Meta-Learning for Monitoring Environment Systems Across the Globe

Abstract

Data sparsity is a key challenge in monitoring climate because of the lack of quality data, problems in sensors, lack of historical data, or financial constraints in certain parts of the world. Thus, monitoring the environment using machine learning becomes a difficult task because classic machine learning algorithms’ main objective is to train a model that uses input features to learn classes. This paradigm requires huge datasets which makes it difficult to train models in tasks where data is sparse. Meta-learning, or learning-to-learn is a learning paradigm which provides an alternative methodology to classic machine learning algorithms. Meta-learning uses machine learning models in various learning episodes and uses this experience to learn in new learning environments. Thus, meta-learning can be used to monitor environment systems by training in scenarios where data is available and leveraging that information in data sparse tasks.