Hands-On Reinforcement Learning for Games

Learning 2 learn

Any good learning model should be trained across a variety of tasks and then generalized to fit the best distribution of those tasks. While we have covered very little with respect to general machine learning, consider the simple image classification problem with a deep learning model. We would typically train such a model with one goal or task, perhaps to identify whether a dataset contains a cat or dog, but not both, and nothing else. With meta learning, the cat/dog dataset would be one training entry in a set of image classification tasks that could cover a broad range of tasks, from recognizing flowers to cars. The following example images demonstrate this concept further:

Example of meta learning applied to image classification ...