| Summary: | "With the fundamental breakthroughs in artificial intelligence and the significant increase of digital healthcare data, there's been enormous interest in AI for healthcare applications. One rapidly developing area is the use of deep neural networks for medical imaging, with applications ranging from diagnosing chest X-rays to the early detection of Alzheimer's to identifying cancer in pathology slides. Despite this variety of applications, there remain some crucial unanswered questions. On the methods side, there's been a premature convergence on a specific model-development strategy: deep neural networks are first trained on natural image data, and then fine-tuned (transferred) to work on the medical data. Maithra Raghu (Cornell University, Google Brain) explores this process and shows that contrary to conventional wisdom, this standard method of model development isn't guaranteed to provide the benefits it's believed to, and she suggests simple and effective alternate methodologies. On the applications side, there's been little exploration of the interaction of these medical AI algorithms with human experts, with existing literature typically evaluating the algorithm in isolation and the human experts in isolation--vastly different from a realistic deployment scenario. Maithra examines the essential question--the role of human experts--which provides new, crucial prediction problems to study and significant benefits through the effective combination of artificial and human intelligence. This session is from the 2019 O'Reilly Artificial Intelligence Conference in San Jose, CA."--Resource description page
|
|---|
