DAUIN - GR-09 - GRAphics and INtelligent Systems - GRAINS
Contrastive and Prototypical Explanations for Computer Vision
Thesis in external company
Reference persons LIA MORRA
External reference persons ALAN PEROTTI
Research Groups DAUIN - GR-09 - GRAphics and INtelligent Systems - GRAINS
Description Deep convolutional neural network are the de-facto state of the art approach for computer vision tasks. However, a major drawback of these models is their black-box nature - that is, the inability to provide human-understandable information about their inner decision process.
The goal of the thesis is to develop a pipeline for explaining black-box classifiers (e.g., CNNs) for computer vision tasks.
Given an image and a trained neural network, the envisioned system will leverage contrastive learning and prototypical networks in order to provide case-based explanation for the model-predicted image label. Furthermore, the system will highlight the highly similar areas of the retrieved images, as well as contrastive explanation highlighting the differences. Prototypical networks are an emerging framework that aims at enhancing the interpretability of deep neural networks by classifying images based on their visual similarity from class prototypes (“this looks like that”). Our goal, however, is not to modify the architecture or the training process, but rather to leverage these techniques to provide post-hoc explanations for a pre-trained model.
Strong programming (Python) and analytical skills are required. Keras/Tensorflow or Pytorch expertise is preferred and will be acquired. The thesis will be co-supervised by the ISI Foundation.
This Looks Like That - https://arxiv.org/abs/1806.10574
SimCLR - https://arxiv.org/abs/2002.05709
Deadline 09/07/2022 PROPONI LA TUA CANDIDATURA