One of the rapidly growing areas in machine learning is meta-learning. Meta-learning or learning to learn focuses on designing models that learn and adapt to the new environment with few training examples. The significant role of meta-leaning is determining suitable algorithms that generate better predictions from the datasets.
Meta-learning assists machine learning by solving the challenges such as high operational costs, the need for large datasets for training, and trials that take a long time to fit the best model. Merits of meta-learning are higher model precision accuracy, faster and cheaper training process, building more generalized models, more adaptability to environmental changes, optimized model architecture and hyper-parameters, and faster AI systems.
Common meta-learning approaches are model-based utilize networks with external or internal memory, metrics-based determine learning effective distance metrics, and optimization-based provide explicitly optimizing model parameters for fast learning. Model based meta-learning models are Memory-Augmented Neural Networks and Meta Networks. Metrics-based meta-learning models are Convolutional Siamese Neural Network, Relation Network, Matching and Prototypical Networks. Optimization-Based meta-learning models are LSTM Meta-Learner, Temporal Discreteness and Reptile.
Recently, Deep meta-learning has been employed to enable the ability of the networks to learn new concepts quickly. Application fields of meta-learning in AI are few-shot learning, robotics, intelligent medicine, and unsupervised learning. Some other applications area of meta-learning is Computer Vision and Graphics-such as classification, object detection, object segmentation, density estimation, Language and Speech-such as language modeling and speech recognition, Abstract Reasoning, Meta Reinforcement Learning-such as 2D Navigation and Locomotion using Reinforcement learning, Environmental learning, and Continual Learning. Future scopes of meta-learning are meta-generalization, distribution over tasks, and visual imitation learning.