With different initialization methods [1] or pruning
strategies [2], we can derive various neural networks (NNs) of the same
architecture but different weights. Previous work generally focuses on
exploring the performance and structure difference of NNs [3]. However, this is
insufficient to explain behaviors of different models. A few works combine XAI methods into the
comparison pipeline [4][5]. They generally compare the feature-based explanations
of different models derived from the same instances. In this project, we aim to
step further to compare the inner working of different NNs. How do learned
features evolve in different NNs? Do different NNs learn the same thing in the
same layer?
Requirements:
Please contact Linhao Meng (l.meng1@tue.nl) if you are interested in this project.
References
[1] Narkhede, M.V., Bartakke, P.P. & Sutaone, M.S. A review on weight initialization strategies for neural networks. Artif Intell Rev 55, 291–322 (2022). https://doi.org/10.1007/s10462-021-10033-z
[2] Blalock, Davis, et al. "What is the state of neural network pruning?." Proceedings of machine learning and systems 2 (2020): 129-146. https://doi.org/10.48550/arXiv.2003.03033
[3] Zeng, H., Haleem, H., Plantaz, X., Cao, N. and Qu, H., 2017. Cnncomparator: Comparative analytics of convolutional neural networks. arXiv preprint arXiv:1710.05285.
[4] X. Xuan, X. Zhang, O. Kwon and K. Ma, "VAC-CNN: A Visual Analytics System for Comparative Studies of Deep Convolutional Neural Networks" in IEEE Transactions on Visualization & Computer Graphics, vol. 28, no. 06, pp. 2326-2337, 2022. doi: 10.1109/TVCG.2022.3165347
[5] Meng, L., van den Elzen, S. and Vilanova, A. (2022), ModelWise:
Interactive Model Comparison for Model Diagnosis, Improvement and Selection.
Computer Graphics Forum, 41: 97-108. https://doi.org/10.1111/cgf.14525