数据闭环的另外一個方面就是主動化的模子練習。主動化模子練習有两個关頭技能,第一是主動化模子搜刮(AutoML),第二是延续進修(Continual Learning)。主動化模子搜刮指讓練習體系主動举行模子调優,AutoML在曩昔几年是比力红火的鑽研標的目的,也有很多论文和實践的摸索。baidu利用的是一種基于進化算法改良的方案,重要搜刮模子的超参数,如使命的权重、optimizer的参数等。感樂趣的同窗可以参考论文Population-based training [6]。
而這里提到的延续進修是近来AI鑽研者們更加存眷的课题,深度進修在新的数据延续注入模子練習的進程中,會表现出两個缺點:(1)劫難性遗忘(catastropic forgetting),即學了新的数据今後在旧的数据上輕易產生遗忘;(2)可塑性丧失(loss of plasticity),即模子在屡次練習今後,在新数据上的進修能力變差/慢。
[1] Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D. and Han, S., 2022. BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird’s-Eye View Representation. arXiv preprint arXiv:2205.13542.
[2] Chen, X., Zhang, T., Wang, Y., Wang, Y. and Zhao, H., 2022. Futr3d: A unified sensor fusion framework for 3d detection. arXiv preprint arXiv:2203.10642.
[3] Zoph, B., Ghiasi, G., Lin, T.Y., Cui, Y., Liu, H., Cubuk, E.D. and Le, Q., 2020. Rethinking pre-training and self-training. Advances in neural information processing systems, 33, pp.3833-3845.
[4] Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J. and Krueger, G., 2021, July. Learning transferable visual models from natural language supervision. In International Conference on Machine Learni美白針,ng (pp. 8748-8763). PMLR.
[5] Chen, Q., Wang, J., Han, C., Zhang, S., Li, Z., Chen, X., Chen, J., Wang, X., Han, S., Zhang, G. and Feng, H., 2022. Group detr v2: Strong object detector with encoder-decoder pretraining. arXiv preprint arXiv:2211.03594.
[6] Jaderberg, M., Dalibard, V., Osindero, S., Czarnecki, W.M., Donahue, J., Razavi, A., Vinyals, O., Green, T., Dunning, I., Simonyan, K. and Fernando, C., 2017. Population based training of neural networks. arXiv preprint arXiv:1711.09846.
[7] Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. and Wayne, G., 2019. Experience replay for continual learning. Advances in Neural Information Processing Systems, 32.
[8] Dohare, S., Mahmood, A.R. and Sutton, R.S., 2021. Continual backprop: Stochastic gradient descent with persistent randomness. arXiv preprint arXiv:2108.06325.