- Machine Learning
- Multi-agent Systems
- Artificial Intelligence
- Computer Security
Intra-class Clustering
Many problem domains in supervised machine learning must overcome the challenge of inbalanced class labels. For example, consider that challenge of deciding whether an image contains a cat or a dog. If the set of training examples is imbalanced (10 images of cats and 90 images of dogs), the learning algorithm will likely skew towards the dog classification label. Intra-class clustering is a novel technique to automatically identify sub-cluster within a label space and thereby produce a more balanced distribution of class labels. More details including source code and binary packages for a weka add-on can be found here
Smart Environments
Before smart environments can see large-scale adoption in the real-world several challenges must be overcome.
Security and Privacy
Too often in the rush to get the next greatest technology out the door security and privacy concerns are overlooked. I am currently investigating ways that smart environments can gather and share information from a variety of sources while still protecting that information and giving consumers control over their own data.
Transfer Learning
One major challenge is obtaining sufficient information to understand the current context of the operating environment. This is where transfer learning can come into play. I am currently working on projects which enable the transfer of knowledge between heterogenous sources. This will greatly reduce the burden that is often associated with training the smart environment.
Vizualizations
Another challenge facing smart environments is displaying useful information to the user. We are currently looking into what kinds of health information can be obtained from smart environments and how to show that information to individuals, caregivers, nurses and doctors.
Annotation
Obtaining annotated data can be a difficult and time-consuming process. I have developed the Real-time Annotation Tool (RAT) as a cross-platform tool designed to annotate data streams in real-time. This tool can be found here.
We have also developed a mobile version(Android) for use with the Washington State Psychology Department's Night-Out Task.
Transition Detection
A crucial component of smart environments is the ability to correctly interpret the current context which can then be used to inform any decision the system might make. One such context which may be useful to the systems is understanding when transitions are occuring between activities. This may be an ideal time to deliver notifications, solicit actions or otherwise request the user's attention. I am currently working on a project to detect such activity transitions in real-time using both supervised and unsupervised learning techniques.