has been the state-of-the-art in GAN for image generation since 2017. In this DaSCI reading we’ll use this results to discuss different aspects of GAN, the iterative process by which the authors detected and corrected the limitations of their work, the technological solutions that allowed such results and the difficulties that we may find if we face related tasks.
Action Recognition for Anomaly Detection using Transfer Learning and Weak Supervision
Lecturer: Francisco Luque
Date: 18/01/2021
Abstract: Automatic video surveillance is an emerging research area, where a huge number of publications are appearing everyday. Particularly, action anomaly detection is a fairly relevant task nowadays. The mainstream approach to the problem using deep models consists in transfer learning from action recognition and weakly supervised fine-tuning for anomaly detection. The objective of the current study is to identify the key aspects of this approaches, and assess the importance of each decision on the training process. To this end, we propose a specific pipeline, where a model is defined by three key aspects: the action recognition model, the pretraining dataset and the weakly supervised fine-tuning policy. Furthermore, we perform extensive experiments to validate the impact of each of the previous aspects in the final solution.
Fuzzy Monitoring of In-bed Postural Changes for the Prevention of Pressure Ulcers using Inertial Sensors Attached to Clothing
Lecturer: Edna Rocío Bernal Monroy
Date: 18/01/2021
Abstract: Postural changes while maintaining a correct body position are the most efficient method of preventing pressure ulcers. However, executing a protocol ofpostural changes over a long period of time is an arduous task for caregivers.To address this problem, we propose a fuzzy monitoring system for posturalchanges which recognizes in-bed postures by means of micro inertial sensors attached to patients’ clothes. First, we integrate a data-driven model to classifyin-bed postures from the micro inertial sensors which are located in the socksand t-shirt of the patient. Second, a knowledge-based fuzzy model computes thepriority of postural changes for body zones based on expert-defined protocols.Results show encouraging performance in the classification of in-bed posturesand high adaptability of the knowledge-based fuzzy approach.
DaSCI Readings 2020
COVID-19 study based on chest X-rays of patients
Lecturer: Anabel Gómez
Date: 06/11/2020
Abstract: COVID-19 is becoming one of the most infectious diseases of the 21st century. Due to the importance of its early detection, new ways to detect it are emerging. In this study, we focus on its detection using chest X-rays, pointing out the main problems of the most used data sets for this purpose. We propose a new data set and a new methodology that allows us to detect cases of COVID-19 with an accuracy of 76.18%, which is higher than the accuracies obtained by experts.
Image inpainting using non-adversarial networks. Towards a deeper semantic understanding of images
Lecturer: Guillermo Gómez
Date: 06/11/2020
Abstract: In this study we explore the problem of image inpainting from a non-adversarial perspective. Can we use general generative models to solve problems other than those for which it was trained to? Do models acquire a deeper and transferable knowledge about the nature of the images they generate? We propose a novel methodology for the image inpainting problem using the implicit knowledge acquired in non-adversarial generative models.
Sentiment Analysis based Multi-person Multi-criteria Decision Making (SA-MpMcDM) Methodology
Lecturer: Cristina Zuheros
Date: 30/11/2020
Abstract: Traditional decision making models are limited by pre-defined numerical and linguistic terms. We present the SA-MpMcDM methodology, which allows experts to evaluate through unlimited natural language and even through numerical ratings. We propose a deep learning model to extract the expert knowledge from the evaluations. We evaluate the methodology in a real case study, which we collect into the TripR-2020 dataset
MonuMAI: Architectural information extraction of monuments through Deep Learning techniques
Lecturer: Alberto Castillo
Date: 30/11/2020
Abstract: An important part of art history can be discovered through the visual information in monument facades. However, the analysis of this visual information, i.e, morphology and architectural elements, requires high expert knowledge. An automatic system for identifying the architectural style or detecting the architectural elements of a monument based on one image will certainly help improving our knowledge in art and history.
The aim of this seminary is to introduce the MonuMAI (Monument with Mathematics and Artificial Intelligence) framework published in the related work [1]. In particular, we designed MonuMAI dataset considering the proposed architectural styles taxonomy, developed MonuMAI deep learning pipeline, and built citizen science based MonuMAI mobile app that uses the proposed deep learning pipeline and dataset for performing in real life conditions.
[1] Lamas, Alberto & Tabik, Siham & Cruz, Policarpo & Montes, Rosana & Martínez-Sevilla, Álvaro & Cruz, Teresa & Herrera, Francisco. (2020) MonuMAI: Dataset, deep learning pipeline and citizen science based app for monumental heritage taxonomy and classification. Neurocomputing. doi.org/10.1016/j.neucom.2020.09.041