東華大學圖書館 |

Towards Explainable Artificial Intelligence : = Interpreting Neural Network Classifiers with Probabilistic Prime Implicants = Zu Erklarbarer Kunstlicher Intelligenz.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Towards Explainable Artificial Intelligence :/
其他題名:	Interpreting Neural Network Classifiers with Probabilistic Prime Implicants = Zu Erklarbarer Kunstlicher Intelligenz.
其他題名:	Zu Erklarbarer Kunstlicher Intelligenz.
作者:	Waldchen, Stephan.
面頁冊數:	1 online resource (207 pages)
附註:	Source: Dissertations Abstracts International, Volume: 84-05, Section: B.
Contained By:	Dissertations Abstracts International84-05B.
標題:	Neural networks. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29780938click for full text (PQDT)
ISBN:	9798352976272

Towards Explainable Artificial Intelligence : = Interpreting Neural Network Classifiers with Probabilistic Prime Implicants = Zu Erklarbarer Kunstlicher Intelligenz.
Waldchen, Stephan.

Towards Explainable Artificial Intelligence :Interpreting Neural Network Classifiers with Probabilistic Prime Implicants = Zu Erklarbarer Kunstlicher Intelligenz.Zu Erklarbarer Kunstlicher Intelligenz. - 1 online resource (207 pages)

Source: Dissertations Abstracts International, Volume: 84-05, Section: B.

Thesis (Ph.D.)--Technische Universitaet Berlin (Germany), 2022.

Includes bibliographical references

In this thesis we develop a framework for interpreting the decisions of highly nonlinear classifer functions with a focus on neural networks. Specifcally, we formalise the idea of separating the input parameters into relevant and irrelevant ones as an explicit optimisation problem.First, we describe what is generally understood as a relevance map for classifers and give an overview over the existing methods to produce such maps. We explain how we used relevance maps to detect artefacts in the PASCAL VOC dataset and track the focus of neural agents playing the Atari video games Breakout and Pinball on a human-like level.Towards a formal defnition of relevance maps, we generalise the concept of prime implicants from abductive logic to a probabilistic setting by introducing δ-relevant sets. For a d-ary Boolean function Φ: {0, 1} d → {0, 1} and an assignment to its variables x = (x1, x2, . . . , xd) we consider the problem of fnding those subsets of the variables that are sufcient to determine the function output with a given probability δ. We show that the problem of fnding small δ-relevant sets is NP-hard to approximate with a factor d 1−α for α > 0. The associated decision problem turns out to be NPPP-complete.We further generalise δ-relevant sets from the binary to the continuous domain. This leads naturally to a rate-distortion trade-of between the size of the δ-relevant set (rate) and the change in the classifer prediction (distortion). Relevance maps can then be interpreted as greedy approximations of the rate-distortion function. Evaluating this function even approximately turns out to be NP-hard, so we develop a heuristic solution strategy based convex relaxation of the combinatorial problem and assumed density fltering (ADF) for deep ReLU neural networks. This results in our own explanation method which we call Rate-Distortion Explanations (RDE).To show that the approximations in ADF are necessary, we give a complete characterisation of families of probability distributions that are invariant under the action of ReLU neural network layers. We demonstrate that the only invariant families are either degenerate or amount to sampling.Subsequently, we propose and discuss several benchmark tests and numerical evaluation methods for relevance maps. We compare RDE to a representative collection of established relevance methods and demonstrate that it outperforms competitors for a wide range of tasks.Finally, we discuss how the knowledge of the true data distribution is crucial for any existing explanation method. We criticise our own method over potential artefacts and introduce a stronger, information theoretical requirement based on the conditional entropy. A novel approach, called Arthur-Merlin-regularisation along with a new framework is developed. The framework is then extended to realistic algorithms and data sets, and we discuss under which assumptions the guarantees still hold.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798352976272Subjects--Topical Terms:

677449
Neural networks.
Index Terms--Genre/Form:

542853
Electronic books.

Towards Explainable Artificial Intelligence : = Interpreting Neural Network Classifiers with Probabilistic Prime Implicants = Zu Erklarbarer Kunstlicher Intelligenz.
LDR:07867nmm a2200373K 4500 001 2355063
005 20230515064601.5
006 m o d
007 cr mn ---uuuuu
008 241011s2022 xx obm 000 0 eng d
020 $a 9798352976272
035 $a (MiAaPQ)AAI29780938
035 $a (MiAaPQ)TechBerlin_1130317230
035 $a AAI29780938
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Waldchen, Stephan. $3 3695453
245 1 0 $a Towards Explainable Artificial Intelligence : $b Interpreting Neural Network Classifiers with Probabilistic Prime Implicants = Zu Erklarbarer Kunstlicher Intelligenz.
246 3 1 $a Zu Erklarbarer Kunstlicher Intelligenz.
264 0 $c 2022
300 $a 1 online resource (207 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 84-05, Section: B.
500 $a Advisor: Pokutta, Sebastian ; Stannat, Wilhelm ; Skutella, Martin ; Petersen, Philipp ; Marques-Silva, Joao.
502 $a Thesis (Ph.D.)--Technische Universitaet Berlin (Germany), 2022.
504 $a Includes bibliographical references
520 $a In this thesis we develop a framework for interpreting the decisions of highly nonlinear classifer functions with a focus on neural networks. Specifcally, we formalise the idea of separating the input parameters into relevant and irrelevant ones as an explicit optimisation problem.First, we describe what is generally understood as a relevance map for classifers and give an overview over the existing methods to produce such maps. We explain how we used relevance maps to detect artefacts in the PASCAL VOC dataset and track the focus of neural agents playing the Atari video games Breakout and Pinball on a human-like level.Towards a formal defnition of relevance maps, we generalise the concept of prime implicants from abductive logic to a probabilistic setting by introducing δ-relevant sets. For a d-ary Boolean function Φ: {0, 1} d → {0, 1} and an assignment to its variables x = (x1, x2, . . . , xd) we consider the problem of fnding those subsets of the variables that are sufcient to determine the function output with a given probability δ. We show that the problem of fnding small δ-relevant sets is NP-hard to approximate with a factor d 1−α for α > 0. The associated decision problem turns out to be NPPP-complete.We further generalise δ-relevant sets from the binary to the continuous domain. This leads naturally to a rate-distortion trade-of between the size of the δ-relevant set (rate) and the change in the classifer prediction (distortion). Relevance maps can then be interpreted as greedy approximations of the rate-distortion function. Evaluating this function even approximately turns out to be NP-hard, so we develop a heuristic solution strategy based convex relaxation of the combinatorial problem and assumed density fltering (ADF) for deep ReLU neural networks. This results in our own explanation method which we call Rate-Distortion Explanations (RDE).To show that the approximations in ADF are necessary, we give a complete characterisation of families of probability distributions that are invariant under the action of ReLU neural network layers. We demonstrate that the only invariant families are either degenerate or amount to sampling.Subsequently, we propose and discuss several benchmark tests and numerical evaluation methods for relevance maps. We compare RDE to a representative collection of established relevance methods and demonstrate that it outperforms competitors for a wide range of tasks.Finally, we discuss how the knowledge of the true data distribution is crucial for any existing explanation method. We criticise our own method over potential artefacts and introduce a stronger, information theoretical requirement based on the conditional entropy. A novel approach, called Arthur-Merlin-regularisation along with a new framework is developed. The framework is then extended to realistic algorithms and data sets, and we discuss under which assumptions the guarantees still hold.
520 $a In dieser Arbeit entwickeln wir ein Framework fur die Interpretation der Entscheidungen hochgradig nichtlinearer Klassifizierungsfunktionen mit Schwerpunkt auf neuronalen Netzen. Insbesondere formalisieren wir die Idee der Trennung der Eingabeparameter in relevante und irrelevante Parameter als explizites Optimierungsproblem.Zunachst beschreiben wir, was man allgemein unter einer Relevanzkarte fur Klassifikatoren versteht, und geben einen Uberblick uber die bestehenden Methoden zur Erstellung solcher Karten. Wir erlautern, wie wir Relevanzkarten verwendet haben, um Artefakte im PASCAL-VOC-Datensatz zu finden und um den Fokus neuronaler Agenten beim Spielen der Atari-Videospiele Breakout und Pinball auf einer menschenahnlichen Ebene zu verfolgen.Auf dem Weg zu einer formalen Definition von Relevanzkarten verallgemeinern wir das Konzept der primaren Implikanten aus der abduktiven Logik auf eine probabilistische Umgebung, indem wir δ-relevante Mengen einfuhren. Fur eine d-are boolesche Funktion Φ : {0, 1}d → {0, 1} und eine Zuordnung zu ihren Variablen x = (x1, x2, . . . , xd) betrachten wir das Problem, diejenigen Teilmengen der Variablen zu finden, die ausreichen, um die Funktionsausgabe mit einer gegebenen Wahrscheinlichkeit δ zu bestimmen. Wir zeigen, dass das Problem, kleine δ-relevante Mengen zu finden, NP-schwer mit einem Faktor d1−α fur α > 0 zu approximieren ist. Das zugehorige Entscheidungsproblem erweist sich als NPPP-vollstandig.Wir verallgemeinern ferner δ-relevante Mengen von der binaren zur kontinuierlichen Domane. Dies fuhrt auf naturliche Art zu einem Raten-Verzerrungs-Kompromiss zwischen der Grose der δ-relevanten Menge (Rate) und der Veranderung der Klassifikation (Verzerrung). Relevanzkarten konnen dann als gierige Approximationen der Raten-Verzerrungs-Funktion interpretiert werden. Diese Funktion auch nur approximativ zu evaluieren ist NP-schwer, daher entwickeln wir eine heuristische Losungsstrategie, die auf einer konvexen Relaxation des kombinatorischen Problems und dem Assumed Density Filtering (ADF) fur tiefe neuronale ReLU-Netze basiert. Dies fuhrt zu unserer eigenen Erklarungsmethode, die wir Rate-Distortion Explanations (RDE) nennen.Um zu zeigen, dass die Naherungen in ADF notwendig sind, geben wir eine vollstandige Charakterisierung von Familien von Wahrscheinlichkeitsverteilungen an, die unter ReLU neuronalen Netzwerkschichten invariant sind. Wir zeigen, dass die einzigen invarianten Familien entweder degeneriert sind oder auf Stichproben hinauslaufen.Anschliesend schlagen wir mehrere Benchmark-Tests und numerische Bewertungsmethoden fur Relevanzkarten vor und diskutieren sie. Wir vergleichen die RDE mit einer reprasentativen Auswahl etablierter Relevanzmethoden und zeigen, dass sie bei einer Vielzahl von Aufgaben besser abschneidet als die Konkurrenz.Schlieslich erortern wir, warum die Kenntnis der wahren Datenverteilung fur jede bestehende Erklarungsmethode entscheidend ist. Wir kritisieren unseren eigenen Ansatz wegen moglicher Artefakte und fuhren eine starkere informationstheoretische Anforderung ein, die auf der bedingten Entropie basiert. Es wird ein neuer Ansatz, die Arthur-Merlin-Regularisierung, zusammen mit einem neuen Framework entwickelt. Das Framework wird dann auf realistische Algorithmen und Datensatze ausgeweitet, und wir diskutieren, unter welchen Annahmen die Garantien noch gelten.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Neural networks. $3 677449
650 4 $a Medical research. $2 bicssc $3 1556686
650 4 $a Game theory. $3 532607
650 4 $a Maps. $3 544078
650 4 $a Feature selection. $3 3560270
650 4 $a Algorithms. $3 536374
650 4 $a Medical imaging. $3 3172799
650 4 $a Computer science. $3 523869
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0574
690 $a 0800
690 $a 0984
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a Technische Universitaet Berlin (Germany). $3 3480527
773 0 $t Dissertations Abstracts International $g 84-05B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29780938 $z click for full text (PQDT)