Meta-learning
Meta-LearningA subfield of machine learning in which automated learning algorithms are applied to the metadata of machine learning experiments.
The core of meta-learning is the "learning (training)" process, which is to study how to make neurons make good use of old knowledge so that they can adjust themselves according to new tasks. The schematic diagram is as follows:
- Initial parameters of the neural network (blue ■);
- Optimizer parameters (pink ★).
There are two main parts in the graph that need to be trained:
- We use the word "model (M)" to refer to our previous neural network, which can now also be understood as a low-level network. The weights of the model are represented by ■ in the figure.
- We use the term “optimizer (O)” or “meta-learner” to refer to the high-level model used to update the weights of the low-level networks. The weights of the optimizer are represented by ★ in the figure.
References
【1】Understanding Meta-Learning from Scratch (Machine Mind)