Browsing by Subject "HPC"

Understanding Machine Learning models’ behaviour is becoming increasingly important as models are growing in complexity. This thesis proposes a framework for validating machine learned signals using performance metrics and model explainability tools, applied to the context of Digital Humanities and Social Sciences. The framework allows for investigation whether the real-world problem that the model tries to represent is well-defined and whether the model accurately captures the phenomena at hand. Explainability techniques such as SHAP, LIME and Gradient-based methods have been used. These produce feature importance scores that the model bases its decisions on. The cases presented in this thesis are related to the research in Computational History and Historical Discourse Analysis with High Performance Computing. The subject of analysis is the large language model BERT fine-tuned on Eighteenth Century Collections Online (ECCO) documents that classifies books into genres. Investigating the performance of the classifier with precision-recall curves suggests that the class signals might be overlapping and not clearly delineated. Further results do not suggest that the noise elements present in the data caused by the OCR digitising process have significant importance for the decision making of the model. The explainability techniques helped uncover the model’s inner workings by showing that the model gets its signal mostly from the beginnings of samples. In a proxy task, a simpler linear model was trained to perform a projection from keywords to genres and showed inconsistency in the explainability method. Different subsets of data have been investigated as given by cells of a confusion matrix, the confidence in prediction probability or additional metadata. Investigating individual samples allows for qualitative analysis as well as more detailed signal understanding.