Foundation of Meta-learning for Multi-dimension classification and application

Machine Learning (ML) has become a prominent backbone of Artificial Intelligence (AI). ML algorithms solve different large-scale science and engineering problems and yield explainable outputs. Numerous research studies have investigated ML to seek an optimized solution that solves the problems. The general theme characterizing this dissertation is to address typical machine learning limitations during the learning processes. An important goal of this research is to contribute to machine learning limitations such as lack of samples in the training process, time efficiency, and low accuracy performance. Meta-learning has played the main role in this dissertation to alleviate the limitation of lack of samples in the training process.

Meta-learning (MTL) is the process of learning to learn providing a new direction for scientists to generate a rich model. MTL provides different learning models and training processes for the conventional classification step. One of the popular models is zero-shot learning (ZSL) in which we learn from seen classes to recognize unseen classes.

We start off with recognizing unseen animals and class categories using zero-shot learning. %which has become popular and challenging in advanced machine learning.Then, we apply the solution to different data domains from computer vision to signal processing, and from 2D images to 3d point cloud models. Finally, we are inspired by zero-shot learning to propose an efficient classification algorithm called meta-semantic learning (Meta-SeL) which efficiently recognizes and classifies 3D objects appeared in 3D point clouds models.