Supervisors: JAN LUNGREN (midsweden university), ULF JENNEHAG (RISE) This research investigates methods for improving the trustworthiness, robustness, and interpretability of machine learning systems, with particular emphasis on understanding how training data and learned representations influence model behaviour. The work combines data-centric explainable AI, representation learning, counterfactual analysis, robustness evaluation, and data valuation. Current research directions include controlled generation of synthetic dataset biases for systematic auditing, factor-isolated counterfactual methods for identifying the influence of specific latent properties, Shapley-based analysis of the contribution of training samples or data subgroups, and task-aware representation compression for studying the relationship between retained information and predictive performance. The research also includes self-supervised and contrastive representation learning for industrial and acoustic monitoring applications based on large quantities of weakly labelled or unlabelled sensor data. These studies require repeated training and evaluation of families of deep neural networks across systematically varied datasets, random initialisations, perturbations, representation spaces, and model configurations. Several experiments additionally involve large-scale generation and processing of synthetic image and audio datasets, computation of high-dimensional embeddings, and repeated model retraining for robustness and data-valuation analyses. The requested GPU-accelerated high-performance computing resources are necessary to make these experiments computationally feasible at sufficient scale and to support statistically reliable and reproducible evaluations. The overall objective is to develop general methods for auditing and explaining machine-learning systems by tracing model behaviour back to properties of their training data and learned representations.