Search

Background: In the last years, neural networks have been massively adopted by industry and research in a wide variety of contexts. Neural network milestones are generally reached by scaling up computation, completely disregarding the carbon footprint required for the associated computations. This trend has become unsustainable given the ever-growing use of deep learning, and could cause irreversible damage to the environment of our planet if it is not addressed soon. Objective: In this study, we aim to analyze not only the effects of different energy saving methods for neural networks but also the effects of the moment of intervention, and what makes certain moments optimal. Method: We developed a novel dataset by training convolutional neural networks in 12 different computer vision datasets and applying runtime decisions regarding layer freezing, model quantization and early stopping at different epochs in each run. We then fit an auto-regressive prediction model on the data collected capable to predict the accuracy and energy consumption achieved on future epochs for different methods. The predictions on accuracy and energy are used to estimate the optimal training path. Results: Following the predictions of the model can save 56.5% of energy consumed while also increasing validation accuracy by 2.38% by avoiding overfitting.The prediction model developed can predict the validation accuracy with a 8.4% of error, the energy consumed with a 14.3% of error and the trade-off between both with a 8.9% of error. Conclusions: This prediction model could potentially be used by the training algorithm to decide which methods apply to the model and at what moment in order to maximize the accuracy-energy trade-off., This work has been supported by the GAISSA project (TED2021-130923B-I00, which is funded by MCIN/AEI/ 10.13039/501100011033 and by the European Union ‘‘NextGenerationEU’’/PRTR). The authors also thank the ‘‘Beatriz Galindo’’ Spanish Program (BEAGAL18/00064), and the European Union under the Italian National Recovery and Resilience Plan (NRRP) of NextGenerationEU, partnership on ‘‘Telecommunications of the Future’’ (PE0000001 - program ‘‘RESTART’’)., Peer Reviewed, Postprint (published version)

Antecedents: Donada la naturalesa accelerada de la tecnologia actual, és crucial avaluar l'impacte de la Intel·ligència Artificial (IA) en el consum d'energia. Els projectes tradicionals d'Aprenentatge Automàtic (AA) prioritzen mètriques de precisió sobre l'eficiència energètica, creant una bretxa en la recerca sobre el consum d'energia durant la inferència de models d'IA, especialment en aplicacions de visió per computador. Objectius: Aquesta tesi té com a objectius (i) comprendre les característiques clau dels conjunts de dades i models de visió per computador, (ii) millorar l'eficiència energètica de la inferència mitjançant la comparació de models optimitzats i no optimitzats, i (iii) proporcionar pautes per optimitzar el rendiment i la sostenibilitat dels models. Mètode: Realitzem un experiment controlat per avaluar sistemàticament l'impacte de diverses tècniques d'optimització de PyTorch (és a dir, quantificació dinàmica, torch.compile, poda local i poda global) en la qualitat d'inferència dels components d'AA, que inclou precisió, temps, i consum d'energia. Concretament, ens centrem en 75 models reals d'HuggingFace per a tasques de classificació d'imatges i detecció d'objectes. Això implica recopilar conjunts de dades i models de Hugging Face, realitzar experiments d'inferència repetits i analitzar les dades recollides per descobrir informació sobre les compensacions entre l'eficàcia del model i l'eficiència energètica per a ambdues tasques. Resultats: La quantificació dinàmica demostra reduccions significatives en el temps d'inferència i el consum d'energia per a la classificació d'imatges, mentre que la poda local no mostra cap impacte positiu. A més, torch.compile demostra un bon equilibri entre la correcció del model i l'eficiència energètica. Per a la detecció d'objectes, totes les tècniques d'optimització redueixen notablement el temps d'inferència, encara que ho fan a costa de la correcció del model. Conclusions: Destacant l'eficàcia de diversos mètode, Background: Given the fast-paced nature of today's technology, it is crucial to assess the impact of Artificial Intelligence (AI) on energy consumption. Traditional Machine Learning (ML) projects prioritize accuracy metrics over energy efficiency, creating a gap in research regarding energy consumption during AI model inference, especially in Computer Vision (CV) applications. Aims: This thesis aims to (i) understand key features of CV datasets and models, (ii) improve inference energy efficiency by comparing optimized and non-optimized models, and (iii) provide guidelines for optimizing model performance and sustainability. Method: We conduct a controlled experiment to systematically evaluate the impact of various optimization techniques from PyTorch (namely, dynamic quantization, torch.compile, local pruning, and global pruning) on the inference quality of ML components, which includes accuracy, time, and energy consumption. Specifically, we focus on 75 real models from HuggingFace for image classification and object detection tasks. It involves gathering datasets and models from Hugging Face, conducting repeated inference experiments, and analyzing the collected data to uncover insights into the trade-offs between model effectiveness and energy efficiency for both tasks. Results: Dynamic quantization demonstrates significant reductions in inference time and energy consumption for image classification, while local pruning shows no positive impact. Additionally, torch.compile demonstrates a good balance between model correctness and energy efficiency. For object detection, all optimization techniques notably reduce inference time, though at the expense of model correctness. Conclusions: By highlighting the effectiveness of diverse optimization methods, this study provides practitioners with crucial insights for informed decision-making. Practical guidelines are provided, emphasizing the importance of selecting optimization strategies tailored to specific task requi

This thesis explores the integration of domain-specific knowledge into the design of machine learning (ML) systems, by examining the specific use case of chess board and pieces recognition. Using deep learning models, we compare brute force approaches to algorithmic methods that leverage chess-specific context, such as previous board positions and legal moves. We have analyzed and compared the performance of the algorithms in terms of accuracy, prediction time, energy consumption, and model file size. Our study demonstrates that algorithmic approaches significantly outperform brute force methods across most metrics, achieving higher prediction accuracy and reduced computational costs. Specifically, our proposed algorithm achieves a board prediction accuracy of 96.85%, a square accuracy of 99.89%, and a median prediction time of 0.353 seconds per board, marking a substantial improvement over state-of-the-art models in the field of chessboard recognition. These results underline the importance of incorporating domain knowledge into ML model design, yielding gains not only in predictive accuracy but also in computational efficiency.

© Owner/Author | ACM 2024. This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in: 2024 IEEE/ACM International Workshop on Methodological Issues with Empirical Studies in Software Engineering, WSESE 2024: 16 April 2024, Lisbon, Portugal: proceedings, http://dx.doi.org/10.1145/3643664.3648204, The rapidly evolving fields of Machine Learning (ML) and Artificial Intelligence have witnessed the emergence of platforms like Hugging Face (HF) as central hubs for model development and sharing. This experience report synthesizes insights from two comprehensive studies conducted on HF, focusing on carbon emissions and the evolutionary and maintenance aspects of ML models. Our objective is to provide a practical guide for future researchers embarking on mining software repository studies within the HF ecosystem to enhance the quality of these studies. We delve into the intricacies of the replication package used in our studies, highlighting the pivotal tools and methodologies that facilitated our analysis. Furthermore, we propose a nuanced stratified sampling strategy tailored for the diverse HF Hub dataset, ensuring a representative and comprehensive analytical approach. The report also introduces preliminary guidelines, transitioning from repository mining to cohort studies, to establish causality in repository mining studies, particularly within the ML model of HF context. This transition is inspired by existing frameworks and is adapted to suit the unique characteristics of the HF model ecosystem. Our report serves as a guiding framework for researchers, contributing to the responsible and sustainable advancement of ML, and fostering a deeper understanding of the broader implications of ML models., This work is supported by the project TED2021-130923B-I00, funded by MCIN/AEI/10.13039/501100011033 and the European Union Next Generation EU/PRTR., Peer Reviewed, Postprint (author's final draft)

As machine learning systems evolve, they face the challenge of operating in dynamic environments where model performance can degrade over time due to model drift. Therefore, the work continues once the model is in production; continuous monitoring and adjustment are essential to maintain optimal performance. As a model receives new data, it might exhibit a different probability distribution than the data used for training, necessitating strategies to detect and address drift effectively. This thesis addresses these challenges within a chess broadcasting system developed by the GESSI research group. This system leverages machine learning for real-time chess game state detection. Despite its capabilities, it is not immune to data and concept drift challenges, which can diminish its effectiveness in production environments. Moreover, this work aims to enhance the efficiency and scalability of the chess game state recognition system by optimizing both the data storage solution and the processing pipeline. These enhancements enable the system to perform consistently and reliably in production, with improved data accessibility and streamlined operations. To address these issues, this thesis presents three main contributions: (1) an enhanced pipeline for image validation that reduces human error and increases efficiency, (2) the migration of the existing data storage from CSV files to a MongoDB database, optimizing scalability and data accessibility, and (3) the integration of drift detection mechanisms, utilizing both autoencoders and production model embeddings, to monitor and mitigate drift in the system's pipeline. Through this research, we demonstrate that embedding-based drift detection is viable for high-dimensional data such as images. Furthermore, we show that the production model can serve as a reliable and resource-efficient alternative to custom autoencoders for feature extraction, offering comparable performance in drift detection tasks. Our findings highlight

Agile software development (ASD) aims to deliver high-quality software. However, software quality often receives less attention in ASD due to focus on functionality. Elicit requirements for, design, validate, implement, and evaluate an evidence-based quality-aware ASD process that organizations can use for better integration of quality management. Using design science research, we collaborated for over 1.5 years with four organizations applying ASD. Involving a total of 35 practitioners, we designed new evidence-based quality-aware ASD processes, which were validated through simulated scenarios, implemented, and evaluated by the practitioners. A general ASD process model, derived from the companyspecific models, was created; it shows activities, artifacts, roles, and development phases where evidence can be used to improve software quality. The new process model contributed to higher quality awareness, improved bottleneck identification, release readiness, and quality issues prioritization in the studied companies. We also learned that companies must be mindful of the rules that determine the use of evidence. Integrating the use of quality related evidence into the ASD process can make it quality-aware, support software quality management, and enable timely reaction through fast feedback loops. The evidence-based quality-aware ASD process is applicable in any Scrum-like ASD process. Applicability in other forms of ASD, however, requires further research., This work was supported by the Q-Rapids project though the European Union’s Horizon 2020 Research and Innovation Program under Grant 732253., Peer Reviewed, Postprint (published version)

El nombre de sistemes que incorporen components d'intel·ligència artificial (IA) està creixent notablement els últims anys. Aquestes tecnologies poden tenir un fort impacte mediambiental, a causa del seu consum energètic, durant el seu desenvolupament i funcionament. Aquest projecte presenta l'eina GAISSALabel. Una eina que permet avaluar el consum energètic dels components d'IA i proporcionar una etiqueta d'eficiència energètica. Aquesta etiqueta permet entendre les implicacions mediambientals i, eventualment, facilitar la disminució d'aquest impacte. Seguint una metodologia àgil en el desenvolupament, s'analitzen els requisits de GAISSALabel, s'especifica, dissenya, implementa i desplega aquesta eina. S'analitza com l'eina es pot integrar en el procés de desenvolupament de sistemes basats en IA. D'aquesta manera, es presenten exemples d'ús que inclouen la generació d'etiquetes energètiques i recomanacions per reduir el consum d'energia. Per acabar, es conclou que GAISSALabel facilita la conscienciació i reducció de l'impacte mediambiental dels sistemes basats en IA., The number of systems incorporating artificial intelligence (AI) components has been growing significantly in recent years. These technologies can have a strong environmental impact due to their energy consumption during development and operation. This project introduces the tool GAISSALabel. A tool that allows evaluating the energy consumption of AI components and providing an energy efficiency label. This label helps understand the environmental implications and, ultimately, facilitates the reduction of this impact. Following an agile methodology in development, the requirements of GAISSALabel are analyzed, and the tool is specified, designed, implemented, and deployed. The tool's integration into the development process of AI-based systems is examined. In this way, usage examples are presented, including the generation of energy labels and recommendations to reduce energy consumption. In conclusion, it is noted that GAISSALabel facilitates awareness and reduction of the environmental impact of AI-based systems.

Recently, the proliferation of commercial ML-based services has given rise to new job roles, such as ML engineers. Despite being highly sought-after in the job market, ML engineers are difficult to recruit, possibly due to the lack of specialized academic curricula for this position at universities. To address this gap, in the past two years, we have supplemented traditional Computer Science and Data Science university courses with a project-based course on MLOps focused on the fundamental skills required of ML engineers. In this paper, we present an overview of the course by showcasing a couple of sample projects developed by our students. Additionally, we share the lessons learned from offering the course at two different institutions., This work is partially supported by the NRRP Initiative – Next Generation EU ("FAIR - Future Artificial Intelligence Research", code PE00000013, CUP H97G22000210007); the Complementary National Plan PNC-I.1 ("DARE - DigitAl lifelong pRevEntion initiative", code PNC0000002, CUP B53C22006420001), and the project TED2021- 130923B-I00, funded by MCIN/AEI/10.13039/50110001 1033 and the European Union Next Generation EU/PRTR., Peer Reviewed, Postprint (author's final draft)

Background: The usage of complex artificial intelligence (AI) models demands expensive computational resources. While currently, available high-performance computing environments can support such complexity, the deployment of AI models in mobile devices, which is an increasing trend, is challenging. Environments with low computational resources imply limitations in the design decisions during the AI-enabled software engineering lifecycle that balance the trade-off between the accuracy and the complexity of the mobile applications. Objective: Our objective is to systematically assess the trade-off between accuracy and complexity when deploying complex AI models (e.g. neural networks) to mobile devices in pursuit of greener AI solutions. We aim to cover (i) the impact of the design decisions on the achievement of high-accuracy and low resource-consumption implementations; and (ii) the validation of profiling tools for systematically promoting greener AI. Method: We implement neural networks in mobile applications to solve multiple image and text classification problems on a variety of benchmark datasets. We then profile and model the accuracy, storage weight, and time of CPU usage of the AI-enabled applications in operation with respect to their design decisions. Finally, we provide an open-source data repository following the EMSE open science practices and containing all the experimentation, analysis, and reports in our study. Results: We find that the number of parameters in the AI models makes the time of CPU usage scale exponentially in convolutional neural networks and logarithmically in fully-connected layers. We also see the storage weight scales linearly with the number of parameters, while the accuracy does not. For this reason, we argue that a good practice for practitioners is to start small and only increase the size of the AI models when their accuracy is low. We also find that Residual Networks (ResNets) and Transformers have a higher baseline cost in t, This work has been partially supported by the GAISSA project (TED2021-130923B-I00, which is funded by MCIN/AEI/10.13039/501100011033 and by the European Union “NextGenerationEU”/PRTR); and, by the “Beatriz Galindo” Spanish Program (BEAGAL18/00064)., Peer Reviewed, Postprint (author's final draft)

Nowadays, AI-based systems have achieved out-standing results and have outperformed humans in different domains. However, the processes of training AI models and inferring from them require high computational resources, which pose a significant challenge in the current energy efficiency societal demand. To cope with this challenge, this research project paper describes the main vision, goals, and expected outcomes of the GAISSA project. The GAISSA project aims at providing data scientists and software engineers tool-supported, architecture-centric methods for the modelling and development of green AI-based systems. Although the project is in an initial stage,we describe the current research results, which illustrate the potential to achieve GAISSA objectives., This paper has been funded by the GAISSA Spanish research project (ref. TED2021-130923B-I00; MCIN/AEI/10.13039/501100011033)., Peer Reviewed, Objectius de Desenvolupament Sostenible::7 - Energia Assequible i No Contaminant, Objectius de Desenvolupament Sostenible::13 - Acció per al Clima, Postprint (author's final draft)

ML pipelines, as key components of ML systems, shall be developed following quality assurance techniques. Unfortunately, it is often the case in which they present maintainability issues, due to the experimentatal nature of data collection and ML model construction. To address this problem, this work in progress proposes initial metrics to measure the presence of code smells in ML pipelines. These metrics reflect good software engineering practices for code quality of ML pipelines., This work was supported by the “Spanish Ministerio de Ciencia e Innovación” under project / funding scheme PID2020-117191RB-I00 / AEI/10.13039/501100011033., Peer Reviewed, Postprint (author's final draft)

Als darrers anys, l'Aprenentatge Profund ha assolit moltes fites en una àmplia varietat de camps. Aquestes fites generalment s'aconsegueixen mitjançant l'escalatge de la computació, sense tenir en compte la petjada de carboni que això podria causar. Presentem un marc de treball per reduir el consum d'energia en l'entrenament de models d'Aprenentatge Profund mentre sense deixar de valorar la precisió, basat en un sistema de 4 passos: monitorització, predicció, entrada de l'usuari, acció. Aquest sistema permet prendre decisions interactives durant l'entrenament basades en les dades recopilades fins a aquest punt. Per desenvolupar el sistema de predicció, creem un nou conjunt de dades entrenant una xarxa neuronal convolucional en 12 conjunts de dades múltiples vegades, aplicant els mètodes triats en diferents moments de l'entrenament. El sistema desenvolupat pot predir la precisió de validació amb un 8.4\% d'error, el consum d'energia amb un 14.3\% d'error i l'equilibri entre ambdós amb un 8.9\% d'error. La eina implementada es va mostrar a un grup de voluntaris que la van valorar amb una mitjana de 4.3 sobre 5 en utilitat percebuda i 3.8 sobre 5 en facilitat d'ús percebuda., In the last years, Deep Learning has achieved loads of milestones in a wide variety of ambits. Those milestones are generally reached by scaling up computation, disregarding the carbon footprint that this could cause. We present a framework to reduce the energy consumption of training DL models while still assessing the accuracy, based on a 4-step system: monitoring, prediction, user input, action. This system allows for interactive decisions during training based on the collected data up to that point. To develop the prediction system, we create a novel dataset on training a convolutional neural network on 12 datasets multiple times, each one applying the chosen methods at different moments of the training. The system developed can predict the validation accuracy with a 8.4\% of error, the energy consumed with a 14.3\% of error and the trade-off between both with a 8.9\% of error. The tool implemented was shown to a group of volunteers that rated it with an average of 4.3 out of 5 on perceived usefulness and a 3.8 out of 5 on perceived ease of use.

Background: The rise of machine learning (ML) systems has exacerbated their carbon footprint due to increased capabilities and model sizes. However, there is scarce knowledge on how the carbon footprint of ML models is actually measured, reported, and evaluated. Aims: This paper analyzes the measurement of the carbon footprint of 1,417 ML models and associated datasets on Hugging Face. Hugging Face is the most popular repository for pretrained ML models. We aim to provide insights and recommendations on how to report and optimize the carbon efficiency of ML models. Method: We conduct the first repository mining study on the Hugging Face Hub API on carbon emissions and answer two research questions: (1) how do ML model creators measure and report carbon emissions on Hugging Face Hub?, and (2) what aspects impact the carbon emissions of training ML models? Results: Key findings from the study include a stalled proportion of carbon emissions-reporting models, a slight decrease in reported carbon footprint on Hugging Face over the past 2 years, and a continued dominance of NLP as the main application domain reporting emissions. The study also uncovers correlations between carbon emissions and various attributes, such as model size, dataset size, ML application domains and performance metrics. Conclusions: The results emphasize the need for software measurements to improve energy reporting practices and the promotion of carbon-efficient model development within the Hugging Face community. To address this issue, we propose two classifications: one for categorizing models based on their carbon emission reporting practices and another for their carbon efficiency. With these classification proposals, we aim to encourage transparency and sustainable model development within the ML community., This work is partially supported by the project TED2021- 130923B-I00, funded by MCIN/AEI/10.13039/501100011033 and the European Union Next Generation EU/PRTR., Peer Reviewed, Postprint (author's final draft)

Current research in the computer vision field mainly focuses on improving Deep Learning (DL) correctness and inference time performance. However, there is still little work on the huge carbon footprint that has training DL models. This study aims to analyze the impact of the model architecture and training environment when training greener computer vision models. We divide this goal into two research questions. First, we analyze the effects of model architecture on achieving greener models while keeping correctness at optimal levels. Second, we study the influence of the training environment on producing greener models. To investigate these relationships, we collect multiple metrics related to energy efficiency and model correctness during the models’ training. Then, we outline the trade-offs between the measured energy efficiency and the models’ correctness regarding model architecture, and their relationship with the training environment. We conduct this research in the context of a computer vision system for image classification. In conclusion, we show that selecting the proper model architecture and training environment can reduce energy consumption dramatically (up to 98.83%) at the cost of negligible decreases in correctness. Also, we find evidence that GPUs should scale with the models’ computational complexity for better energy efficiency., This work is part of the GAISSA project (TED2021-130923B-I00), which is funded by MCIN/AEI/10.13039/501100011033 and by the European Union "NextGenerationEU”/PRTR., Peer Reviewed, Postprint (author's final draft)

Les aplicacions d'aprenentatge automàtic (ML) han experimentat un ràpid creixement en els darrers anys, la qual cosa ha provocat un augment del consum d'energia associat a la formació i el desplegament de models ML, fet que ha generat preocupacions sobre el seu impacte ambiental. En aquest projecte de recerca, pretenem investigar l'eficiència energètica dels models ML durant les etapes d'optimització i inferència. Investigarem específicament l'efecte de les estratègies d'optimització de ML com la quantificació i la poda sobre l'energia consumida i la correctesa del model ML. També veurem com els diferents marcs d'optimització, com PyTorch i TensorFlow, afecten aquestes mètriques tant en les fases d'optimització com d'inferència. A més, realitzarem una avaluació contextual de l'eficiència energètica per a diversos models de ML en diversos dominis, inclosos el processament del llenguatge natural, la visió per ordinador i la generació de codi. Mitjançant l'exploració de sis models de ML, abordarem qüestions de recerca relacionades amb l'energia consumida, la correctesa i l'optimització de la compensació. Tot això es farà mitjançant dos scripts de Python: l'script d'optimització, que optimitza els sis models i desa totes les versions al disc, i l'script d'inferència, que agafa els models desats per l'script d'optimització i fa inferència sobre ells. Tots dos scripts també generen un fitxer csv amb els resultats de cada fase. Els resultats del projecte mostren que els marcs, dominis, models i estratègies d'optimització de ML tenen un efecte significatiu en el consum d'energia tant en optimització com en inferència: TensorFlow consumeix menys energia que PyTorch però PyTorch aconsegueix una major correctesa en inferència, mentre que PyTorch funciona millor. en optimització, i la poda consumeix menys energia que la quantificació. Al mateix temps, els models de visió per ordinador són molt eficients energèticament, mentre que els models de codi requereixen una mica més d'en, Machine learning (ML) applications have witnessed rapid growth in recent years, leading to increased energy consumption associated with training and deployment of ML models, which has raised concerns about their environmental impact. In this research project, we aim to investigate the energy efficiency of ML models during the optimization and inference stages. We will specifically investigate the effect of ML optimization strategies such as quantization and pruning on the energy consumed and ML model correctness. We will also look at how different optimization frameworks, such as PyTorch and TensorFlow, affect these metrics both in the optimization and inference phases. Additionally, we will conduct a context-aware evaluation of energy efficiency for diverse ML models across various domains, including Natural Language Processing, Computer Vision and Code Generation. Through the exploration of six ML models, we will address research questions related to the energy consumed, correctness, and trade-off optimization. All of this will be done using two Python scripts: the optimization script, which optimizes the six models and saves all the versions in disk, and the inference script, which takes the models saved by the optimization script and performs inference on them. Both these scripts also generate a csv file with the results of each phase. The results of the project show that ML frameworks, domains, models and optimization strategies have a significant effect on the consumption of energy in both optimization and inference: TensorFlow consumes less energy than PyTorch but PyTorch achieves a higher correctness in inference, while PyTorch performs better in optimization, and pruning consumes less energy than quantization. At the same time, Computer Vision models are very energy efficient while code models are slightly more energy demanding than text models. The findings of this research project provide insights into optimizing the energy consumed in ML models, contribu

Context and motivation: AI-based systems (i.e., systems integrating some AI model or component) are becoming pervasive in society. A number of characteristics of AI-based systems challenge classical requirements engineering (RE) and raise questions yet to be answered. Question: This vision paper inquires the role that RE should play in the development of AI-based systems with a focus on three areas: roles involved, requirements’ scope and non-functional requirements. Principal Ideas: The paper builds upon the vision that RE shall become the cornerstone in AI-based system development and proposes some initial ideas and roadmap for these three areas. Contribution: Our vision is a step towards clarifying the role of RE in the context of AI-based systems development. The different research lines outlined in the paper call for further research in this area., This paper is part of the project TED2021-130923B-I00, funded by MCIN/AEI/10.13039/501100011033 and the European Union “NextGenera-tionEU”/PRTR., Peer Reviewed, Postprint (author's final draft)

Building and maintaining production-grade ML-enabled components is a complex endeavor that goes beyond the current approach of academic education, focused on the optimization of ML model performance in the lab. In this paper, we present a project-based learning approach to teaching MLOps, focused on the demonstration and experience with emerging practices and tools to automatize the construction of ML-enabled components. We examine the design of a course based on this approach, including laboratory sessions that cover the end-to-end ML component life cycle, from model building to production deployment. Moreover, we report on preliminary results from the first edition of the course. During the present year, an updated version of the same course is being delivered in two independent universities; the related learning outcomes will be evaluated to analyze the effectiveness of project-based learning for this specific subject., This work is partially supported by the project TED2021-130923B-I00, funded by MCIN/AEI/10.13039/501100011033 and the European Union Next Generation EU/PRTR., Peer Reviewed, Postprint (author's final draft)

Nowadays there are lots of emerging use cases and fields to artificial intelligence/machine learning technologies. One of several relevant fields is that of image recognition model training, which itself has several potential applications already in development. This project studies this field through a specific case study: a machine learning model built to recognise and differentiate between chess piece images. For models to accurately discern what they are being presented with it is pivotal to meticulously assemble and refine the data they receive. In this project, based on a previously existing dataset of chess board images, we delve with the acquisition of curated chessboard images, each representing a spectrum of game states and scenarios. This is done in two ways: firstly, by capturing images from physical chess boards during in-person tournaments, and secondly by recreating digital versions of real professional chess games in a physical artificial environment in order to enrich our dataset substantially. This foundational step furnishes the bedrock components of our data collection pipeline, which is the second part of the project. Once the previous step has been completed we proceed by constructing a preprocessing pipeline tailor-fitted for preparing data for model training, automatizing the meticulous labeling of raw data and curating data consistently. In summary, this thesis is wholly devoted to the betterment and refinement of a machine learning dataset for chess board image recognition, enhancing data collection and automatizing data preprocessing workflows. Insights gleaned from this project bear significance for the continued development of computer vision systems.

Aquest treball de fi de grau investiga l'ús d'arquitectures d'aprenentatge profund per a la tasca de digitalització automàtica de partides d'escacs. El treball previ és escàs i, tot i que ha mostrat resultats prometedors, les tècniques encara necessiten millores addicionals per permetre'n un ús pràctic en producció. La nostra primera contribució es va fer comparant el rendiment de ResNet50, NASNet Mobile, Mobilenet i VGG16 en la classificació de les peces d'escacs a partir d'una imatge de la peça. L'estudi va trobar que totes les arquitectures van ser capaces d'assolir una precisió satisfactòria, arribant al 98\% d'exactitud amb ResNet50. Aquestes arquitectures es van entrenar per desenvolupar un sistema capaç de recuperar les posicions en un tauler d'escacs a partir d'una imatge. En les millors condicions, el sistema pot digitalitzar una imatge d'un tauler d'escacs en 4 segons amb una exactitud del 75\%. Els resultats de les proves de producció destaquen la importància de l'ús de coneixement del domini que tinguem a la nostra disposició i una extracció cuidadosa de les dades, ja que creiem que la majoria de les imprecisions que el sistema cometeix es podrien resoldre amb la introducció d'informació del domini i una forma més òptima d'obtenir les caselles del tauler. L'estudi també proporciona informació sobre el consum d'energia durant l'entrenament de les Xarxes Neuronals Convolucionals., This bachelor's thesis investigates the use of deep learning architectures for the task of automatic digitization of chess games. Previous work is scarce and, although it has shown promising results, the techniques still need further enhancements to allow a practical use in production. Our first contribution was done by comparing the performance of ResNet50, NASNet Mobile, Mobilenet, and VGG16 when classifying chess pieces given a piece image. The study found that all the architectures were able to retrieve a high satisfactory accuracy, reaching a 98\% accuracy with ResNet50. Those architectures were trained in order to develop a system able to retrieve the positions in a chess board given just an image. In the best conditions, the system is able to digitise a chess board image in 4 seconds with a 75\% accuracy. The results from the production testing highlight the importance of the use of domain knowledge and a careful extraction of the data, as we believe that the majority of imprecisions the system makes could be solved with the introduction of domain information and a more optimal way to get the squares from the board. The study also gives some insight about the energy consumption of the Convolutional Neural Networks training.

El auge de los sistemas de machine learning (ML), la mejora de sus capacidades y el mayor tamaño de los sistemas, ha incrementado el impacto medioambiental de los modelos ML. Sin embargo, la información sobre cómo se mide, comunica y evalúa la huella de carbono de los modelos de ML es escasa. Este proyecto, basado en un análisis de 1.417 modelos de ML y conjuntos de datos asociados en Hugging Face, el repositorio más popular para modelos de ML preentrenados, tiene como objetivo proporcionar una solución integrada para comprender, informar y optimizar la eficiencia de carbono de los modelos de ML. Además, implementamos una aplicación web que genera etiquetas de eficiencia energética para modelos de ML y permite visualizar sus emisiones de carbono. Con menos del 1% de los modelos en Hugging Face proporcionando información sobre las emisiones de carbono, el proyecto subraya la necesidad de mejorar las prácticas de reporte energético y la promoción del desarrollo de modelos eficientes en carbono dentro de la comunidad Hugging Face. Para abordar esta cuestión, ofrecemos una herramienta web que produce etiquetas de eficiencia energética para modelos de ML, una contribución que fomenta la transparencia y el desarrollo de modelos sostenibles dentro de la comunidad de ML. Permite la creación de etiquetas energéticas, al tiempo que proporciona valiosas visualizaciones de los datos de emisiones de carbono. Esta solución integrada constituye un paso importante hacia prácticas de IA más sostenibles medioambientalmente., The rise of machine learning (ML) systems has increased their environmental impact due to the enhanced capabilities and larger model sizes. However, information about how the carbon footprint of ML models is measured, reported, and evaluated remains scarce and scattered. Aims: This project, based on an analysis of 1,417 ML models and associated datasets on Hugging Face, the most popular repository for pretrained ML models, aims to provide an integrated solution for understanding, reporting, and optimizing the carbon efficiency of ML models. Moreover, we implement a web-based application that generates energy efficiency labels for ML models and visualizes their carbon emissions. With less than 1% of models on Hugging Face currently reporting carbon emissions, the project underscores the need for improved energy reporting practices and the promotion of carbon-efficient model development within the Hugging Face community. To address this, we offer a web-based tool that produces energy efficiency labels for ML models, a contribution that encourages transparency and sustainable model development within the ML community. It enables the creation of the energy labels, while also providing valuable visualizations of carbon emissions data. This integrated solution serves as an important step towards more environmentally sustainable AI practices.

Background: When using deep learning models, one of the most critical vulnerabilities is their exposure to adversarial inputs, which can cause wrong decisions (e.g., incorrect classification of an image) with minor perturbations. To address this vulnerability, it becomes necessary to retrain the affected model against adversarial inputs as part of the software testing process. In order to make this process energy efficient, data scientists need support on which are the best guidance metrics for reducing the adversarial inputs to create and use during testing, as well as optimal dataset configurations. Aim: We examined six guidance metrics for retraining deep learning models, specifically with convolutional neural network architecture, and three retraining configurations. Our goal is to improve the convolutional neural networks against the attack of adversarial inputs with regard to the accuracy, resource utilization and execution time from the point of view of a data scientist in the context of image classification. Method: We conducted an empirical study using five datasets for image classification. We explore: (a) the accuracy, resource utilization, and execution time of retraining convolutional neural networks with the guidance of six different guidance metrics (neuron coverage, likelihood-based surprise adequacy, distance-based surprise adequacy, DeepGini, softmax entropy and random), (b) the accuracy and resource utilization of retraining convolutional neural networks with three different configurations (one-step adversarial retraining, adversarial retraining and adversarial fine-tuning). Results: We reveal that adversarial retraining from original model weights, and by ordering with uncertainty metrics, gives the best model w.r.t. accuracy, resource utilization, and execution time. Conclusions: Although more studies are necessary, we recommend data scientists use the above configuration and metrics to deal with the vulnerability to adversarial inputs of deep l, This work was supported by the GAISSA Spanish research project (ref. TED2021-130923B-I00; MCIN/AEI/10.13039/501100011033), the “UNAM-DGECI: Iniciación a la Investigación (verano otoño 2021)” scholarship provided by Universidad Nacional Autónoma de México (UNAM), the “Beatriz Galindo” Spanish Program BEAGAL18/00064, the Austrian Science Fund (FWF): I 4701-N and the project Continuous Testing in Production (ConTest) funded by the Austrian Research Promotion Agency (FFG): 888127., Peer Reviewed, Objectius de Desenvolupament Sostenible::7 - Energia Assequible i No Contaminant, Objectius de Desenvolupament Sostenible::13 - Acció per al Clima, Postprint (published version)

Individuals with ADHD face significant challenges in their daily lives due to difficulties with attention, hyperactivity, and impulsivity. These challenges are especially pronounced in the workplace or educational settings, where the ability to sustain attention and manage time effectively is crucial for success. Virtual reality (VR) software has emerged as a promising tool for improving productivity in individuals with ADHD. However, the effectiveness of such software depends on the identification of potential distractions and timely intervention. The proposed computer screenshot classification approach addresses this need by providing a means for identifying and analyzing potential distractions within VR software. By integrating Convolutional Neural Networks (CNNs), Optical Character Recognition (OCR), and Natural Language Processing (NLP), the proposed approach can accurately classify screenshots and extract features, facilitating the identification of distractions and enabling timely intervention to minimize their impact on productivity. The implications of this research are significant, as ADHD affects a substantial portion of the population and has a significant impact on productivity and quality of life. By providing a novel approach for studying, detecting, and enhancing productivity, this research has the potential to improve outcomes for individuals with ADHD and increase the efficiency and effectiveness of workplaces and educational settings. Moreover, the proposed approach holds promise for wider applicability to other productivity studies involving computer users, where the classification of screenshots and feature extraction play a crucial role in discerning behavioral patterns., Les persones amb TDAH s’enfronten a reptes importants en la seva vida diària a causa de les dificultats d’atenció, hiperactivitat i impulsivitat. Aquests reptes són especialment pronunciats al lloc de treball o en entorns educatius, on la capacitat de mantenir l’atenció i gestionar el temps de manera eficaç és crucial per a l’èxit. El software de realitat virtual (RV) s’ha revelat com a eina prometedora per millorar la productivitat de les persones amb TDAH. Tanmateix, l’eficàcia del software esmentat depèn de la identificació de distraccions potencials i de la intervenció oportuna. L’enfocament de classificació de captures de pantalla d’ordinador proposat aborda aquesta necessitat proporcionant un mitjà per identificar i analitzar les distraccions potencials dins del programari de RV. Mitjançant la integració de xarxes neuronals convolucionals (CNN), el reconeixement òptic de caràcters (OCR) i el processament del llenguatge natural (NLP), l’enfocament proposat pot classificar amb precisió les captures de pantalla i extreure’n característiques, facilitant la identificació de les distraccions i permetent una intervenció oportuna per minimitzar-ne l’impacte en la productivitat. Les implicacions d’aquesta investigació són importants, ja que el TDAH afecta una part substancial de la població i té un impacte significatiu a la productivitat i la qualitat de vida. En proporcionar un enfocament nou per estudiar, detectar i millorar la productivitat, aquesta investigació té el potencial de millorar els resultats per a les persones amb TDAH i augmentar l’eficiència i l’eficàcia dels llocs de treball i els entorns educatius. A més, l’enfocament proposat promet una aplicabilitat més gran a altres estudis de productivitat en què participin usuaris d’ordinadors, en què la classificació de captures de pantalla i l’extracció de característiques tenen un paper crucial a l’hora de discernir patrons de comportament.

Aquest projecte optimitza el procés de facturació a la companyia Amphora Logistics. Millora la precisió i l'eficiència, prediu els pesos dels transportistes i optimitza els costos. Se separa en dos blocs principals: integració i anàlisi de dades. S'analitza el procés de facturació actual, se n'identifiquen les limitacions i es proposa un enfocament integrat. L'extracció de dades automatitzada de les factures dels transportistes i l'anàlisi estadística ajuden els treballadors de finances a controlar els costos d'enviament mitjançant l'ecosistema d'AWS. Una anàlisi comparativa produeix noves tarifes i transportistes òptims en funció de la regió nacional o país i el pes, fent una simulació per avaluar-ne l'impacte. Els models predictius (regressió lineal, xarxes neuronals) prediuen el pes brut i volumètric dels transportistes, utilitzant els productes i el pes brut enviat. Pel que fa a l'impacte d'aquesta tesi, la integració ha reduït un 70\% el temps de facturació, l'anàlisi òptima estalvia un 2.55\% en costos i els models predictius tenen una precisió superior al 90\% al 77.7\% dels enviaments. Aquestes solucions milloren l'eficiència, redueixen els costos i milloren la presa de decisions a Amphora Logistics., This thesis optimizes the billing process in Amphora Logistics Company. It improves accuracy and efficiency, predicts carrier weights, and optimizes costs. Two main blocks are addressed: data integration and analysis. The current billing process is analyzed, limitations identified, and an integrated approach proposed. The automated data extraction from carrier bills and statistical analysis help finance workers in controlling shipping costs, using AWS ecosystem. A comparative analysis leads to new rates and optimal carriers for each national region or country and weight, making a simulation to evaluate their impact. The predictive models (linear regression, neural networks) predict carriers gross and volumetric weights using products and gross weight shipped. As for the impact of this thesis, the integration reduced billing time by 70\%, the optimal analysis saved 2.55\% in costs and predictive models achieved over 90\% accuracy in 77.7\% of orders. These approaches enhance efficiency, reduce costs, and improve decision-making at Amphora Logistics.

Aquesta memòria descriu el treball fet per a l'empresa tecnològica HP, que es dedica a la fabricació i venda d'una àmplia gamma de productes tecnològics, incloent-hi impressores. HP volia estudiar la influència de les característiques d'un fitxer en el temps d'impressió d'aquest, amb l'objectiu d'optimitzar els processos implicats. El projecte té l'objectiu de veure la influència de determinats factors en el temps de processat al driver d'un fitxer, abans d'enviar-lo a la impressora, a través del desenvolupament d'un model de regressió lineal i una anàlisi de clústers. Aquests dos enfocaments proporcionaran resultats que caldrà interpretar de manera diferent i que es complementaran entre ells. Aquesta tesi explica el procés de creació d'aquestes dues investigacions: l'obtenció de les dades, el tractament d'aquestes, la creació del model predictiu i de l'anàlisi, i la comparació dels resultats. De manera addicional, també s'ha desenvolupat una eina que permeti seleccionar un fitxer i obtenir-ne el temps d'impressió, integrant els codis d'obtenció de característiques i model de regressió lineal. Finalment, hi ha una validació d'aquesta eina per part de l'empresa i companys, on es comprova la seva bona funcionalitat i utilitat per als desenvolupadors de l'empresa. També es proposen aplicacions i idees de millora, per tal de fer l'eina més útil i precisa., This report describes the work carried out for the technological company HP, which specializes in manufacturing and selling a wide range of technological products, including printers. HP wanted to study the influence of a file's characteristics on its printing time, with the aim of optimizing the processes involved. The project's goal is to examine the influence of certain factors on the processing time at the driver of a file, before sending it to the printer, through the development of a linear regression model and a cluster analysis. These two approaches will provide results that will need to be interpreted differently and will complement each other. This thesis explains the process of creating these two investigations: the acquisition of data, the processing of this data, the creation of the predictive model and analysis, and the comparison of results. Additionally, a tool has also been developed that allows a user to select a file and obtain its printing time, integrating the codes for feature extraction and linear regression model. Finally, there is a validation of this tool by the company and colleagues, where its good functionality and usefulness for the company's developers is verified. Also, applications and improvement ideas are proposed, in order to make the tool more useful and accurate.

As an individual who has collected training activity data through Strava and Garmin APIs over the years, a substantial database has been amassed containing valuable information that has the potential to enhance performance and achieve athletic goals. Through personal observation, it has been recognized that optimizing training is not solely about exerting maximum effort daily or pushing the body to its limits. Instead, a balanced routine encompassing various training types and appropriate rest periods is crucial. While these observations are based on personal intuition, it is reasonable to assume that patterns exist within the aforementioned database, enabling the inference of current physical condition and the development of an optimal training routine geared towards personal performance improvement. Furthermore, like any extensive database, this dataset contains latent information that can be further explored and comprehended through visualizations and analytical techniques. Leveraging the power of data science, it is possible to delve deeper into the data, uncovering trends, patterns, and correlations that may not be immediately apparent. Such insights can then be employed to adjust training regimens and fine-tune approaches to cater to individual needs more effectively. The main objective of the project is to develop a software system application that allows any user to upload the data of their training activities and use it for inferring their fitness status. Thus, it allows any athlete who has registered their training activities to upload it and see the consequences of their weekly workouts on the performance for the following week, along with other relevant results.

With the immense progress in Machine Learning in the past decades, General Machine Learning(GLM) models have increased to perform many different tasks such, grammar correction , classification of items or explain pieces of code. One of these new fields that GLM models can achieve is the Automated Program Repair(APR) helping developers by generating patches to automatically repair software bugs. Typically these APR techniques are based on a pre-trained neural network mode, trained by pairs of codes with errors and their correction. These models take as input a buggy code and return a regenerated code with the defects repaired. Nowadays these are one of the most effective techniques for automatic correction, but these DeepLearning-based APR approaches still present many errors on a large portion of the codes given to the models. Alternatively Code language models (CLMs), has become a good solution when APR problems are involved, this CLMs differ from the DeepLearing(DL) ones because CLMS are not trained from a set of buggy codes and correct ones, these models are trained from a huge source of unlabeled codes for general code language modeling tasks. The aim of this research is to know the capabilities and the shortcomings of Code Language Models in the automatic program repair field and see how developers can take advantage of these new techniques and use them in order to improve their programming skills, saving time and money

A medida que la tecnología sigue evolucionando a una velocidad exponencial, las empresas luchan por mantenerse al día con la aparición de nuevas tecnologías, lo que inevitablemente les lleva a tener uno o más sistemas heredados. Las industrias muy restringidas, como la química, se ven especialmente afectadas por este problema, ya que cada cambio es más largo y difícil de realizar. Las empresas hacen todo lo posible por aprovechar las tecnologías y técnicas de diseño modernas durante la fase de diseño para prolongar la vida útil de sus productos. Sin embargo, estos esfuerzos resultan insignificantes si el producto no se mantiene correctamente a lo largo del tiempo. Este fue el enfoque usado cuando se construyó RAD hace veinticinco años, una aplicación web diseñada para ayudar a medir la eficiencia y la variabilidad de una planta química a partir de datos relacionados con la producción. Se construyó utilizando las tecnologías más novedosas de la época, pero no se mantuvo adecuadamente y poco a poco se transformó en un sistema heredado. Por ello, una gran variedad de problemas surgieron respecto su capacidad de mantenimiento, ampliación y compatibilidad. Se prohibió la modificación y extensión del sistema debido a las malas decisiones de diseño tomadas en la fase inicial. Esto provocó que no se pudiera satisfacer las necesidades actuales de los usuarios. Además, la aplicación web no funcionaba correctamente en el navegador de preferencia, Google Chrome. Por último, y lo más urgente, se ha anunciado recientemente que el servidor Windows donde se almacena el sistema heredado se renovará este año. Esto supone un riesgo importante para la funcionalidad de RAD, ya que el sistema heredado no funcionará correctamente en el nuevo servidor Windows. Por estos motivos, el equipo de desarrollo se sintió apresurado para revivir el sistema heredado antes de que fuera demasiado tarde, y aquí es donde entra en juego este trabajo de final de grado. El objetivo de este trabajo de final, As technology continues to evolve at an exponential pace, companies struggle to keep up with the appearance of new technologies, which inevitably leads them to have one or more legacy systems. Highly restricted industries, such as the chemical industry, are especially affected by this issue, as every change is longer and harder to do. Organizations try their best to leverage modern technologies and design techniques during the design phase to prolong the lifespan of their products. However, these efforts become negligible if the product is not maintained correctly over time. This was the approach taken when building RAD twenty-five years ago, a web application designed to help measure the efficiency and variability of a chemical plant based on production-related data. It was built using the newest technologies of the time, however, it wasn't maintained properly, and it slowly transformed into a legacy system. This brought a wide variety of problems in regard to maintainability, extendability, and compatibility. The web application became prohibited from any further modification or extension due to the poor design choices made in the initial design phase. This lead to not being able to meet the up-to-date needs of the users. In addition, the web application wouldn't function correctly on the preferred web browser, Google Chrome. Lastly, and most importantly, it recently got announced that the Windows server where the legacy system is hosted will be renewed this year. This poses a significant risk to the correct functionality of RAD as the legacy system won't function correctly on the new Windows server. For these reasons, the development team was urged to revive the legacy system before it was too late, which is where this thesis comes in. The aim of this thesis is to help the development team solve these issues with the legacy system in the desired timeframe of this year. This will be done by re-implementing from scratch its core features, the operations, using the

Due to recent industry needs, synergies between data science and software engineering are starting to be present in data science and engineering academic programs. Two synergies are: applying data science to manage the quality of the software (software analytics) and applying software engineering best practices in data science projects to ensure quality attributes such as maintainability and reproducibility. The lack of these synergies on academic programs have been argued to be an educational problem. Hence, it becomes necessary to explore how to teach software analytics and software engineering best practices in data science programs. In this context, we provide hands-on for conducting laboratories applying project-based learning in order to teach software analytics and software engineering best practices to data science students. We aim at improving the software engineering skills of data science students in order to produce software of higher quality by software analytics. We focus in two skills: following a process and software engineering best practices. We apply project-based learning as main teaching methodology to reach the intended outcomes. This teaching experience shows the introduction of project-based learning in a laboratory, where students applied data science and best software engineering practices to analyze and detect improvements in software quality. We carried out a case study in two academic semesters with 63 data science bachelor students. The students found the synergies of the project positive for their learning. In the project, they highlighted both utility of using a CRISP-DM data mining process and best software engineering practices like a software project structure convention applied to a data science project., This paper was partly funded by a teaching innovation project of ICE@UPC-BarcelonaTech (entitled ‘‘Audiovisual and digital material for data engineering, a teaching innovation project with open science’’), and the ‘‘Beatriz Galindo’’ Spanish Program BEA-GAL18/00064., Peer Reviewed, Postprint (published version)

Software organisations aim to develop and maintain high-quality software systems. Due to large amounts of behaviour data available, software organisations can conduct data-driven software maintenance. Indeed, software quality assurance and improvement programs have attracted many researchers' attention. Bayesian Networks (BNs) are proposed as a log analysis technique to discover poor performance indicators in a system and to explore usage patterns that usually require temporal analysis. For this, an action research study is designed and conducted to improve the software quality and the user experience of a web application using BNs as a technique to analyse software logs. To this aim, three models with BNs are created. As a result, multiple enhancement points have been identified within the application ranging from performance issues and errors to recurring user usage patterns. These enhancement points enable the creation of cards in the Scrum process of the web application, contributing to its data-driven software maintenance. Finally, the authors consider that BNs within quality-aware and data-driven software maintenance have great potential as a software log analysis technique and encourage the community to deepen its possible applications. For this, the applied methodology and a replication package are shared., Junta de Andalucía, Grant/Award Number: P20‐00091; AEI, Grant/Award Number: PID2019‐106758GB‐32/AEI/10.13039/501100011033; Spanish project, Grant/Award Number: PDC2021‐121195‐I00; Spanish Program, Grant/Award Number: BEAGAL18/00064, Peer Reviewed, Postprint (published version)

The evaluation of Deep Learning (DL) models has traditionally focused on criteria such as accuracy, F1 score, and related measures. The increasing availability of high computational power environments allows the creation of deeper and more complex models. However, the computations needed to train such models entail a large carbon footprint. In this work, we study the relations between DL model architectures and their environmental impact in terms of energy consumed and CO2 emissions produced during training by means of an empirical study using Deep Convolutional Neural Networks. Concretely, we study: (i) the impact of the architecture and the location where the computations are hosted on the energy consumption and emissions produced; (ii) the trade-off between accuracy and energy efficiency; and (iii) the difference on the method of measurement of the energy consumed using software-based and hardware-based tools., This work has been supported by the Spanish project PID2020-117191RB-I00 funded by MCIN/AEI/10.13039/501100011033., Peer Reviewed, Postprint (published version)

L'aprenentatge automàtic, un tipus particular d'intel·ligència artificial, s'utilitza cada vegada més en els sistemes de software modern. A més de requerir molta energia, el seu ús augmenta les emissions de CO2, té un elevat cost financer i consumeix molts recursos. Encara que existeixen nombrosos estudis que analitzen les possibilitats del ML, molt pocs d'ells se centren en les seves característiques ecològiques, com l'ús d'energia o la seva petjada de carboni. Aquest estudi tracta d'augmentar la conscienciació dels costos associats al desplegament d'un sistema de ML en algunes de les plataformes d'allotjament en el núvol més avançades. Per a això, duem a terme un ampli estudi empíric per a avaluar i contrastar les emissions i la latencia de quatre proveïdors de núvol diferents, Amazon Web Services, Azure, Heroku i Railway, Machine learning, a particular type of artificial intelligence, is increasingly being used in modern software systems. In addition to requiring a lot of energy, its use raises CO2 emissions, has a high financial cost, and consumes many resources. Even while there are numerous studies that look at the possibilities of ML, very few of them concentrate on its green features, like energy consumption or its carbon footprint. This bachelor thesis tries to increase understanding of the costs associated with deploying an ML system into some of the state-of-the-art cloud hosting platforms. In order to do this, we conduct an extensive empirical study to assess and contrast the emissions and latency of four different cloud providers: Amazon Web Services, Azure, Heroku, and Railway.

The main goal of this study is to analyze the impact of the model architecture and training strategy when building green computer vision systems. We divide this goal into two research questions. In the first one, we analyze the impact of the model architecture on achieving greener models while keeping accuracy at optimal levels. In the second one, we study the influence of training strategy on producing greener systems.

To remain competitive, organizations are challenged to make informed design decisions in order to construct software systems with similar architectural needs. They use reference architectures to achieve interoperability of (parts of) their software, standardization of software systems among multiple actors, and faster development with templates and guidelines for designing. In this chapter, we give an overview of domain-independent reference architectures and standards available in the gray literature for practitioners, classified by technology (cloud computing, big data, Internet of Things, and artificial intelligence) and analyzed based on their characteristics (e.g., purpose and contents). The discussed reference architectures are either published by standardization organizations such as ISO/IEC and NIST or fostered by large companies such as Microsoft and IBM., This work has been partially funded by the “Beatriz Galindo” Spanish Program BEAGAL18/00064 and by the DOGO4ML Spanish research project (ref. PID2020-117191RB-I00)., Peer Reviewed, Postprint (author's final draft)

Architecting AI-based systems entails making some decisions that are particular to this type of systems. Therefore, it becomes necessary to gather all necessary knowledge to inform such decisions, and to articulate this knowledge in a form that facilitates knowledge transfer among different AI projects. In this exploratory paper, we first present the main results of a literature survey in the area, and then we propose a preliminary ontology for architectural decision making, which we exemplify using a subset of the papers selected in the literature review. In the discussion, we remark on the variety of decision types and system contexts, highlighting the need to further investigate the current state of research and practice in this area. Besides, we summarize our plans to move along this research area by widening the literature review and incorporating more AI-related concepts to this first version of the ontology., This paper has been funded by the Spanish Ministerio de Ciencia e Innovación under project / funding scheme PID2020-117191RB-I00 / AEI/10.13039/501100011033., Peer Reviewed, Postprint (author's final draft)

Un dels perills contra els quals la majoria de governs estan lluitant avui dia és l'escalfament global, el qual amenaça la qualitat de vida dels diferents éssers vius que habiten el planeta. Aquest és causat per la gran quantitat de gasos d'efecte hivernacle que s'emeten en el dia a dia, els quals provenen de diferents fonts. Un dels sectors que més participa en l'emissió d'aquests gasos és la producció d'energia. En els darrers anys, s'han desenvolupat diferents solucions per reduir les emissions que es formen de la producció d'energia, ja sigui augmentant el nombre de fonts d'energia renovable o bé disminuint el consum energètic que diferents elements del nostre dia a dia realitzen. En aquest treball s'ha desenvolupat una aplicació web per poder calcular l'eficiència energètica tant d'edificis com de sistemes software i permet a l'usuari visualitzar-ne la classificació energètica de diferents característiques d'aquests. D'aquesta manera, l'aplicació permet a un usuari conèixer en quins aspectes el producte flaqueja en quant a eficiència energètica. A més, aquesta permet obtenir un informe amb els resultats dels càlculs realitzats. Aquesta està formada per un backend desenvolupat amb el framework Django i un frontend desenvolupat amb Flutter. A més, aquest projecte utilitza la metodologia Scrum. L'àmbit en què es troba aquesta aplicació encara està en desenvolupament, per la qual cosa permet tenir una alternativa a les que actualment existeixen. A més, per part dels edificis aquesta aporta un càlcul senzill de fer però que alhora mostra informació important sobre aquests. D'altra banda, per als sistemes software permet conèixer el consum que aquest realitza al nostre ordinador quan és executat., One of the dangers that most governments are struggling with today is global warming, which threatens the quality of life of the various living things that inhabit the planet. This is caused by the large amount of greenhouse gases that are emitted on a daily basis, which come from different sources. One of the sectors most involved in the emission of these gases is energy production. In recent years, different solutions have been developed to reduce the emissions that are formed from energy production, either by increasing the number of renewable energy sources or by reducing the energy that different elements consume. The application developed allows to calculate the energy efficiency of both buildings and software systems and allows the user to view the energy classification of different characteristics of these. As a result, the application allows a user to know in which aspects their product is weakening in terms of energy efficiency. Moreover, it allows to get an report with the results obtained. This applicaction consists of a backend developed with the Django framework and a frontend developed with Flutter. In addition, this project uses the Scrum methodology. The area in which this application is located is still under development, so this application can be an alternative to those that currently exist. In addition, on the part of the buildings it provides a simple calculation to perform but at the same time shows important information about them. On the other hand, for the software systems it allows to know the consumption that this one makes in our computer when it is executed.

AI-based systems are software systems with functionalities enabled by at least one AI component (e.g., for image-, speech-recognition, and autonomous driving). AI-based systems are becoming pervasive in society due to advances in AI. However, there is limited synthesized knowledge on Software Engineering (SE) approaches for building, operating, and maintaining AI-based systems. To collect and analyze state-of-the-art knowledge about SE for AI-based systems, we conducted a systematic mapping study. We considered 248 studies published between January 2010 and March 2020. SE for AI-based systems is an emerging research area, where more than 2/3 of the studies have been published since 2018. The most studied properties of AI-based systems are dependability and safety. We identified multiple SE approaches for AI-based systems, which we classified according to the SWEBOK areas. Studies related to software testing and software quality are very prevalent, while areas like software maintenance seem neglected. Data-related issues are the most recurrent challenges. Our results are valuable for: researchers, to quickly understand the state-of-the-art and learn which topics need more research; practitioners, to learn about the approaches and challenges that SE entails for AI-based systems; and, educators, to bridge the gap among SE and AI in their curricula., This work has been partially funded by the “Beatriz Galindo” Spanish Program BEAGAL18/00064 and by the DOGO4ML Spanish research project (ref. PID2020-117191RB-I00), Peer Reviewed, Postprint (author's final draft)

Este proyecto consiste en el desarrollo de dos sistemas de software para la empresa de iluminación Vibia. Primero, un gestor de incidencias que servirá para que sus clientes puedan encontrar soluciones a sus problemas sin la necesidad de contactar con un agente, agilizando así el proceso. Segundo, un gestor de contenidos donde se almacenarán todos los archivos y recursos usados por la empresa en todas sus páginas webs, incluida la desarrolla en este proyecto, clasificados y organizados por etiquetas. Esta memoria incluye todos los pasos a seguir, desde la planificación y la gestión hasta la implementación y la validación, necesarios para desarrollar estos dos sistemas., This project consists of the development of two software systems for the lighting company Vibia. First, an incident manager that will help their clients find solutions to their problems without the need to contact an agent, thus speeding up the process. Second, a content manager where all the files and resources used by the company on all its web pages will be stored, including the one developed in this project, classified and organized by tags. This report includes all the steps to follow, from planning and management to implementation and validation, necessary to develop these two systems.

El Machine Learning (o ML) es una de las ramas principales de la ingeniería basada en un aprendizaje automático donde las máquinas aprenden a detectar patrones en los datos para hacer predicciones. El ML es una de las tecnologías más novedosas y prometedoras actualmente, pero a su vez existen aún muchas incógnitas que resolver para llevar esta tecnología a su máximo exponente. En este proyecto se pretende integrar modelos de machine learning en una aplicación donde se nos permita detectar y clasificar la personalidad humana en función de mensajes de texto escritos por los propios usuarios. A su vez se realizará un estudio comparativo del efecto que puede llegar a tener el despliegue de dicha aplicación en una plataforma u otra, evaluando factores como el rendimiento o la escalabilidad., Machine Learning (or ML) is one of the main branches of engineering based on machine learning where machines learn to detect patterns in data to make predictions. ML is one of the newest and most promising technologies at present, but at the same time there are still many unknowns to be solved in order to take this technology to its maximum exponent. This project aims to integrate machine learning models in an application that allows us to detect and classify human personality based on text messages written by the users themselves. At the same time, a comparative study will be carried out on the effect that the deployment of this application on one platform or another may have, evaluating factors such as performance and scalability.

La Universitat Politècnica de Catalunya (UPC), de la mateixa manera que la resta d'universitats importants a escala mundial, és una institució amb una gran estructura, la qual està composta de moltes peces, essencialment de persones. Doncs, és crucial tenir una bona gestió de tota aquesta perquè l'organització funcioni correctament. Actualment, aquesta gestió es realitza a través dels seus diferents òrgans de govern, els quals s'encarreguen, tant de forma global, com a nivell més específic, de prendre les decisions. És per això que, avui dia, la universitat disposa d'una sèrie de sistemes que s'encarreguen de l'administració de tots els seus òrgans de govern, tant pel que fa a la seva organització com a nivell de presa de decisions. Aquests sistemes, però, en l'àmbit organitzatiu, requereixen cert temps per a mantenir les seves dades, degut a que aquestes han de ser afegides o modificades una per una de manera manual, el que ha arribat a suposar un problema, tenint dades al sistema que no corresponen amb l'actualitat. Doncs, en aquest projecte, s'ha realitzat el desenvolupament d'una aplicació web que permet un manteniment de totes les dades corresponents a l'estructura de qualsevol òrgan UPC de manera àgil i fàcil, la qual a més a més s'encarrega d'actualitzar algunes de les dades de manera automàtica, podent així disposar d'informació real en tot moment sense haver d'invertir el temps que comporten els sistemes actuals., The Polytechnic University of Catalonia (UPC), like any other major university worldwide, is an institution with a pretty large structure, that is made up of many parts, essentially people. Therefore, it is really important to have a good management of all of it for the organization to work correctly. Presently, this management is carried out through its different governing bodies, which are responsible, both globally and at a more specific level, for taking the decisions across the institution. This is why, nowadays, the university has a series of systems that are responsible for the administration of all its governing bodies, both at the structure level and at the decision-making level. These systems, however, especially at the organizational level, require a certain amount of time to maintain their data, as these have to be added or modified one by one manually, which has become a problem, as the data in the system does not correspond to reality. This is why, in this project, we have developed a web application that allows the maintenance of all the data corresponding to the structure of any UPC body in an agile and easy way, which is also responsible for updating part of the data automatically, thus being able to have real information at all times without having to invest the time involved in the current systems.

Machine Learning based Software Systems (MLSS) are becoming increasingly pervasive in today’s society and can be found in virtually every domain. Building MLSS is challenging due to their interdisciplinary nature. MLSS engineering encompasses multiple disciplines, of which Data Engineering and Software Engineering appear as most relevant. The DOGO4ML project aims at reconciling these two disciplines for providing a holistic end-to-end framework to develop, operate and govern MLSS and their data. It proposes to combine and intertwine two software cycles: the DataOps and the DevOps lifecycles. The DataOps lifecycle manages the complexity of dealing with the big data needed by ML models, while the DevOps lifecycle is in charge of building the system that embeds these models. In this paper, we present the main vision and goals of the project as well as its expected contributions and outcomes. Although the project is in its initial stage, the progress of the research undertaken so far is detailed., This paper has been funded by the Spanish Ministerio de Ciencia e Innovación under project / funding scheme PID2020-117191RB-I00 / AEI/10.13039/501100011033., Peer Reviewed, Postprint (published version)

Aquest projecte consisteix a desenvolupar un software per entitats de lleure que s'encarrega de la gestió del personal (monitors, participants i responsable), equips i activitats. L'objectiu és poder reduir els recursos físics que les entitats utilitzen per complir amb la normativa vigent. S'ha construït des de zero una aplicació mòbil i web per als monitors i un servidor format per una APIRest connectat a una base de dades relacional. L'aplicació està pensada per la consulta de recursos i la web per l'administració. La memòria inclou tots els passos a seguir per desenvolupar el sistema, des de la planificació i la gestió fins a la implementació i la validació., This project consists of depeloping a software for the management of the staff (counselours, participants, heads) in entities that organize leisure activities. The objective is to reduce the physical resources used by these entities to follow current regulations. A mobile and web applications have been developed from scratch for the counselours and a server format for an APITest connected to a relational database. The mobile App is made to look up resources and the web for administration. The report includes every step to develop the system, from planning and management to implementation and validation.

L'objectiu d'aquest projecte és facilitar el desenvolupament de projecte sobre Machine Learning per a que pugui ser entenedor per a tothom relacionat amb aquest. Per a realitzar això, s'ha desenvolupat una plantilla que conté un decorador introduïnt MLFlow per a la monitorització d'aquests projectes, així com un seguit de mecanismes que fa que executar-lo estigui a l'abast de tothom. Finalment, s'han duit a terme un seguit de proves amb el projecte inicial juntament amb la plantilla final generada., The objective of this purpose is to facilitate the development of projects related to Machine Learning so that they can be understood by everyone related to it. To do this, a template has been developed that contains a decorator introducing MLFlow for monitoring these projects, as well as a series of mechanisms that allow anyone to run it. Finally, some tests have been carried out with the initial project together with the final generated template.