Symbolic Regression for Efficient Time Series Modelling on Ultra Low Power Embedded Platforms

Context

Symbolic regression is a subfield of machine learning (ML) that aims to find the analytical expression that fits the data set. Widely used in the modelling of physical phenomena [1], symbolic regression provides a way to fit complex data with a concise mathematical representation, allowing for compact machine learning model design and the possibility of model interpretation. Both properties are crucial for the highly integrated AI systems in industrial applications, where the lack of efficiency and interpretability of classical deep learning models limits their application. This work aims to explore the applicability of symbolic regression for building hardware-efficient ML models.

Goals

In this work the research of the state of the art in the symbolic regression, existing methodologies and frameworks will be conducted. Then the methodology for the hardware-aware symbolic regression must be developed. To prove the methodology experiments on the selection of the synthetic and real-life application datasets for the sequence modelling will be performed. Finally, the optimized networks have to be evaluated on real hardware following followed by the analysis of the applicability of symbolic regression methods for hardware-aware ML.

[1] Angelis, Dimitrios, Filippos Sofos, and Theodoros E. Karakasidis. “Artificial Intelligence in Physical Sciences: Symbolic Regression Trends and Perspectives.” Archives of Computational Methods in Engineering 30, no. 6 (July 1, 2023): 3845–65. https://doi.org/10.1007/s11831-023-09922-z.

Requirements

• Experience with Deep Learning (DL) and one of the DL Frameworks like PyTorch or Keras/TensorFlow
• Experience with Python and C++
• Motivation and independent working style