|
Subject
Transformer architecture has achieved state-of-the-art results in many NLP(Natural Language Processing) tasks and significantly improved the performance of the language processing models. The effect is comparable to what had been done to image understanding since 2012 by convolution neural networks(CNNs). So there is an increasing convergence of computer vision and NLP with much more efficient class of architectures. At the end of 2021 we have transformers entering the top of well-known computer vision benchmarks, such as image classification on ImageNet and object detection on COCO. Explainable Artificial Intelligence(XAI) is also a very popular research area at the moment. XAIespecially explainable machine learningwill be essential if future warfighters are to understand, appropriately trust, and effectively manage an emerging generation of artificially intelligent machine partners. The purpose of the XAI methods is to make the machine learning model no longer a black box, so that people can understand how AI is taking its decision and on what part of the input data it bases its decisions upon. The strengths and weaknesses of the model should also be captured by XAI methods so that experts can work on improving the model. However, most of the currently proposed XAI algorithms are discussed on CNNs. So, we will explore whether different XAI methods are practical for different datasets on the Transformer models.
Kind of work
The students need to learn state-of-the-art Transformer models in computer vision and several XAI methods. Then they should visualize the heatmaps of these XAI methods on the Transformer models and the CNNs, comparing the results on two architectures with some error metrics for classification and localization.
Framework of the Thesis
[1] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30. [2] Yang F, Du M, Hu X. Evaluating explanation without ground truth in interpretable machine learning[J]. arXiv preprint arXiv:1907.06831, 2019. [3] Zhou B, Khosla A, Lapedriza A, et al. Learning deep features for discriminative localization[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 2921-2929.
Number of Students
1-2
Expected Student Profile
The student should have a background in machine learning, and python programming knowledge of deep learning (e.g., PyTorch)
|
|