MIT EECS | Takeda Undergraduate Research and Innovation Scholar
DNN Structures to Ensure Safety for Modeling Dynamical Systems
- Artificial Intelligence and Machine Learning
Dynamical systems describe the temporal trajectory of a point in space, and can be applied to physical and chemical systems. Deep neural networks (DNNs) have recently been shown to be able to model dynamical systems efficiently and accurately. However, current DNN implementations fall short in creating dynamical models that preserve the conservation of mass of a system, which can lead to safety problems when applied in real life.
A recently explored method of preserving conservation laws involve a projecting the DNN function in a trajectory that preserves a targeted conservation law. The project being proposed would involve modeling a dynamical system with a DNN and determining whether this projection method produces a stable, efficient dynamical model.
I am participating in SuperUROP to learn how proper self-guided research is done. Over the past two summers, I have completed a machine learning UROP and had an NLP research paper accepted to a workshop. I hope to continue to learn the skills that only come with doing real research in order to prepare myself for grad school and help me decide whether I would like to pursue it as a career.