Skip to main content
Browse by:

+DSvLE: Simulation-based machine learning for computational microscopy

Event Image
Icon calendar
Wednesday, October 20, 2021
Icon time
4:00 pm - 5:00 pm
Icon speaker
Srinivas Turaga, PhD; Computation & Theory, HHMI Janelia Research Campus

Supervised machine learning is a well-established paradigm for training deep neural networks. However, there are many settings where it is challenging or impossible to collect ground truth data for supervised learning. In this session, Dr. Srinivas Turaga will describe how to use the framework of auto-encoders and combine differentiable physics-based simulations with deep networks to solve two problems in computational microscopy.

1. Single molecule localization microscopy is fluorescence based super-resolution microscopy technique. Dr. Turaga will explain how a biophysically realistic simulation of such microscopy data, along with the appropriate neural network architecture, and a new spatial point process based loss function can enable exciting new capabilities for such microscopy.
Speiser, Muller, et al, "Deep learning enables fast and dense single-molecule localization with high accuracy", Nature Methods 2021.

2. Programmable microscopy is enabled by newly available programmable optical elements. Such microscopes offer the possibility of optimizing the imaging process to the biological specimen to achieve the best possible trade off of spatial resolution, temporal resolution, and signal to noise. Dr. Turaga will show how to build a differentiable wave-optics simulation of a programmable microscope, and to combine it with a new neural network architecture to engineer microscope parameters for fast 3D snapshot microscopy of whole brain neural activity in larval zebrafish.

Deb et al, "Programmable 3D snapshot microscopy with Fourier convolutional networks", arxiv 2021.

Presented by Srinivas Turaga, PhD; Computation & Theory, HHMI Janelia Research Campus

This session is part of the Duke+DataScience (+DS) program virtual learning experiences (vLEs). To learn more, please visit