Analyzing state-of-the-art CNN’s explainability focusing on food classification

Radeva, PetiaRódenas Cumplido, JavierBergadà Salsen, Joan2022-02-162022-02-162020-06-21https://hdl.handle.net/2445/183203Treballs Finals de Grau d'Enginyeria Informàtica, Facultat de Matemàtiques, Universitat de Barcelona, Any: 2020, Director: Petia Radeva i Javier Ródenas Cumplido[en] Convolutional Neural Networks (CNNs) are Deep Learning algorithms that can be applied to a wide range of environments with a high performance. We will study all the elements that form the CNNs, learn how they work, how to train them and how to diagnose its performance. We will also work the state-of-the-art visual explainability algorithms developing how they create their heatmaps. In the practical part we will use CNNs in order to classify food images into their classes and we will apply the studied explainability algorithms to understand the predictions made by the neural networks. Finally, we will perform a both qualitative and quantitative comparison between the explanations given by the applied algorithms.80 p.application/pdfcatmemòria: cc-nc-nd (c) Joan Bergadà Salsen, 2020http://creativecommons.org/licenses/by-nc-nd/3.0/es/Xarxes neuronals convolucionalsReconeixement òptic de formesProgramariTreballs de fi de grauVisió per ordinadorAprenentatge automàticAlimentsConvolutional neural networksOptical pattern recognitionComputer softwareComputer visionMachine learningBachelor's thesesFoodAnalyzing state-of-the-art CNN’s explainability focusing on food classificationinfo:eu-repo/semantics/bachelorThesisinfo:eu-repo/semantics/openAccess