GedankenNet: Self-Supervised AI Model for Computational Imaging
Researchers from the UCLA Samueli School of Engineering have developed GedankenNet, an artificial intelligence-based model for computational imaging and microscopy. Unlike existing models, GedankenNet does not require training with experimental objects or real data.
The Challenge in Applying AI to Microscopy
Artificial intelligence has transformed imaging in various fields, but applying it to microscopy presents persistent challenges. Current AI models rely on human supervision and large-scale, pre-labeled datasets, making experiments laborious and costly. Moreover, these models struggle with processing new sample types or experimental setups.
Borrowing from Einstein’s Thought Experiment
The UCLA team drew inspiration from Albert Einstein’s Gedanken experiment approach. They developed GedankenNet using conceptual thought experiments and the laws of physics that govern electromagnetic wave propagation in space.
Without any real-world experiments or actual data, GedankenNet reconstructed microscopic images using random artificial holograms created solely from imagination.
The researchers then tested GedankenNet using 3D holographic images of human tissue samples. In its first attempt, GedankenNet successfully reconstructed microscopic images of human tissue samples and Pap smears from their holograms.
Compared to existing methods that rely on supervised learning with large-scale experimental data, GedankenNet showcased superior generalization to unseen samples without any prior information. It also accurately represented the physics of wave equations, generating output light waves that faithfully depicted 3D light propagation in space.
This breakthrough demonstrates the potential of self-supervised AI for learning from thought experiments, opening up new opportunities for developing physics-compatible and broadly applicable neural network models.
The paper was authored by Professor Aydogan Ozcan and his research team, including graduate students Luzhe Huang and Hanlong Chen, as well as postdoctoral scholar Tairan Liu from the UCLA Electrical and Computer Engineering Department.
GedankenNet has the potential to revolutionize computational imaging and microscopy by eliminating the need for laborious and costly experiments and enabling the processing of new sample types and experimental setups.