Subject

AI explainability and interpretability has recently become an important research topic due to new regulations and expectations regarding AI trustworthiness, transparency, fairness etc. In the computer vision domain, most explanations consist of assigning relevance values to pixels or regions in the image, to highlight which part of the image is used or most important for a classification / detection / other decision of the model. This is achieved by propagating gradients or intermediate values in the network back to the input image, resulting in a relevance heatmap.
A recent state-of-the-art architecture in computer vision is the Transformer network, which originated from the natural language processing field, that improved the state-of-the-art in a plethora of vision tasks. Although there are a few early works on Transformer explainability, these are limited to basic model architectures. We would like to extend these methods to newer architectures and possibly improve their effectiveness.

Kind of work

The goal of the thesis is to extend existing explanation methods to new visual Transformer architectures. More specifically, the work will (i) evaluate existing explanation methods for Transformers, (ii) extend them to other architectures such as the Swin Transformer or MViT and possibly (iii) suggest novel methods or architectures to improve explainability.

Framework of the Thesis

H. Chefer, S. Gur, L. Wolf, “Generic Attention-model Explainability for Interpreting Bi-Modal and Encoder-Decoder Transformers”, ICCV, 2021.
H. Fan et al., “Multiscale Vision Transformers”, ICCV, 2021.
Z. Liu et al., “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows”, ICCV, 2021.

Number of Students

1-2

Expected Student Profile

Programming background (e.g., Python)
Background in machine learning
Experience with machine learning frameworks (e.g., any of TensorFlow, PyTorch, Flax)