Generative Adversarial Networks (GAN) learn to synthesize novel samples for a given data distribution. While GANs can train on diverse data of various modalities, the most successful use cases to date apply GANs to computer vision tasks. Despite significant advances in training algorithms and network architectures, GANs still struggle to consistently generate high-quality outputs after training. We present a series of papers that improve GAN output inference qualitatively and quantitatively. The first chapter, Alpha Model Domination, addresses a related subfield of Multiple Choice Learning, which -- like GANs -- aims to generate diverse sets of outputs. The next chapter, CoachGAN, introduces a real-time refinement method for the latent input space that improves inference quality for pretrained GANs. The following two chapters introduce finetuning methods for arbitrary, end-to-end differentiable GANs. The first, PuzzleGAN, proposes a self-supervised puzzle-solving task to improve global coherence in generated images. The latter, Trained Truncation Trick, improves upon a common inference heuristic by better maintaining output diversity while increasing image realism. Our final work, Two Second StyleGAN Projection, reduces the time for high-quality, image-to-latent GAN projections by two orders of magnitude. We present a wide array of results and applications of our method. We conclude with implications and directions for future work.

Generative Adversarial Networks (GAN) learn to synthesize novel samples for a given data distribution. While GANs can train on diverse data of various modalities, the most successful use cases to date apply GANs to computer vision tasks. Despite significant advances in training algorithms and network architectures, GANs still struggle to consistently generate high-quality outputs after training. We present a series of papers that improve GAN output inference qualitatively and quantitatively. The first chapter, Alpha Model Domination, addresses a related subfield of Multiple Choice Learning, which -- like GANs -- aims to generate diverse sets of outputs. The next chapter, CoachGAN, introduces a real-time refinement method for the latent input space that improves inference quality for pretrained GANs. The following two chapters introduce finetuning methods for arbitrary, end-to-end differentiable GANs. The first, PuzzleGAN, proposes a self-supervised puzzle-solving task to improve global coherence in generated images. The latter, Trained Truncation Trick, improves upon a common inference heuristic by better maintaining output diversity while increasing image realism. Our final work, Two Second StyleGAN Projection, reduces the time for high-quality, image-to-latent GAN projections by two orders of magnitude. We present a wide array of results and applications of our method. We conclude with implications and directions for future work.

In einer Zeit von steigenden Studierendenzahlen und zunehmender Nachfrage nach individuellen Übungsmöglichkeiten, sehen sich Lehrende mit der Herausforderung konfrontiert, mit wenig Ressourcen möglichst viele Lernmöglichkeiten bereitzustellen. Computergestützte Übungen können hierfür eine ressourcenschonende Möglichkeit darstellen, und die direkte Interaktion zwischen Lehrenden und Lernenden um eine flexible und digitale Komponente ergänzen. Für eine Variation in den Single-Choice-Aufgaben kann ein Generator verwendet werden. Diese Arbeit untersucht den aktuellen Stand der Generierung von Multiple -Choice-Aufgaben und stellt einige Softwarelösungen vor. Im Anschluss wird die Implementierung eines angepassten Generators exemplarisch erläutert, anhand eines Prototyps in der Programmiersprache Python. Der Prototyp übernimmt eine Aufgabenschablone, die mehrere Antworten und Parameter enthalten kann. Daraus können viele unterschiedliche Aufgaben generiert werden.

The multiple-choice test question is a popular item format used for tests ranging from classroom assessments to professional licensure exams. The popularity of this format stems from its administration and scoring efficiencies. The most common multiple-choice format consists of a stem that presents a problem to be solved accompanied by a single correct answer and two, three, or four incorrect answers. A well-constructed item using this format can result in a high quality assessment of an examinee's knowledge, skills and abilities. However, for some complex, higher-order knowledge, skills and abilities, a single correct answer is often insufficient. Test developers tend to avoid using multiple correct answers out of a concern about the increased difficulty and lower discrimination of such items. However, by avoiding the use of multiple correct answers, test constructors may inadvertently create validity concerns resulting from incomplete content coverage and construct irrelevant variance. This study explored an alternative way of implementing multiple-choice questions with two or more correct answers by specifying in each question the number of answers examinees should select instead of using the traditional guideline to select all that apply. This study investigated the performance of three operational exams that use a standard multiple-choice format where the examinees are told how many answers they are to select. The collective statistical performance of multiple-choice items that included more than one answer that is keyed as correct was compared with the performance of traditional single-answer, multiple-choice (SA) items within each exam. The results indicate that the multiple-answer, multiple-choice (MA) items evaluated from these three exams performed at least as well as to the single-answer questions within the same exams.

While large language models like GPT-3 have achieved impressive results in the zero-, one-, and few-shot settings, they still significantly underperform on some tasks relative to the state of the art (SOTA). For many tasks it would be useful to have answer options explicitly listed out in a multiple choice format, decreasing computational cost and allowing the model to reason about the relative merits of possible answers. We argue that the reason this hasn't helped models like GPT-3 close the gap with the SOTA is that these models struggle with symbol binding - associating each answer option with a symbol that represents it. To ameliorate this situation we introduce index prompting, a way of leveraging language models trained on code to successfully answer multiple choice formatted questions. When used with the OpenAI Codex model, our method improves accuracy by about 18% on average in the few-shot setting relative to GPT-3 across 8 datasets representing 4 common NLP tasks. It also achieves a new single-model state of the art on ANLI R3, ARC (Easy), and StoryCloze, suggesting that GPT-3's latent "understanding" has been previously underestimated.

No description available.

No description available.

This thesis investigated whether out­come after in-hospital cardiac arrest patients could be improved by a cardiopulmonary resuscitation (CPR) educational intervention focusing on all hospital healthcare professionals. Annually in Sweden, approximately 3000 in-hospital patients suffer a cardiac arrest in which CPR is attempted, and which 900 will survive. The thesis is based on five papers: Paper I was a methodological study concluding in a reliable multiple choice questionnaire (MCQ) aimed at measuring CPR knowledge. Paper II was an intervention study. The intervention consisted of educating 3144 healthcare professionals in CPR. The MCQ from Paper I was answered by the healthcare professionals both before (82% response rate) and after (98% response rate) education. Theoretical knowledge improved in all the different groups of healthcare professionals after the intervention. Paper III was an observational laboratory study investigating the practical CPR skills of 74 healthcare professionals’. Willingness to use an automated external defibrillator (AED) improved generally after educa­tion, and there were no major differences in CPR skills between the different healthcare professions. Paper IV investigated, by use of a questionnaire, the attitudes to CPR of 2152 healthcare professionals (82% response rate). A majority of healthcare professionals reported a positive attitude to resuscitation. Paper V was a register study of patients suffering from cardiac arrest. The intervention tended not to reduce the delay to start of treatment or to increase overall survival. However, our results suggested indirect signs of an improved cerebral function among survivors. In conclusion, CPR education and the introduction of AEDs in-hospital – improved healthcare professionals knowledge, skills, and attitudes – did not improve patients’ survival to hospital discharge, but the functional status among survivors improved.

D.Litt.et Phil. (African Languages) / 1. The aim of the study The primary aim of this study was to ascertain whether or not a multiple-choice test can effectively assess English mother-tongue competence. Because the testing of language is at issue, the study was approached from an applied linguistic perspective. 2. The method of investigation 2.1. Uterature study. A review of the literature relating to the following topics was performed: (i) mother-tongue competence; (ii) language skills and in particular, the skills inherent in reading; (iii) the communicative approach to language teaching and testing; (iv) multiple-choice testing; and (v) test validity. 2.2. Empirical research work. Each of the four South African provincial education departments within the Department of Education and Culture: House of Assembly was called upon to compile three traditional English First Language papers for the trial matriculation examinations. A single item bank test of one hundred and fifty pretested multiple-choice questions was compiled for the same examination period. Pupils from the four provinces wrote the traditional papers which were set for their provinces. In other words, the pupils from each province wrote a separate set of traditional papers. In contrast, the same item bank test was written by all 9456 matriculants involved in the project. . 3. Findings The study revealed that the skills inherent in the four language modes of reading, writing, speaking and listening overlap each other to such an extent that it is virtually impossible to separate them for testing purposes. The validity coefficients of the combined scores of the three traditional papers and the total scores of the item bank test were consistently satisfactory for all four education

This study investigated the difficulty order of 38 grammar structures obtained from an analysis of multiple-choice items using a Rasch analysis. The order was compared with the order predicted by processability theory and the order in which the structures appear in junior and senior high school textbooks in Japan. Because processability theory is based on natural speech data, a sentence repetition test was also conducted in order to compare the result with the order obtained from the multiple-choice tests and the order predicted by processability theory. The participants were 872 Japanese university students, whose TOEIC scores ranged from 200 to 875. The difficulty order of the 38 structures was displayed according to their Rasch difficulty estimates: The most difficult structure was subjunctive and the easiest one was present perfect with since in the sentence. The order was not in accord with the order predicted by processability theory, and the difficulty order derived from the sentence repetition test was not accounted for by processability theory either. In other words, the results suggest that processability theory only accounts for natural speech data, and not elicited data. Although the order derived from the repetition test differed from the order derived from the written tests, they correlated strongly when the repetition test used ungrammatical sentences. This study tentatively concluded that the students could have used their implicit knowledge when answering the written tests, but it is also possible that students used their explicit knowledge when correcting ungrammatical sentences in the repetition test. The difficulty order of grammatical structures derived from this study was not in accord with the order in which the structures appear in junior and senior high school textbooks in Japan. Their correlation was extremely low, which suggests that there is no empirical basis for textbook makers'/writers' policy regarding the ordering of grammar items. This study also demonstrated the difficulty of writing items testing the knowledge of the same grammar point that show similar Rasch difficulty estimates. Even though the vocabulary and the sentence positions were carefully controlled and the two items looked parallel to teachers, they often displayed very different difficulty estimates. A questionnaire was administered concerning such items, and the students' responses suggested that they seemed to look at the items differently than teachers and what they notice and how they interpret what they notice strongly influences item difficulty. Teachers or test-writers should be aware that it is difficult to write items that produce similar difficulty estimates and their own intuition or experience might not be the best guide for writing effective grammar test items. It is recommended to pilot test items to get statistical information about item functioning and qualitative data from students using a think-aloud protocol, interviews, or a questionnaire. / CITE/Language Arts