2023-Japan-Barbastathis

Conference Video|Duration: 35:10
January 27, 2023
Please login to view this video.
  • Video details
    If you point your camera to a scene, and the camera registers nothing meaningful—does it mean that nothing was really there? Hardly! Even if a human observer cannot detect and interpret it, much information may still have been recorded in the pixels. How, then, should one capture and decode it to reveal the hidden scene? 

    With my research group, we have decoded several challenging scenes with important applications for industry. For example, we have peeked inside integrated circuits non-invasively to work out if their manufactured topology matches the design file; quantified mechanical effects in the retinal fibrous structures and vasculature to forecast glaucoma progression; and measured the particle size distribution in drying powders toward early detection of undesired agglomeration events.

    In all these cases, even the most advanced state-of-the-art imaging methods cannot capture the relevant phenomena with sufficient fidelity or economy.  It is a unique feature of our work that physical models are explicitly weaved into data-driven models. Thus, our algorithms perform well in test cases, and are also interpretable and resilient. We have also demonstrated significant savings: for example, reduction by two orders of magnitude in total scanning and computation time.

    Funding acknowledgments: Parts of this work were funded by the United States Intelligence Advanced Research Projects Activity (IARPA); Singapore’s National Research Foundation (NRF); and by Takeda Development Centre Americas, Inc. (successor in interest to Millennium Pharmaceuticals, Inc.)

Locked Interactive transcript
Please login to view this video.
  • Video details
    If you point your camera to a scene, and the camera registers nothing meaningful—does it mean that nothing was really there? Hardly! Even if a human observer cannot detect and interpret it, much information may still have been recorded in the pixels. How, then, should one capture and decode it to reveal the hidden scene? 

    With my research group, we have decoded several challenging scenes with important applications for industry. For example, we have peeked inside integrated circuits non-invasively to work out if their manufactured topology matches the design file; quantified mechanical effects in the retinal fibrous structures and vasculature to forecast glaucoma progression; and measured the particle size distribution in drying powders toward early detection of undesired agglomeration events.

    In all these cases, even the most advanced state-of-the-art imaging methods cannot capture the relevant phenomena with sufficient fidelity or economy.  It is a unique feature of our work that physical models are explicitly weaved into data-driven models. Thus, our algorithms perform well in test cases, and are also interpretable and resilient. We have also demonstrated significant savings: for example, reduction by two orders of magnitude in total scanning and computation time.

    Funding acknowledgments: Parts of this work were funded by the United States Intelligence Advanced Research Projects Activity (IARPA); Singapore’s National Research Foundation (NRF); and by Takeda Development Centre Americas, Inc. (successor in interest to Millennium Pharmaceuticals, Inc.)

Locked Interactive transcript