Deep learning algorithms have become a standard tool in many industries ranging from e-commerce to healthcare. Their ability to translate data into intelligent predictions completely transformed these industries, increasing productivity and controlling costs. These successes sometimes obscure inherent vulnerability of these models such as difficulty in generalizing in the presence of distribution shifts and ineffectiveness in estimating confidence of their prediction. In this webinar, the detail of these vulnerabilities will be illustrated and possible algorithmic solutions to these challenges will be described.

• Better understand the landscape of clinical AI applications
• Identify some intrinsic vulnerabilities of these approaches
• Learn about potential algorithmic solutions

Regina Barzilay
Professor, Electrical Engineering and Computer Science Department
MIT