Skip to content
Recently, the intersection of physics, mathematics, and machine learning has promoted the use of machine learning frameworks to optimize physical models, further encouraging researchers to develop many exciting new machine learning models (such as Neural ODEs, Hamiltonian Neural Networks, etc.) that draw on concepts from physics.
Researchers from Stanford University’s Shanhui Fan group are particularly interested in the perspective that physics itself can serve as a computational engine.In other words, they are interested in physical systems that can be used as hardware accelerators or dedicated simulation processors for fast and efficient machine learning computations.
The corresponding author of this article, Professor Shanhui Fan from Stanford University
In their recent paper published in Science Advances, they demonstrated that the physical properties of waves can be directly mapped to the time dynamics of RNNs.Using this connection, the researchers showed that acoustic/optical systems (numerical models developed through PyTorch) can be trained to accurately classify vowels from recordings of human speakers.Essentially, this involves inputting vowel waveforms into a physical model and allowing the optimizer to add and remove materials at 1000 different points within the domain, essentially acting as the model’s weights.
Because this machine learning model corresponds to a physical system, it means that a trained material distribution can be
