1,720,965 research outputs found
mit einem Fokus auf die Sprachdomäne
No full textLearning machines, like complex Deep Neural Networks (DNNs), are employed in critical infrastructures such as the medical or financial domains. These models affect human lives and therefore humans must be able to inspect them thoroughly. Unfortunately, there is currently a trade-off between the complexity of a neural model and the ability of humans to explain or interpret its decisions or representations. This thesis contributes to efforts that counter the opaqueness of DNNs, in particular in the language domain.
Explainability: I first consider feature attribution methods that attribute the output activation of a DNN to its input features. In a first contribution, two established explainers are combined into a hybrid, coined Pattern-Guided Integrated Gradients (PGIG). PGIG inherits important properties from its parent methods and performs favorably in a large-scale experiment in which the parent methods and numerous prior approaches are included. Problematic edge cases are identified, too. The feature attribution methods considered above produce explanations for the input as well as for the hidden layers of the DNN they aim to explain. However, non-linear mappings as well as dimensionality reductions and expansions make it difficult to inspect such intermediate explanations. In a second contribution, the possibility for humans to make sense of hidden explanations is demonstrated for a prominent class of DNNs, so called Graph Convolutional Networks (GCNs). A new method that traces and visualizes the cross-layer hidden state dynamics in GCNs is proposed and evaluated.
Efficiency: Several neural explainability methods, including the proposed PGIG explainer, are computationally very expensive. This is why, in a third contribution, the possibility to efficiently model the explanations of such expensive explainers is explored. To this end, efficient feature attribution modelling with Empirical Explainers is proposed. Empirical Explainers learn from data to predict the attribution maps of expensive explainers. Experiments show that Empirical Explainers model their expensive counterparts with significant accuracy, at a fraction of the cost. They could thus become a viable efficient alternative in applications that tolerate an approximation error. In a fourth contribution, I demonstrate how feature attribution methods can be exploited to also increase efficiency on the human side. A system is presented that aids the human expert in analysing machine translation engines by exposing systematic differences between human and machine translations.
Interpretability: The feature attribution maps considered above are informative in particular when the input space is fully human-understandable. This generally is not the case for distributed language vector spaces from which models in the language domain consume inputs. In a fifth contribution, I present a novel interpretability method that maps language vectors onto human-understandable concepts. In a sixth contribution, the method is adapted to cross-lingual settings, to make it available for low-resource languages.Verfahren des maschinellen Lernens, insbesondere tiefe neuronale Netze (TNN), werden mittlerweile in vielen kritischen Infrastrukturen eingesetzt, z.B. in der medizinischen Domäne oder dem Finanzsektor. Die Netze, die dort zum Einsatz kommen, haben einen nicht unerheblichen Einfluss auf das Leben von Menschen. Diese müssen daher unbedingt in der Lage sein, die TNNs eingehend zu inspizieren. Eine Inspektion wird erschwert durch die hohe Komplexität von TNNs, da ein Spannungsverhältnis besteht zwischen Komplexität und Erklärbarkeit, bzw. Interpretierbarkeit. Diese Arbeit leistet Beiträge, die der Undurchsichtigkeit solcher komplexer Modelle entgegenwirken, insbesondere in der Sprachdomäne.
Erklärbarkeit: Zunächst berücksichtige ich Methoden, die die Ausgaben eines TNNs der Eingabe zuordnen. Zwei solche Attributionsmethoden werden in einem ersten Beitrag in einem Hybrid kombiniert, Pattern-Guided Integrated Gradients (PGIG). PGIG erbt wichtige theoretische Eigenschaften und in experimentellen Auswertungen übertrifft es die Originale, sowie auch zahlreiche andere Methoden. Jedoch werden auch Grenzfälle identifiziert, in denen der Hybrid unterlegen bleibt.Attributionsmethoden generieren Erklärungen nicht nur für die Ein- und Ausgabeschichten eines TNNs, sondern auch für versteckte Schichten. Diese zu inspizieren ist jedoch sehr herausfordernd wegen nicht-linearen Abbildungen und Dimensionsänderungen. In einem zweiten Beitrag werden solche Inspektionen für sogenannte Graph Convolutional Networks (GCNs) mit einem neuen Ansatz zur Visualisierung versteckter Dynamiken ermöglicht.
Effizienz: Viele Methoden der neuronalen Erklärbarkeit, einschließlich PGIG, sind sehr rechenaufwendig. Aus diesem Grund wird in einem dritten Beitrag die Möglichkeit erörtert, rechenaufwendige Attributionen empirisch zu approximieren. Experimente suggerieren, dass dies mit signifikanter Akkuratheit bei einem Bruchteil des Rechenaufwandes möglich ist. Der Ansatz könnte so die Kosten neuronaler Erklärbarkeit in Applikationen deckeln, die einen Approximationsfehler tolerieren. In einem vierten Beitrag demonstriere ich, wie Methoden der neuronalen Erklärbarkeit genutzt werden können, um auch die Effizienz auf der menschlichen Seite zu steigern, insbesondere bei der Diagnose von Modellen zur maschinellen Übersetzung. Das vorgestellte System nutzt Attributionsmethoden, um systematische Unterschiede zwischen menschlichen und maschinellen Übersetzungen aufzuklären.
Interpretierbarkeit: Attributionsmethoden sind insbesondere dann informativ, wenn die Eingabe interpretierbar ist. Dies ist für die nun üblichen Sprach-Repräsentationen in verteilten Vektorräumen jedoch i.d.R. nicht der Fall. In einem fünften Beitrag präsentiere ich eine Methode zur Interpretierbarkeit, die solche vektoriellen Repräsentationen auf für Menschen verständliche Konzepte abbildet. In einem sechsten Beitrag wird die Methode in eine mehrsprachige Version konvertiert
Going Beyond Counting First Authors in Author Co-citation Analysis
Full text linkThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
Variations on the Author
Full text link“Variations on the Author” discusses two of Eduardo Coutinho’s recent films (Um Dia na Vida, from 2010, and Últimas Conversas, posthumously released in 2015) and their contribution to the general question of documentary authorship. The director’s filmography is characterized by a consistent yet self-effacing form of authorial self-inscription: Coutinho often features as an interviewer that rather than express opinions propels discourses; an interviewer that is good at listening. This mode of self-inscription characterizes him as an author who is not expressive but who is nonetheless markedly present on the screen. In Um Dia na Vida, however, Coutinho is completely absent form the image, while Últimas Conversas, on the contrary, includes a confessional prologue that moves the director from the margins to the center of his films. This article examines the ways in which these works stand out in the filmography of a director who offers new insights into the notion of cinematic authorship
Appropriate Similarity Measures for Author Cocitation Analysis
Full text linkWe provide a number of new insights into the methodological discussion about author cocitation analysis. We first argue that the use of the Pearson correlation for measuring the similarity between authors’ cocitation profiles is not very satisfactory. We then discuss what kind of similarity measures may be used as an alternative to the Pearson correlation. We consider three similarity measures in particular. One is the well-known cosine. The other two similarity measures have not been used before in the bibliometric literature. Finally, we show by means of an example that our findings have a high practical relevance.information science;Pearson correlation;cosine;similarity measure;author cocitation analysis
Dispelling the Myths Behind First-author Citation Counts
Full text linkWe conducted a full-scale evaluative citation analysis study of scholars in the XML research field to explore just how different from each other author rankings resulting from different citation counting methods actually are, and to demonstrate the capability of emerging data and tools on the Web in supporting more realistic citation counting methods. Our results contest some common arguments for the continued
use of first-author citation counts in the evaluation of scholars, such as high correlations between author rankings by first-author citation counts and other citation
counting methods, and high costs of using more realistic citation counting methods that are not well-supported by the ISI databases. It is argued that increasingly available digital full text research papers make it possible for citation analysis studies to go beyond what the ISI databases have directly supported and to employ more
sophisticated methods
koamabayili/VECTRON-author-checklist: VECTRON author checklist
No full textWe have done our best to complete the author checklist relating to the use of animals in the hut study. Note that the objective for the hut study was to evaluate the IRS treatment applications for residual efficacy against Anopheles mosquitoes, including the local An. coluzzii mosquito population. Cows were only used to attract mosquitoes into the huts and no tests were carried out directly on the cows. The author checklist is intended for use with studies where experiments are carried out on animals, which is why we have had such difficulty in completing this for the hut study, as many of the questions do not relate to how the cows were used
Author-wise bibliometric analysis based on entropy.
No full textAuthor-wise bibliometric analysis based on entropy.</p
- …
