[intermediate] AutoML for Generic Computer Vision Tasks
In recent years, deep networks have reached remarkable performance on computer vision tasks, and we have seen a fast increase in their popularity. This success is attributed to decades of research in many aspects of the field, ranging from new architectures to learning strategies. The difficulty of deep network design has prompted a new field, named AutoML. AutoML aims to automate the design of deep learning components by spending machine computing time instead of human research time. AutoML for generic computer vision tasks has long been pursued in machine learning research. This course will discuss the advanced AutoML research works, including the automated search for neural architecture (NAS), normalization-activation operations, dropout patterns, data augmentations, training hyperparameters, loss functions, the whole ML algorithms, etc.
- Introduction to Deep Learning for Computer Vision Tasks
- General Principles of AutoML Algorithms
- Automated Search of Network Architectures, Learning Strategies, etc.
 Elsken, Thomas, Jan Hendrik Metzen, and Frank Hutter. “Neural Architecture Search: A survey.” The Journal of Machine Learning Research 20.1 (2019): 1997-2017.
 Cubuk, Ekin D., et al. “AutoAugment: Learning Augmentation Strategies from Data.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.
 Real, Esteban, et al. “AutoML-Zero: Evolving Machine Learning Algorithms from Scratch.” International Conference on Machine Learning. PMLR, 2020.
 Liu, Hanxiao, et al. “Evolving Normalization-Activation Layers.” arXiv preprint arXiv:2004.02967 (2020).
 Dai, Xiaoliang, et al. “FBNetV3: Joint Architecture-Recipe Search Using Neural Acquisition Function.” arXiv e-prints (2020): arXiv-2006.
 He, Xin, Kaiyong Zhao, and Xiaowen Chu. “AutoML: A Survey of the State-of-the-Art.” Knowledge-Based Systems 212 (2021): 106622.
 Pham, Hieu, and Quoc V. Le. “AutoDropout: Learning Dropout Patterns to Regularize Deep Networks.” arXiv preprint arXiv:2101.01761 1.2 (2021): 3.
 Li, Hao, et al. “AutoLoss-Zero: Searching Loss Functions from Scratch for Generic Tasks.” arXiv preprint arXiv:2103.14026 (2021).
Basic machine learning and computer vision knowledge.
Dr. Jifeng Dai is an Executive Research Director at SenseTime. He received his bachelor’s degree from Tsinghua University in 2009 and his doctoral degree from 2014. He was a visiting scholar at UCLA from 2012 to 2013. His research interests include object detection, segmentation, and deep learning algorithms in computer vision. He has published more than 30 papers in top conferences and journals and has gained more than 14,000 citations. He won the first prize in the COCO object detection challenge for two consecutive years. He serves as an Editorial Board Member of IJCV, Area Chair of CVPR 2021 and ECCV 2020, Senior PC Member of AAAI 2018, and Young Scientist of Zhiyuan Institute of Artificial Intelligence in Beijing.