Continual Learning and Catastrophic Forgetting
HTML-код
- Опубликовано: 7 окт 2024
- A lecture that discusses continual learning and catastrophic forgetting in deep neural networks. We discuss the context, methods for evaluating algorithms, and algorithms based on regularization, dynamic architectures, and Complementary Learning Systems. Specifically, we discuss data permutation tasks, incremental task learning, multimodal learning, the Learning without Forgetting algorithm, Elastic Weight Consolidation, Progressive Neural Networks, and Generative replay.
This lecture is from Northeastern University's CS 7150 Summer 2020 class on Deep Learning, taught by Paul Hand.
The notes are available at: khoury.northeas...
References:
Parisi et al. 2019:
Parisi, German I., Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. "Continual lifelong learning with neural networks: A review." Neural Networks (2019).
Chen and Liu 2018:
Chen, Zhiyuan, and Bing Liu. "Lifelong machine learning." Synthesis Lectures on Artificial Intelligence and Machine Learning 12, no. 3 (2018): 1-207. Chapter 4.
Kemker et al. 2017:
Kemker, Ronald, Marc McClure, Angelina Abitino, Tyler L. Hayes, and Christopher Kanan. "Measuring catastrophic forgetting in neural networks." In Thirty-second AAAI conference on artificial intelligence. 2018.
van de Ven and Tolias 2019:
van de Ven, Gido M., and Andreas S. Tolias. "Three scenarios for continual learning." arXiv preprint arXiv:1904.07734 (2019).
Li and Hoiem 2018:
Li, Zhizhong, and Derek Hoiem. "Learning without forgetting." IEEE transactions on pattern analysis and machine intelligence 40, no. 12 (2017): 2935-2947.
Kirkpatrick et al. 2017:
Kirkpatrick, James, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan et al. "Overcoming catastrophic forgetting in neural networks." Proceedings of the national academy of sciences 114, no. 13 (2017): 3521-3526.
Rusu et al. 2016:
Rusu, Andrei A., Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. "Progressive neural networks." arXiv preprint arXiv:1606.04671 (2016).
Shin et al. 2017:
Shin, Hanul, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. "Continual learning with deep generative replay." In Advances in Neural Information Processing Systems, pp. 2990-2999. 2017.
this is a gem on the internet...more people should know about these videos!
Agree !
Very nice lecture giving a good overview. Thanks a lot.
One remark: There are actually regions in the brain that continuously generate new neurons, in particular the detate gyrus in the hippocampus, and we have actually built a model for avoiding catastrophic interference on that, see Wiskott, L., Rasch, M. & Kempermann, G. A functional hypothesis for adult hippocampal neurogenesis: avoidance of catastrophic interference in the dentate gyrus, Hippocampus, 2006, 16, 329-343.
this video is genius why isn't it famous ???
Man , here in 2024 and I just discovered this concepts an vid . Amazing !
Great aproach to the problem, best explanation i´ve found.
Thank you Paul Very Much for this brilliant summary of the "Continual Learning" topic , you saved my day!
indeed very good video. well and easy explained.
Excellent content to get a quick overview.
this video was really informative and well described.
Brilliant lecture!
Paul, Do you have a patreon? Your videos are awesome.
god-tier video
Amazing video!
Dont know much about this subject. What if you would not need old training data but generate it with the current neural network and the knowledge which objects or things where related. You could dream up the old skill and include the new situation. Nvm.. this has been covered in the video..
Hi dear, nice explanation.. do you have python code for this task?
Paul, although a very good explanation but dude c'mon why so low volume!
Great talk! Thank you for the video.
Just two comments regarding typing on ruclips.net/video/vjaq03IYgSk/видео.html:
1. During initialization, wouldn't Y_o be Y^hat_o? Because that is the output of the network
2. In the argmin formula isn't Y_o the same as Y_n?