This Week in ML & AI avatar

Dexa/This Week in ML & AI

Learn more
Published Jul 22, 2019

Retinal Image Generation for Disease Discovery with Stephen Odaibo - TWIML Talk #284

Explore the transformative power of AI in ophthalmology with Dr. Stephen Odaibo, as he delves into the critical intersection of GANs and data quality in medical imaging, and the promising potential for AI-driven solutions in addressing healthcare challenges and shortages worldwide.
Episode Highlights
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) logo

Topics covered

Popular Clips

Episode Highlights

  • GANs Insights

    Stephen Odaibo shares his insights on the application of GANs in medical imaging, emphasizing the importance of domain-specific knowledge. He argues that the future of machine learning lies in interdisciplinary teams with deep expertise in both machine learning and the specific field of application, such as healthcare. Stephen's journey with GANs began with a theoretical question about their potential for data augmentation, leading him to explore their capabilities and limitations.

    It's an enormous value in today, in ML, and I think we're all going to. I think people are starting to get that the initial big hype and excitement that computers are going to take over is starting to dampen, and people are starting to realize that the only real way forward is going to be, is going to be domain specific, and it's going to be with really integrated interdisciplinary teams and ideally, people who have detailed, expert level knowledge of both the ML side, as well as whatever other field that they're looking to apply ML to, be it agriculture, be it transportation, be it healthcare.

    ---

    He highlights the necessity of understanding the domain well, as machine learning is inherently heuristic, and theories must encapsulate the general properties of data 1 2.

       

    Image Augmentation

    In exploring GANs for image augmentation, Stephen Odaibo discusses the nuanced role of data in machine learning. He challenges the notion that more data is always better, suggesting that the type and specificity of data are crucial for solving particular problems. Stephen emphasizes that understanding the data landscape, including its unique characteristics and subfamilies, is essential for effective machine learning applications.

    If you're trying to capture the essence of a certain neighborhood of the data distribution, you're better off sampling from that area than coming up with saying, I have 10 million images of this broad, ill defined data class, when in reality they are truly subfamilies, multiple subfamilies within that whole area.

    ---

    He believes that the human visual cortex and experience will remain crucial in guiding these processes for the foreseeable future 3 4.

Related Episodes