abstract: One of the measures to determine the intelligence of a system is through Question Answering, as it requires a system to comprehend a question and reason using its knowledge base to accurately answer it. Qualitative word problems are an important subset of such problems, as they require a system to recognize and reason with qualitative knowledge expressed in natural language. Traditional approaches in this domain include multiple modules to parse a given problem and to perform the required reasoning. Recent approaches involve using large pre-trained Language models like the Bidirection Encoder Representations from Transformers for downstream question answering tasks through supervision. These approaches however either suffer from errors between multiple modules, or are not interpretable with respect to the reasoning process employed. The proposed solution in this work aims to overcome these drawbacks through a single end-to-end trainable model that performs both the required parsing and reasoning. The parsing is achieved through an attention mechanism, whereas the reasoning is performed in vector space using soft logic operations. The model also enforces constraints in the form of auxiliary loss terms to increase the interpretability of the underlying reasoning process. The work achieves state of the art accuracy on the QuaRel dataset and matches that of the QuaRTz dataset with additional interpretability. / Dissertation/Thesis / Masters Thesis Computer Science 2020

Thesis advisor: Jan Engelbrecht / Along with the proliferation of digital data collected using sensor technologies and a boost of computing power, Deep Learning (DL) based approaches have drawn enormous attention in the past decade due to their impressive performance in extracting complex relations from raw data and representing valuable information. At the same time, though, rooted in its notorious black-box nature, the appreciation of DL has been highly debated due to the lack of interpretability. On the one hand, DL only utilizes statistical features contained in raw data while ignoring human knowledge of the underlying system, which results in both data inefficiency and trust issues; on the other hand, a trained DL model does not provide to researchers any extra insight about the underlying system beyond its output, which, however, is the essence of most fields of science, e.g. physics and economics. The interpretability issue, in fact, has been naturally addressed in physics research. Conventional physics theories develop models of matter to describe experimentally observed phenomena. Tasks in DL, instead, can be considered as developing models of information to match with collected datasets. Motivated by techniques and perspectives in conventional physics, this thesis addresses the issue of interpretability in general information modeling. This thesis endeavors to address the two drawbacks of DL approaches mentioned above. Firstly, instead of relying on an intuition-driven construction of model structures, a problem-oriented perspective is applied to incorporate knowledge into modeling practice, where interesting mathematical properties emerge naturally which cast constraints on modeling. Secondly, given a trained model, various methods could be applied to extract further insights about the underlying system, which is achieved either based on a simplified function approximation of the complex neural network model, or through analyzing the model itself as an effective representation of the system. These two pathways are termed as guided model design (GuiMoD) and secondary measurements, respectively, which, together, present a comprehensive framework to investigate the general field of interpretability in modern Deep Learning practice. Remarkably, during the study of GuiMoD, a novel scheme emerges for the modeling practice in statistical learning: Algebraic Learning (AgLr). Instead of being restricted to the discussion of any specific model structure or dataset, AgLr starts from idiosyncrasies of a learning task itself and studies the structure of a legitimate model class in general. This novel modeling scheme demonstrates the noteworthy value of abstract algebra for general artificial intelligence, which has been overlooked in recent progress, and could shed further light on interpretable information modeling by offering practical insights from a formal yet useful perspective. / Thesis (PhD) — Boston College, 2021. / Submitted to: Boston College. Graduate School of Arts and Sciences. / Discipline: Physics.

Agile is an idea that has spread far within the corporate world and was originally designed for use in small, single- team project within IT. To this, limited is known about agile in larger settings and the purpose of this thesis is to explore the translation of agile in a large organization. Conducting an ethnographic study at SEB Pension & Försäkring we illustrate that the translation of agile imply adaption. We identify three processes of translation; (1) adaption through unifying the understanding of agile, (2) adaption through testing of agile elements, and (3) adaption through negotiations. The ethnography indicates that translation of agile in large organizations is challenging and individuals struggle to convey the essence of the idea, ending up in discussions through various interventions. Due to the popularity of agile the contributions is of value for organizations that attempts to become agile. The thesis is limited by the restricted time of conducting ethnographic studies. Further research is needed to explore the translation of agile in larger settings and to provide validity for the three processes of translation.

The ubiquitous presence of automated decision-making systems that have a performance comparable to humans brought attention towards the necessity of interpretability for the generated predictions. Whether the goal is predicting the system’s behavior when the input changes, building user trust, or expert assistance in improving the machine learning methods, interpretability is paramount when the problem is not sufficiently validated in real applications, and when unacceptable results lead to significant consequences. While for humans, there are no standard interpretations for the decisions they make, the complexity of the systems with advanced information-processing capacities conceals the detailed explanations for individual predictions, encapsulating them under layers of abstractions and complex mathematical operations. Interpretability for deep learning classifiers becomes, thus, a challenging research topic where the ambiguity of the problem statement allows for multiple exploratory paths. Our work focuses on generating natural language interpretations for individual predictions of deep learning text classifiers. We propose a framework for extracting and identifying the phrases of the training corpus that influence the prediction confidence the most through unsupervised key phrase extraction and neural predictions. We assess the contribution margin that the added justification has when the deep learning model predicts the class probability of a text instance, by introducing and defining a contribution metric that allows one to quantify the fidelity of the explanation to the model. We assess both the performance impact of the proposed approach on the classification task as quantitative analysis and the quality of the generated justifications through extensive qualitative and error analysis. This methodology manages to capture the most influencing phrases of the training corpus as explanations that reveal the linguistic features used for individual test predictions, allowing humans to predict the behavior of the deep learning classifier.

Machine learning models are becoming more and more powerful and accurate, but their good predictions usually come with a high complexity. Depending on the situation, such a lack of interpretability can be an important and blocking issue. This is especially the case when trust is needed on the user side in order to take a decision based on the model prediction. For instance, when an insurance company uses a machine learning algorithm in order to detect fraudsters: the company would trust the model to be based on meaningful variables before actually taking action and investigating on a particular individual. In this thesis, several explanation methods are described and compared on multiple datasets (text data, numerical), on classification and regression problems. / Maskininlärningsmodellerna blir mer och mer kraftfulla och noggranna, men deras goda förutsägelser kommer ofta med en hög komplexitet. Beroende på situationen kan en sådan brist på tolkning vara ett viktigt och blockerande problem. Särskilt är det fallet när man behöver kunna lita på användarsidan för att fatta ett beslut baserat på modellprediktionen. Till exempel, ett försäkringsbolag kan använda en maskininlärningsalgoritm för att upptäcka bedrägerier, men företaget vill vara säkert på att modellen är baserad på meningsfulla variabler innan man faktiskt vidtar åtgärder och undersöker en viss individ. I denna avhandling beskrivs och förklaras flera förklaringsmetoder, på många dataset av typerna textdata och numeriska data, på klassificerings- och regressionsproblem.

As a complicated lethal medical emergency, sepsis is not easy to be diagnosed until it is too late for taking any life saving actions. Early prediction of sepsis in ICUs may reduce inpatient mortality rate. Although deep learning models can make predictions on the outcome of ICU stays with high accuracies, the opacity of such neural networks decreases their reliability. Particularly, in the ICU settings where the time is not on doctors' side and every single mistake increase the chances of patient's mortality. Therefore, it is crucial for the predictive model to provide some sort of reasoning in addition to the prediction it provides, so that the medical staff could avoid actions based on false alarms. To address this problem, we propose to add an attention layer to a deep recurrent neural network that can learn the relative importance of each of the parameters of the multivariate data of the ICU stay. Our approach sheds light on providing explainability through attention mechanism. We compare our method with some of the state-of-the-art methods and show the superiority of our approach in terms of providing explanations.

Importance prioritised coding is a principle aimed at improving the interpretability (or image content recognition) versus bit-rate performance of image coding systems. This can be achieved by (1) detecting and tracking image content or regions of interest (ROI) that are crucial to the interpretation of an image, and (2)compressing them in such a manner that enables ROIs to be encoded with higher fidelity and prioritised for dissemination or transmission. Traditional image coding systems prioritise image data according to an objective measure of distortion and this measure does not correlate well with image quality or interpretability. Importance prioritised coding, on the other hand, aims to prioritise image contents according to an 'importance map', which provides a means for modelling and quantifying the relative importance of parts of an image. In such a coding scheme the importance in parts of an image containing ROIs would be higher than other parts of the image. The encoding and prioritisation of ROIs means that the interpretability in these regions would be improved at low bit-rates. An importance prioritised image coder incorporated within the JPEG 2000 international standard for image coding, called IMP-J2K, is proposed to encode and prioritise ROIs according to an 'importance map'. The map can be automatically generated using image processing algorithms that result in a limited number of ROIs, or manually constructed by hand-marking OIs using a priori knowledge. The proposed importance prioritised coder coder provides a user of the encoder with great flexibility in defining single or multiple ROIs with arbitrary degrees of importance and prioritising them using IMP-J2K. Furthermore, IMP-J2K codestreams can be reconstructed by generic JPEG 2000 decoders, which is important for interoperability between imaging systems and processes. The interpretability performance of IMP-J2K was quantitatively assessed using the subjective National Imagery Interpretability Rating Scale (NIIRS). The effect of importance prioritisation on image interpretability was investigated, and a methodology to relate the NIIRS ratings, ROI importance scores and bit-rates was proposed to facilitate NIIRS specifications for importance prioritised coding. In addition, a technique is proposed to construct an importance map by allowing a user of the encoder to use gaze patterns to automatically determine and assign importance to fixated regions (or ROIs) in an image. The importance map can be used by IMP-J2K to bias the encoding of the image to these ROIs, and subsequently to allow a user at the receiver to reconstruct the image as desired by the user of the encoder. Ultimately, with the advancement of automated importance mapping techniques that can reliably predict regions of visual attention, IMP-J2K may play a significant role in matching an image coding scheme to the human visual system.

The purpose of this thesis is to develop an interpretable model that gives predictions for what factors impacted a shooter’s results. Experiment is our chosen research method. Our three independent variables are weapon movement, trigger pull force and heart rate. Our dependent variable is shooting accuracy. A random forest regression model is trained with the experiment data to produce predictions of shooting accuracy and to show correlation between independent and dependent variables. Our method shows that an increase in weapon movement, trigger pull force and heart rate decrease the predicted accuracy score. Weapon movement impacted shooting results the most with 53.61%, while trigger pull force and heart rateimpacted shooting results 22.20% and 24.18% respectively. We have also shown that LIME can be a viable method to give explanations on how the measured factors impacted shooting results. The results from this thesis lay the groundwork for better training tools for target shooting using explainable prediction models with sensors.

The thesis deals with definability of certain Zariski geometries, introduced by Zilber, in the theory of algebraically closed fields. I axiomatise a class of structures, called 'abstract linear spaces', which are a common reduct of these Zariski geometries. I then describe what an interpretation of an abstract linear space in an algebraically closed field looks like. I give a new proof that the structure "quantum harmonic oscillator", introduced by Zilber and Solanki, is not interpretable in an algebraically closed field. I prove that a similar structure from an unpublished note of Solanki is not definable in an algebraically closed field and explain the non-definability of both structures in terms of geometric interpretation of the group law on a Galois cohomology group H¹(k(x), μ_n). I further consider quantum Zariski geometries introduced by Zilber and give necessary and sufficient conditions that a quantum Zariski geometry be definable in an algebraically closed field. Finally, I take an attempt at extending the results described above to complex-analytic setting. I define what it means for quantum Zariski geometry to have a complex analytic model, an give a necessary and sufficient conditions for a smooth quantum Zariski geometry to have one. I then prove a theorem giving a partial description of an interpretation of an abstract linear space in the structure of compact complex spaces and discuss the difficulties that present themselves when one tries to understand interpretations of abstract linear spaces and quantum Zariski geometries in the compact complex spaces structure.

Context. Software testing is the process, where a particular software product, or a system is executed, in order to find out the bugs, or issues which may otherwise degrade its performance. Software testing is usually done based on pre-defined test cases. A test case can be defined as a set of terms, or conditions that are used by the software testers to determine, if a particular system that is under test operates as it is supposed to or not. However, in numerous situations, test cases can be so many that executing each and every test case is practically impossible, as there may be many constraints. This causes the testers to prioritize the functions that are to be tested. This is where the ability of topic models can be exploited. Topic models are unsupervised machine learning algorithms that can explore large corpora of data, and classify them by identifying the hidden thematic structure in those corpora. Using topic models for test case prioritization can save a lot of time and resources. Objectives. In our study, we provide an overview of the amount of research that has been done in relation to topic models. We want to uncover various quality criteria, evaluation methods, and metrics that can be used to evaluate the topic models. Furthermore, we would also like to compare the performance of two topic models that are optimized for different quality criteria, on a particular interpretability task, and thereby determine the topic model that produces the best results for that task. Methods. A systematic mapping study was performed to gain an overview of the previous research that has been done on the evaluation of topic models. The mapping study focused on identifying quality criteria, evaluation methods, and metrics that have been used to evaluate topic models. The results of mapping study were then used to identify the most used quality criteria. The evaluation methods related to those criteria were then used to generate two optimized topic models. An experiment was conducted, where the topics generated from those two topic models were provided to a group of 20 subjects. The task was designed, so as to evaluate the interpretability of the generated topics. The performance of the two topic models was then compared by using the Precision, Recall, and F-measure. Results. Based on the results obtained from the mapping study, Latent Dirichlet Allocation (LDA) was found to be the most widely used topic model. Two LDA topic models were created, optimizing one for the quality criterion Generalizability (TG), and one for Interpretability (TI); using the Perplexity, and Point-wise Mutual Information (PMI) measures respectively. For the selected metrics, TI showed better performance, in Precision and F-measure, than TG. However, the performance of both TI and TG was comparable in case of Recall. The total run time of TI was also found to be significantly high than TG. The run time of TI was 46 hours, and 35 minutes, whereas for TG it was 3 hours, and 30 minutes.Conclusions. Looking at the F-measure, it can be concluded that the interpretability topic model (TI) performs better than the generalizability topic model (TG). However, while TI performed better in precision, Conclusions. Looking at the F-measure, it can be concluded that the interpretability topic model (TI) performs better than the generalizability topic model (TG). However, while TI performed better in precision, recall was comparable. Furthermore, the computational cost to create TI is significantly higher than for TG. Hence, we conclude that, the selection of the topic model optimization should be based on the aim of the task the model is used for. If the task requires high interpretability of the model, and precision is important, such as for the prioritization of test cases based on content, then TI would be the right choice, provided time is not a limiting factor. However, if the task aims at generating topics that provide a basic understanding of the concepts (i.e., interpretability is not a high priority), then TG is the most suitable choice; thus making it more suitable for time critical tasks.