IC Colloquium: Cameras for better machine learning, and machine learning for better cameras
By: Kristina Monakhova - MIT
IC Faculty candidate
Abstract
Cameras are everywhere powering everything from self‑driving cars to medical diagnostics and scientific discovery, creating massive economic value and saving (or failing to save) human lives. But automation has fundamentally changed their purpose: people don’t look at these images, algorithms do. Our current cameras are optimized for the wrong goal: sharp, high‑resolution images for human consumption, rather than information‑dense images to be used by algorithms. Using machine learning, we can design better, more capable imaging systems; with better imaging systems, we could have more robust, better-informed intelligent systems.
In this talk, I will demonstrate how we can make imaging systems more capable by using physics-informed machine learning, which combines imaging system physics with deep learning. First, I will show how we can make tiny, lensless cameras have image quality comparable to lensed cameras. Next, I will show how with our algorithms, we can push the limit of what cameras can see in the extreme dark by an order of magnitude, enabling photorealistic videos of moving objects on a clear, moonless night with no external illumination (submillilux). In addition, I will demonstrate how we can design cheap, compact, and capable computational cameras and microscopes that capture higher-dimensional information, such as 3D and multiple wavelengths of light, which could be useful for high-level tasks. My future research agenda expands upon this to design optics and algorithms together with higher level tasks in order to create the next generation of intelligent cameras and microscopes that are optimized for high‑level insights rather than images.
Bio
Kristina Monakhova is a postdoctoral fellow at MIT, supported by the MIT Postdoctoral Fellowship for Engineering Excellence. She received her PhD from UC Berkeley in Electrical Engineering and Computer Sciences in 2022, where she was a member of Laura Waller’s Computational Imaging research group. Her research focuses on making more capable cameras and microscopes through the co-design of imaging systems and algorithms. Her research lies at the intersection of signal processing, machine learning, and optics. She is a recipient of the NSF Graduate Research Fellowship and the NDSEG Fellowship.
More information
IC Faculty candidate
Abstract
Cameras are everywhere powering everything from self‑driving cars to medical diagnostics and scientific discovery, creating massive economic value and saving (or failing to save) human lives. But automation has fundamentally changed their purpose: people don’t look at these images, algorithms do. Our current cameras are optimized for the wrong goal: sharp, high‑resolution images for human consumption, rather than information‑dense images to be used by algorithms. Using machine learning, we can design better, more capable imaging systems; with better imaging systems, we could have more robust, better-informed intelligent systems.
In this talk, I will demonstrate how we can make imaging systems more capable by using physics-informed machine learning, which combines imaging system physics with deep learning. First, I will show how we can make tiny, lensless cameras have image quality comparable to lensed cameras. Next, I will show how with our algorithms, we can push the limit of what cameras can see in the extreme dark by an order of magnitude, enabling photorealistic videos of moving objects on a clear, moonless night with no external illumination (submillilux). In addition, I will demonstrate how we can design cheap, compact, and capable computational cameras and microscopes that capture higher-dimensional information, such as 3D and multiple wavelengths of light, which could be useful for high-level tasks. My future research agenda expands upon this to design optics and algorithms together with higher level tasks in order to create the next generation of intelligent cameras and microscopes that are optimized for high‑level insights rather than images.
Bio
Kristina Monakhova is a postdoctoral fellow at MIT, supported by the MIT Postdoctoral Fellowship for Engineering Excellence. She received her PhD from UC Berkeley in Electrical Engineering and Computer Sciences in 2022, where she was a member of Laura Waller’s Computational Imaging research group. Her research focuses on making more capable cameras and microscopes through the co-design of imaging systems and algorithms. Her research lies at the intersection of signal processing, machine learning, and optics. She is a recipient of the NSF Graduate Research Fellowship and the NDSEG Fellowship.
More information
Practical information
- General public
- Free
Contact
- George Candea