Summer 2009 Project Areas Overview

Computer Vision

The Computer Vision Laboratory is conducting exciting new research in Optical Character Recognition. Our methods, strangely enough, do not care what letters look like, but only use the pattern of repetitions.
For example, we could recognize the string "01221221331" as "Mississippi" if we knew that every letter had been replaced by a digit in a consistent way. The reason is that Mississippi is the only word that
follows this pattern of repetitions.

Our work this summer will extend our new techniques to foreign languages that have alphabets different that English, such as Greek and Russian. If you are interested in doing OCR and watching it work on
other foreign languages, please come join us!

Medical Safety

Medical errors cause approximately 98,000 patients to die each year. An Institute of Medicine report suggests that the delivery of healthcare must fundamentally change to address medical error. In particular, it is now widely recognized that medical errors result from systemic rather than individual failures, leading the Institute of Medicine report to advocate the development of healthcare systems that directly address patient safety. In particular, the IOM report states, "what is most disturbing is the absence of real progress… in information technology to improve clinical processes."

To address this concern, we have begun to investigate how current research in process definition and execution and in software verification and analysis can be applied and extended to help reduce errors and improve safety and efficiency in medical processes. In particular, we are exploring how to extend some current software engineering research techniques to define, monitor, coordinate, analyze, and improve the safety and efficiency of medical processes. Working with experts on medical safety and building upon our experience with process languages and with system verification and analysis, we plan to develop visual process representations for critical paths of care that capture, not only the standard paths, but also describe the exceptional situations that can arise and the inherent concurrency and multi-tasking frequently undertaken by extremely busy healthcare providers, and to provide the basis for careful analysis and evaluation leading to safety and efficiency improvements.

Specific projects include:

Developing and helping visualize the results from discrete event simulations will help medical professionals make decisions about resource allocations. For example, to help reduce the waiting time for an emergency room visit, which currently is six hours, would it be better to increase the number of beds, number of doctors, or number of nurses?

Several of the analysis tools find traces through the process definitions that seem to lead to errors. Since the error prone paths tend to be rather convoluted, techniques to help track down the "real" cause of the error need to be developed.

Validating a process definition involves comparing that definition to a statement of the requirements. Our PROPEL tool allows these requirements to be stated by selecting from parameterized English phrases, which map to finite-state automata. By interacting with medical professionals, PROPEL requirements need to be developed for the Emergency Room process definition, as well as developing improvements to the tool itself.

Scientific Workflow: The Analytic Web

The Internet has changed forever the way in which science will be done. Worldwide scientific collaborations are beginning to use internet access to create opportunities for scientists to make data available to worldwide communities, thereby enabling expedited collaborations among geographically distributed researchers. While this creates opportunities through the broader availability of more comprehensive scientific analyses, it also creates risks arising from uncertainty about the way in which internet-accessed data were produced and the appropriate ways in which they can be used. Scientific workflow is the application of workflow technology to scientific endeavours, and is becoming recognized as a valuable approach for assisting scientists in accessing and analyzing data.

Working with UMass computer scientists and Harvard Forest ecologists and building upon the existing SciWalker Analytic Web prototype software system, a more complete version of the system for an ecological domain will be developed.

Artificial Intelligence

Research in the Autonomous Learning Laboratory centers on learning in machines and animals. We have been focusing on reinforcement learning (RL), whose main ideas go back a very long way. Especially exciting these days are the connections between temporal difference algorithms, which were developed in this lab, and the brain's reward systems that rely on the neurotransmitter dopamine. Of particular interest at present is what psychologists call intrinsically motivated behavior, meaning behavior that is done for its own sake rather than as a step toward solving a specific problem of clear practical value. What we learn during intrinsically motivated behavior is essential for our development as competent autonomous entities able to efficiently solve a wide range of practical problems as they arise. Recent work in the lab on intrinsically motivated RL is aimed at allowing artificial agents to construct and extend hierarchies of reusable skills that form the building blocks for open-ended learning.

Some possible summer projects include:

Developing good computer game characters using apprenticeship learning. One of the hard aspects of creating entertaining computer games is creating opponents having various levels of abilities. Although people have tried to apply RL to this in the past, it is too slow to do from scratch, and there are far too many variables. As a result, game AI is usually based on some sort of search. However, one thing game companies are not short of is expert players. It should be possible to use traces from their games to do apprenticeship learning (as in Abbeell's helicopter paper, not just imitation because we do not have the reward function), and then use RL to further improve the player's performance.

Deep learning (or Ng's Self-Taught Learning) for feature extraction in Go. These algorithms are essentially general hierarchical feature extractions. Go is a good, visual but manageably complex testbed to see if they can extract feature hierarchies that correspond to known features of the game.

The Wavelet Basis. Like RBFs, but also orthogonal. This would be a nice project but you would need to sift through the math and come up with a fixed basis based on wavelets and compare it to RBFs and the Fourier Basis.

Building skill hierarchies in Wargus. Wargus is a popular strategy game domain that the MAX-Q people have just induced a hierarchy for. It might be a good intrinsic motivation domain that a student could take a crack at. It's Warcraft-y and so quite fun.

Robotics

REU candidates will have the opportunity to perform experiments along side PhD students that study how robots can learn and develop. Our experimental work includes learning how to grasp and manipulate objects, how to interpret interactions with human beings and acquire gestural means of communication, and how "play" can lead to "common sense." In addition, we study new methods for programming robots that are inspired by growth and develpment in human infants, socially situated learning, and learning and adaptation in distributed robots and human-robot teams.We are developing these technologies on Dexter, a large upper-body humanoid, uBot-5, a small and unique concept in mobile manipulation, and a distributed sensor array with an eye toward technologies for personal robotics and future healthcare applications.

Computer Architecture

The Architecture and Language Implementation Lab has projects for
REU students in the general area of tools related to describing
computer instructions and architectures and in using the
descriptions to generate simulators, assemblers, disassemblers,
linkers, debuggers, and compiler components. We are further
interested in building graphical user interfaces to these tools,
in writing specific descriptions, measuring performance, etc.
Most coding is in Java. Much of the work requires a degree of
comfort with computers at the instruction set level (i.e.,
assembly of machine level programming). Compiler experience is
a plus!

Technologies for Teaching & Learning

The RIPPLES group is developing, deploying and assessing multimedia learning technologies and investigating how to use these technologies to support effective teaching and learning inside and outside the classroom. RIPPLES is also evaluating the application of multimedia technologies to collaborative work and outreach to learning and other communities.
Our current research on learning technologies addresses:
- automatic capture of classroom activities from multiple media streams
-- producing navigable classroom records for later study and review
-- distributing materials in real time to enable active, constructivist learning
- extending the RIPPLES MANIC technologies to create a constructivist learning environment by adding tools for extended search, collaboration, and notation
- investigating new technologies for asynchronous multimedia delivery, including Java, SMIL, and other technologies

Mobile Systems: Networking and Security

We are looking for students interested in mobile systems and mobile systems forensics. We operate a large testbed of vehicles and WiFi access points. The vehicles are equipped with computers, radios, and 3G data connections. The APs are operated by the Town on our behalf. We are looking for students who wish to gain experience working with mobile systems and measuring performance of systems and applications and characteristics of wireless channels. Similarly, we want to work with students who want to learn more about mobile systems privacy and criminal investigation.