Subject

The field of deep learning has seen remarkable progress in recent years, enabling state-of-the-art performance in many vision tasks such as object recognition, segmentation, and detection. However, the high complexity of these models often makes them opaque to humans, limiting our ability to interpret their decisions and understand their inner workings. Visual explanations have emerged as a promising approach to address this issue, providing intuitive and interpretable insights into the model's decision-making process.

Kind of work

Despite visual explanations‘ benefits, the quality and reliability of these visual explanations remain a crucial concern. Many studies have shown that these explanations can be highly sensitive to the choice of input image, the model architecture, and the algorithm used to generate them. Moreover, visual explanations are often evaluated using sensitive metrics, making it difficult to compare different methods or draw objective conclusions.
In this thesis proposal, we aim to address these challenges and evaluate the quality of visual explanations for deep learning models in vision tasks. We plan to investigate the following research questions:
1. What are the current state-of-the-art techniques for generating visual explanations, and how do they compare in terms of accuracy, interpretability, and robustness?
2. How can we develop objective metrics to evaluate the quality and reliability of visual explanations, and how do they correlate with human judgments?
3. How do different factors, such as the choice of input image, the model architecture, and the complexity of the task, affect the performance of visual explanation methods?

Framework of the Thesis

To answer these questions, we will conduct a comprehensive literature review, develop novel evaluation metrics and benchmarks, and perform extensive experiments on several deep learning models and vision tasks. We believe that this research will contribute to a better understanding of the strengths and limitations of visual explanations for deep learning models and help researchers and practitioners to develop more reliable and interpretable AI systems.

Number of Students

1

Expected Student Profile

Good knowledge of Machine Learning, AI and data processing. Good programming skills in Python (particularly PyTorch)