AI Talk: Privacy, beer, marketing and doodling

March 29th, 2019 / By V. “Juggy” Jagannathan, PhD

This week’s AI Talk…

Federated Learning to overcome privacy issues

Anyone working with health information knows that getting access to clinical data for any purpose is a non-trivial exercise. PHI—protected health information—is protected by regulations, in particular by the Health Insurance Portability and Accountability Act (HIPAA). So, if you need patient data from multiple clinical sites, even if the entities want to collaborate, one cannot pool the data resources to run machine learning algorithms, as the PHI in the data is protected and cannot be shared. One can de-identify the data, but that is a very expensive proposition. One approach to this conundrum is to get the data directly from the patient who opts into the research program and provides access to their data. This is the approach taken by National Institutes of Health “All of Us” research program, created as part of the precision medicine initiative started at the end of the last administration. All of Us seeks to recruit hundreds of thousands of participants over the next decade. The other approach is a federated one, covered quite nicely by this Technology Review article. The basic idea is this: Each hospital or entity trains a partial deep learning model with the data that they have. The partial model (with no PHI) can then be combined into a global model. Google published a blog on this idea a few years ago, but recently MIT researchers developed a variation on the approach called “Split Learning.” In Split Learning, the machine learning happens in almost independent silos, with strategic integration of the model layer and clever ways to propagate and minimize errors. They show that, using this approach, the model learned is just as good as pooled data with the bonus of individual institutions not needing to have huge computing resources to run their training. Of course, the other main reason for this approach is you protect PHI as it remains within the walls of a particular institution.

Optimizing the last mile

There is an AI conference organized by MIT Technology Review happening this week. Lots of talks and presentations on various AI related advances and predictions. This one caught my attention as being quite practical. In this report, a company called Wise Systems is developing optimization solutions using machine learning for the last mile. By analyzing logistical data on delivery, they can figure out what time deliveries should be scheduled for to minimize time spent in transit. Who is using it? As an example Annheuser-Busch—you really need to get your beer in a timely fashion!

AI Marketing by Microsoft

While on the topic of deploying AI across industries, Microsoft has adopted an interesting tactic. This month it launched what it has dubbed AI Business School. This is an online curriculum targeted at managers and executives in companies who are explaining the strategic implications of adopting AI by discussing a variety of AI case study deployments across different industries.

Want to become a painter? A real artist?

It looks like now anyone can—with a bit of help from an AI-driven software called GauGAN developed by NVIDIA. The deep learning trick that this application harnesses is one called Generative Adversarial Networks (GANs). As their name implies, GANs use two networks, one that generates images and another (a discriminator network) to decide whether the images are real or not. By training on millions of images, the generative part becomes so good as to fool the discriminator. This is the same approach now being used to generate fake videos of celebrities. In this tool, NVIDIA, takes a palette approach to  rendering objects. In a canvas, you pick an object to draw—a rock, a pond, a landscape with trees and the simple objects are transformed into photorealistic images. Take a look at the video posted here—it’s quite stunning! Of course, on a practical front, this might help city planners and architects to create amazing visualizations of their creations.

CMS is launching an AI Health Outcomes Challenge

CMS has just launched an AI Health Outcomes Challenge with a pot of gold worth $1 million. The goal is to predict unplanned hospital and skilled nursing facility (SNF) admissions and adverse events.


The AI-doodles article was from my colleague, Marty Taylor.

I am always looking for feedback and if you would like me to cover a story, please let me know. “See something, say something”!  Leave me a comment below or ask a question on my blogger profile page.

V. “Juggy” Jagannathan, PhD, is Vice President of Research for M*Modal, with four decades of experience in AI and Computer Science research.