Molina Ramirez, Alejandro (2021)
Deep Networks That Know When They Don't Know.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00018525
Dissertation, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
Machine Learning (ML) and Artificial Intelligence (AI) are more present than ever in our society's collective discourse. CEOs, politicians, and fellow citizens all put incredibly high hopes and expectations into AI's future capabilities. In many applications, ranging from the medical field to autonomous robots such as self-driving cars, we are starting to entrust human lives to decisions made by algorithms and machines. With credit scoring algorithms and hiring practices now adopting these new technologies, machine learning can have a profound impact on people’s lives. The expectation of inherent fairness, accuracy, and consistency we have of these algorithms goes beyond even what we expect from fellow humans. Indeed, these expectations are driven by the desire to improve everyone’s quality of life.
Many current machine learning models focus on providing the highest possible accuracy. However, these models are often black boxes that are hard to examine. They are mostly discriminative models that focus on modeling decisions based on the training data, but do not create a model for the data itself. This is important, as we are interested in questioning the training data to detect systematic biases. Furthermore, we are also highly interested in asking the model whether the current data it is processing fits the training data. In other words, is it qualified to make decisions and "knows what it is talking about", or whether it simply "does not know". Therefore, we require a generative model that can answer these, and other, questions. In this thesis, we focus on deep generative models based on probabilistic circuits; a family of statistical models that allows us to answer a wide range of normalized probability queries with guarantees on computational time. We can then ask these generative models about biases, including how confident they are about a particular answer, as they "know when they do not know".
We develop models for count data, extend them to non-parametric models, and models based on dictionaries of distributions. They cover a large variety of use-cases. We then make connections to Deep Neural Networks and show how to build generative models from them with inference guarantees. All these models cover a wide range of use cases, including hybrid domains. Moreover, we present a model that learns from the data making most decisions automatically so that non-experts can also benefit from these powerful tools. This will contribute to the democratization of machine learning.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2021 | ||||
Autor(en): | Molina Ramirez, Alejandro | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Deep Networks That Know When They Don't Know | ||||
Sprache: | Englisch | ||||
Referenten: | Kersting, Prof. Dr. Kristian ; Natarajan, Prof. Dr. Sriraam | ||||
Publikationsjahr: | 2021 | ||||
Ort: | Darmstadt | ||||
Kollation: | xxv, 202 Seiten | ||||
Datum der mündlichen Prüfung: | 23 April 2021 | ||||
DOI: | 10.26083/tuprints-00018525 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/18525 | ||||
Kurzbeschreibung (Abstract): | Machine Learning (ML) and Artificial Intelligence (AI) are more present than ever in our society's collective discourse. CEOs, politicians, and fellow citizens all put incredibly high hopes and expectations into AI's future capabilities. In many applications, ranging from the medical field to autonomous robots such as self-driving cars, we are starting to entrust human lives to decisions made by algorithms and machines. With credit scoring algorithms and hiring practices now adopting these new technologies, machine learning can have a profound impact on people’s lives. The expectation of inherent fairness, accuracy, and consistency we have of these algorithms goes beyond even what we expect from fellow humans. Indeed, these expectations are driven by the desire to improve everyone’s quality of life. Many current machine learning models focus on providing the highest possible accuracy. However, these models are often black boxes that are hard to examine. They are mostly discriminative models that focus on modeling decisions based on the training data, but do not create a model for the data itself. This is important, as we are interested in questioning the training data to detect systematic biases. Furthermore, we are also highly interested in asking the model whether the current data it is processing fits the training data. In other words, is it qualified to make decisions and "knows what it is talking about", or whether it simply "does not know". Therefore, we require a generative model that can answer these, and other, questions. In this thesis, we focus on deep generative models based on probabilistic circuits; a family of statistical models that allows us to answer a wide range of normalized probability queries with guarantees on computational time. We can then ask these generative models about biases, including how confident they are about a particular answer, as they "know when they do not know". We develop models for count data, extend them to non-parametric models, and models based on dictionaries of distributions. They cover a large variety of use-cases. We then make connections to Deep Neural Networks and show how to build generative models from them with inference guarantees. All these models cover a wide range of use cases, including hybrid domains. Moreover, we present a model that learns from the data making most decisions automatically so that non-experts can also benefit from these powerful tools. This will contribute to the democratization of machine learning. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
Status: | Verlagsversion | ||||
URN: | urn:nbn:de:tuda-tuprints-185251 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Künstliche Intelligenz und Maschinelles Lernen |
||||
Hinterlegungsdatum: | 11 Mai 2021 10:15 | ||||
Letzte Änderung: | 18 Mai 2021 07:13 | ||||
PPN: | |||||
Referenten: | Kersting, Prof. Dr. Kristian ; Natarajan, Prof. Dr. Sriraam | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 23 April 2021 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |