数据闭环的另一个方面就是自动化的模型训练。自动化模型训练有两个关键技术,第一是自动化模型搜索(AutoML),第二是持续学习(Continual Learning)。自动化模型搜索指让训练系统自动进行模型调优,AutoML在过去几年是比较红火的研究方向,也有许多论文和实践的探索。百度使用的是一种基于进化算法改进的方案,主要搜索模型的超参数,如任务的权重、optimizer的参数等。感兴趣的同学可以参考论文Population-based training [6]。
而这里提到的持续学习是最近AI研究者们更为关注的课题,深度学习在新的数据持续注入模型训练的过程中,会体现出两个缺陷:(1)灾难性遗忘(catastropic forgetting),即学了新的数据以后在旧的数据上容易发生遗忘;(2)可塑性损失(loss of plasticity),即模型在多次训练以后,在新数据上的学习能力变差/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 Learning (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.