Prof. Lorenzo Torresani

The Visual Learning Group researches methods to acquire models of the real-world from visual data, such as images, videos, and motion sequences. Most of this research lies at the boundary between computer vision, machine learning and computer animation. Projects include learning image representations for visual recognition - the problem of determining what is where in a picture - but also extracting models of human movement from video and motion capture data.

Read more...

Prof. Chris Bailey-Kellogg

Family reunions can be very interesting. Proteins have relatives (across organisms and even within the same organism) that are similar, but also different in significant ways. For example, shown in the figure are one serine protease (blue; function is to chop up other proteins) and one of its inhibitors (red; function is to block the chopping mechanism, as shown). Different proteases recognize different places to chop, while different inhibitors have different degrees of inhibition for different proteases. We are developing techniques to learn from nature's exploration of these families -- generalizing common features of observed family members and characterizing their differences, and relating these to experimental observations about their functions. This enables us to optimize our own new variants with desired functions (e.g., particularly strong and protease-specific inhibitors).

Read more...

Prof. Andrew Campbell

How do we make smartphones even smarter?

That is the central question driving the Smartphone Sensing Group at Dartmouth.

Smartphones are open and programmable and come with a growing number of powerful embedded sensors, such as an accelerometer, digital compass, gyroscope, GPS, microphone, and camera, which are enabling new sensing applications across a wide variety of domains such as social networks, mobile health, gaming, entertainment, education and transportation.

Application delivery channels such as the AppStore and Market have transformed plain old cell phones into app phones, capable of downloading a myriad of applications in an instant.

The Smartphone Sensing Group is turning the everyday smart phone into a cognitive phone by pushing intelligence to the phone and the computing cloud to make inferences about people's behavior, surroundings and their life patterns.

We are developing new software technology for smartphones to sense, learn, visualize, and share information about ourselves, friends, communities, the way we live, and the world we live in.

Some of the sensing algorithms, systems and applications we have developed in collaboration with Tanzeem Choudhury (Cornell University) and others include CenceMe, SoundSense, NeuroPhone, Jigsaw, Darwin Phones, NextPlace, EyePhone, BeWell, Community-Guided Learning , and Community Similarity Networks.

Read more...

Prof. Amit Chakrabarti

Nearest neighbour searching is one of those basic and fascinating theoretical problems in computer science that has a host of applications in problems from very diverse fields. The problem is widely believed to be intractable in high dimensions, but good approximation algorithms were discovered in the late 1990s.

Our research considers the approximate nearest neighbour search (ANNS) problem on the d-dimensional Hamming cube. Building on techniques of information complexity and communication complexity from an earlier research project of ours, we prove the first unrestricted lower bound for this problem. We also improve previous algorithms for this problem, ending up with matching upper and lower bounds on the number of randomised queries required to solve ANNS.

Shown in the figure is a schematic of the reduction bridging communication complexity and ANNS, a key step in our proof.

Read more...

Prof. Tom Cormen

High-performance computing with disk-resident data? Although that might seem like an oxymoron, we are developing a software environment, FG, to enable it.

FG is for asynchronous programs that run on clusters and fit into a pipeline framework. Each pipeline stage corresponds to a function that operates on a buffer. Multiple buffers traverse the pipeline and correspond to blocks in the memory hiearchy. Stages run asynchronously (via threads) in order to make it easy to overlap their operations (computation, communication, and I/O).

Using FG, we have developed programs that can sort well in excess of 100 gigbytes of data. (Image, left, shows Tom's whiteboard covered with deep thoughts about pipelines.)

The TISH project's multidisciplinary research will drive innovation in information-sharing technology that ensures security and privacy while addressing the pragmatic needs of patients, clinical staff, and healthcare organizations to deliver efficient, high-quality care. This multidisciplinary team of investigators will address fundamental challenges in current and emerging areas of information security, as identified by its healthcare partners, and will focus on four research "threads".

• Access control in clinical settings, led by Sean Smith (Computer Science)
• Mobile healthcare (mHealth), led by David Kotz and Tanzeem Choudhury (Computer Science)
• Economic and risk models, led by Eric Johnson (Tuck School of Business)
• Social informatics of IT in healthcare organizations, led by Denise Anthony (Sociology) and Ann Flood (The Dartmouth Institute of Health Policy and Clinical Practice)
•  
• See more about this project and other healthcare-IT projects at the ISTS website.
•  

Prof. Lorie Loeb

Information visualization, real-time energy data retrieval and analysis, animation, social and behavioral sciences, environmental science, statistical analysis, game technology and computation all come together to encourage resource conservation at Dartmouth and beyond. We ask the fundamental question: how can complex information about energy use be displayed in ways that encourage people to change their behavior? Green Lite Dartmouth visualizes complex, real-time energy data in simple, meaningful and dynamic ways. When electric use (plug load and lighting), for example, is low, the polar bear in the display is happy and playful. As electric use goes up, the polar bear becomes more distressed and his well-being is endangered.

Prof. Sean Smith

A trusted third party (TTP) can make it easier to solve multi-party security problems. However, using a TTP in a design has been akin to invoking magic or fairies: tools not possible in the real world.

However, in the last decade, techniques have emerged that can provide foundations for TTPs based in hardware. In our lab, we are exploring ways of building effective trustworthy platforms founded on these techniques and using these platforms for hardware-based TTPs in real-world applications.

Read more...

Prof. Peter Winkler

The basic problem in the Theory of Computing is do determine what kinds of problems can be efficiently solved by computer; in other words, what can algorithms do or not do.

In the 70's, theorists uncovered a wide range of "hard" decision and computation problems which are all intractable—except in the (unlikely) event that they were all tractable! Since then, exciting new techniques involving probabilistically checkable proofs have established that many quantities are as hard to approximate as they are to compute exactly.

Ironically, however, at about the same time theorists discovered a powerful mathematical tool which has greatly expanded the range of quantities which we can approximate efficiently. Included are volumes of high-dimensional objects, partition functions in statistical physics, the permanent of a matrix, the number of linear extensions of an ordered set, generation of contingency tables (used by actuaries to compute insurance rates) and many more. The method, called MCMC (Markov chain Monte Carlo), entails randomly sampling the objects being counted by starting with a non-random object and moving to a random neighboring object.

The mathematical crux of this method is proving "rapid mixing", i.e., showing that the process reaches a random object in sufficiently short time. Practically every branch of mathematics has contributed to the delightful variety of ways that have been devised to show rapid mixing in its many guises.

Shown in the picture is a configuration of the Widom-Rowlinson (WR) model of statistical physics, generated by MCMC. The WR model simulates two gases (here, with red and green molecules), under the constraint that differing molecules may not occupy adjacent points on a grid.

Read more...

Prof. Sean Smith

In a distributed world, how can a stakeholder verify the identity (or other relevant properties) of a remote party?

In the emerging information infrastructure, multiple parties spanning multiple organizations have various opinions about what they might trust, in what contexts. To be effective, the underlying mechanics for transmitting and expressing assertions about these parties needs to be able to provide the right parameters to accommodate this multiplicity of views.

Because it can enable secure communications between parties who do not share a secret beforehand, public key cryptography is a natural and perhaps unique building block. Achieving this vision in the real world, however, requires public key infrastructure (PKI). In our lab, we are exploring a number avenues to make PKI effective in the real world.

Read more...