Ινστιτούτο Πληροφορικής & Τηλεπικοινωνιών

string(13) "dissertations"

Vasileios Gkatsis

Integration of prior knowledge into machine learning systems taking into account explainability requirements and data...

Start date:October 1, 2023
End date:October 31, 2027
University:Department of Informatics and Telecommunications, National and Kapodistrian University of Athens
PhD candidate:Vasileios Gkatsis
Supervisors:Vangelis Karkaletsis George Paliouras
Related groups:INSANE

Abstract:

Machine learning (ML) has transformed domains ranging from image recognition to natural language processing. However, current approaches face limitations when applied on settings of varying data scarcity.
Another challenge relates to how recent models can integrate problem setting constraints, such as natural principles or domain knowledge (including common knowledge, research findings, expert intuition, etc.).

Informed machine learning is the research field that faces this second challenge, by creating hybrid models which combine data-driven techniques with external/prior knowledge. The integration of such knowledge typically occurs at the following stages of a ML pipeline:

Training Data: Prior knowledge is embedded into the training data, e.g. by augmenting datasets with simulated data[9] or crafting specialized input features.
Model Architecture: Knowledge is reflected e.g. in the architecture of neural networks (NNs) parameterized on the specific problem.
Learning Algorithm: Prior knowledge can be integrated using constraints, such as additional regularization terms in the loss function of a model.

The above approaches effectively bridge ML with natural sciences. For example physics-informed neural networks (PINNs) are NNs developed specifically to solve physics-related problems. A last challenge of recent models is the interpretability and explainability of their function and output. An interpretable model allows humans to grasp its inner processes, while an explainable one provides clear justifications for the results produced.

This research will (a) explore the capabilities of integrating prior knowledge into machine learning systems for natural sciences, (b) research novel approaches to enhance explainability in these settings, (c) examine alternative approaches of the above under different data scarcity constraints.

« Back to list