Featured Page

Today's featured page showcases the Strength Distribution, a lightweight machine learning technique we have used for embedding adaptive behavior in devices that participate in an Internet of Things (IoT). 



This video demonstrates a novel machine learning algorithm called Strength Distribution, developed by Prof. William McEneneay and Dr. Rajdeep Singh. The Strength Distribution performs online learning, meaning that it doesn't require a separate training period. It learns as it goes. And as the situation in the world changes, the Strength Distribution quickly adapts. More information on the Strength Distribution can be found in the inventors' presentation at the 12th ICCRTS Conference

The video demonstrates the Strength Distribution's ability to learn a new model quickly and adapt the learned model as the real model changes. Consider a child who has favorite and not so favorite colors. At first, we don't know what the child's preferences are, but we start naming colors, and if we pick a color that the child likes a lot, he gives us lots of tiddly-winks, If we name a color that the child likes a little, he gives us a few tiddly-winks, and if we name a color that the child doesn't like at all he gives us nothing. The Strength Distribution processes this information and after a few rounds, we learn a very good representation of the child's preferences. But children, being fickle, change their preferences without notice. And about half way through the video, the child goes to a completely new set of preferences. All we know for certain is that we aren't getting as many tiddly-winks as we used to. We have to unlearn everything we have learned and learn the new distribution. 

In the video frame, the upper left represents the child's actual color preference (bigger balloons are higher preference) and the upper right represents the learned perception of the child's preference. When the picture at the upper right looks like the picture at the upper left, the child's preferences have been learned. 

YouTube Video


Here is the first physical prototype we built using the Strength Distribution back in 2009. In fact, we built about a dozen of these and tested them in a perimeter breech scenario. Each device had several sensors (motion, 3-D accelerometer, GaAs photocell, and a thermometer). Each device used the Strength Distribution to determine whether to send sensor data to its neighbors, fuse its data with data from neighbors and send the fused consensus, or do nothing and conserve battery power. The constellation of devices accurately detected the presence of an intruder crossing the perimeter, and messages were relayed through the ad hoc network to alert a human responsible for perimeter security.  




Minimalist Project Management

“The number of people having any connection with the project must be restricted in an almost vicious manner. Use a small number of good people (10 percent to 25 percent compared to the so-called normal systems).”

Kelly Johnson (Founder, Lockheed Skunk Works), Management Rule #3
S3 Data Science Logo
S3 Data Science, copyright 2015.