Meta-Learning in Computer Vision
Wenfeng Wang · 2023
Early machine learning was not end-to-end, that is, it did not directly input raw data at the input end, but based on human prior knowledge, preprocessed the raw data and used the extracted features as input. So early machine learning was also known as feature engineering. Due to the fact that the input raw data is manually processed before being used to train the model, it is impossible to avoid situations such as incorrect feature selection, inaccurate feature extraction, and even significant deviation in feature calculation results. With the development of machine learning, especially the emergence of multi-layer neural networks, the end-to-end learning process has been achieved, marking the transition of machine learning from feature engineering to representation learning. Deep learning is further developed on this basis. Deep learning models have the ability to learn features. Because by configuring relevant parameters, multi-layer feedforward neural networks can approximate any function. Metalearning belongs to the category of machine learning. The ultimate goal of machine learning algorithm research is to enable machines to have learning abilities that are close to those of humans. This is a very big challenge! The current mainstream machine learning methods, whether classical machine learning theories or more advanced deep learning models, cannot achieve this. The emergence of meta learning is precisely to enable machines to have learning abilities that are close to those of humans. Machine learning theory is evolving from "mechanical memory" to "active learning and application", which is an inevitable trend. Just like a person's pursuit of a student career. At the beginning, we had relatively little consideration for learning methods, mainly relying on the matching of memory and knowledge to complete our preliminary understanding of the world. When the difficulty of knowledge increases, simple memory and matching can no longer meet the needs of learning, we have to start to pay attention to learning methods. This book will provide a learning-to-learn theory for the machine brain, which will help overcome the limitations of existing machine learning algorithms.