Evaluation of Explainability Methods and Robustness in Image Classification
Raatikainen, Lassi (2022)
Raatikainen, Lassi
2022
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2022-10-06
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202209277303
https://urn.fi/URN:NBN:fi:tuni-202209277303
Tiivistelmä
Modern deep learning architectures have gotten proficient in the task of classifying an image to a correct class. However, the complexity of image classification architectures has made the decision-making process of such models obscure. This is why explainability methods have been introduced to visualize the most relevant parts of an image to a given classification. This thesis studied the evaluation of such explainability methods alongside the robustness of image classification models.
This was investigated through two research questions: "How is the accuracy and stability of explainability methods evaluated?" and "How do image classification models react to different distortions?". For both questions, a literature review was conducted to understand the current knowledge on explainability methods and robustness of image classification models. Consequently, convolutional and vision transformer models were identified as the key architectures within current image classification research. After this, new methods were presented to further develop the understanding of both the evaluation of explainability methods and robustness.
The accuracy of explainability methods was expanded on through a new metric, the Weighting Game, which measures how class-guided explanations are concentrated on top of the correct classes within images. For stability, two new methods were introduced and implemented, which measure how consistent explanations are under transformations or small movement. The experiments revealed that model architecture is important to consider when selecting an explainability method, and that newer explainability methods often do not outperform older ones.
Robustness of image classification models was further studied through the lens of shape distortion robustness. Local thin plate splines were introduced to create distortions to shapes of objects, via bending the silhouette of objects. Results of robustness experiments revealed that the vision transformer architecture was least affected by shape distortions, which aligns with most current research on robustness of image classification models.
As a conclusion, it is argued that current methods to validate explainability methods and robustness of models still require future work, as there are no definitive results or rankings for either category.
This was investigated through two research questions: "How is the accuracy and stability of explainability methods evaluated?" and "How do image classification models react to different distortions?". For both questions, a literature review was conducted to understand the current knowledge on explainability methods and robustness of image classification models. Consequently, convolutional and vision transformer models were identified as the key architectures within current image classification research. After this, new methods were presented to further develop the understanding of both the evaluation of explainability methods and robustness.
The accuracy of explainability methods was expanded on through a new metric, the Weighting Game, which measures how class-guided explanations are concentrated on top of the correct classes within images. For stability, two new methods were introduced and implemented, which measure how consistent explanations are under transformations or small movement. The experiments revealed that model architecture is important to consider when selecting an explainability method, and that newer explainability methods often do not outperform older ones.
Robustness of image classification models was further studied through the lens of shape distortion robustness. Local thin plate splines were introduced to create distortions to shapes of objects, via bending the silhouette of objects. Results of robustness experiments revealed that the vision transformer architecture was least affected by shape distortions, which aligns with most current research on robustness of image classification models.
As a conclusion, it is argued that current methods to validate explainability methods and robustness of models still require future work, as there are no definitive results or rankings for either category.