Knowledge Elicitation in Deep Learning Models

• Main Author: Wagner Luiz da Cunha Ceulin
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: https://hdl.handle.net/10216/152171
• topic: Outras ciências da engenharia e tecnologias; Other engineering and technologies

Though a buzzword in modern problem-solving across various domains, deep learning presents a significant challenge - interpretability. This thesis journeys through a landscape of knowledge elicitation in deep learning models, shedding light on feature visualization, saliency maps, and model distillation techniques. These techniques were applied to two deep learning architectures: convolutional neural networks (CNNs) and a black box package model (Google Vision). Our investigation provided valuable insights into their effectiveness in eliciting and interpreting the encoded knowledge. While they demonstrated potential, limitations were also observed, suggesting room for further development in this field. This work does not just highlight the need for more transparent, more explainable deep learning models, it gives a gentle nudge to developing innovative techniques to extract knowledge. It is all about ensuring responsible deployment and emphasizing the importance of transparency and comprehension in machine learning. In addition to evaluating existing methods, this thesis also explores the potential for combining multiple techniques to enhance the interpretability of deep learning models. A blend of feature visualization, saliency maps, and model distillation techniques was used in a complementary manner to extract and interpret the knowledge from our chosen architectures. Experimental results highlight the utility of this combined approach, revealing a more comprehensive understanding of the models' decision-making processes. Furthermore, we propose a novel framework for systematic knowledge elicitation in deep learning, which cohesively integrates these methods. This framework showcases the value of a holistic approach toward model interpretability rather than relying on a single method. Lastly, we discuss the ethical implications of our work. As deep learning models continue to permeate various sectors, from healthcare to finance, ensuring their decisions are explainable and justified becomes increasingly crucial. Our research underscores this importance, laying the groundwork for creating more transparent, accountable AI systems in the future.