Meta-learning: learning to learn. Generalization of the ability of neural networks for task series

Context

The neural networks have demonstrated their strong capabilities in various domains, such as recognizing objects into images, translating language and playing games. However, a neural network is only trained for one specific task and the ability cannot be adopted in another scenario. A neural network that can recognize handwriting, however, cannot find out the differences between cats and dogs.
The meta-learning method describes the philosophy of "learning to learn". A set of tasks are considered together and the generalization is learned. A machine learning model not only learns the tasks themselves, but also how it can train itself better.
New methods for meta-learning are investigated within the scope of this work, depending on the focus selected. Possible topics are e.g: Meta-learning in multi-task learning environments, imitation learning, adaptive learning, etc.

Taskn

The specific tasks relate to the chosen topic, but the general process of the work can be summarized as follows:

• Literature research on the SOTA technique in the relevant field
• Conceptualization of the work regarding the workload (e.g. classification, regression and reinforcement learning problems)
• Design, training and optimization of neural networks (e.g. feature selection, network pruning & quantization, knowledge distillation), comparison with other SOTA or classical methods
• Generalization of the framework to enable good performance on different tasks
• Evaluation and analysis of the results and writing of the final thesis

Prerequisites

• Motivation and interest in solving technical problems independently
• Knowledge of machine learning, ideally reinforcement learning
• Experience with programming (Python, C++, Java...)
• Analytical, problem-solving and communication skills