Research-old

	Energy-Efficient Algorithms Erik Demaine	The new field of energy-efficient algorithms aims to develop new techniques for solving computational problems with vastly reduced energy consumption—for some problems, by several orders of magnitude—in exchange for a small increase in time and memory requirements. Specifically, we explore how to algorithmically exploit reversible computation, an idea that has been around since the 1970s and has just started to become a practical reality in the latest AMD chips, but for which we have only just begun understanding how to design efficient algorithms. Our preliminary investigations indicate
	Uncovering Clinically Relevant Medical Knowledge John Guttag	The day-to-day practice of medicine is based largely on a combination of the personal experience of those making the decisions and non-patient-specific information derived by applying conventional statistical methods to large clinical trials. With the boom in the collection of clinical information in computationally accessible formats, it is now possible to use advanced machine learning and data mining techniques to put clinical decision making on a sounder more patient-specific basis. That is the mission of CSAIL's Data -driven Medical Research Group. Current projects include risk strat
	Big Data for a Better Life: The Trento Smart City Project Alex Pentland	We have deployed a data sensing and data sharing architecture in the city of Trento in order to `mashup' government, company, and individual mobile data. The goal is to validate the value, monitization, and privacy/ownership issues of Big Data in running a `smart city.' Joint with Telefonica, Telecom Italia, Government of Trento, European Inst. Technology.
	Growing Big Linked Data From Seed: Building a Demo Tim Berners-Lee	The goal of Linked Data is to replace traditional app-data silos with a universal integration platform to provide globally contextualized information using global identifiers, authentication, authorization, storage, and privacy. The architecture separates application from data giving users control of the data and where it’s stored, independent of the choice of application.
	Vision Machine: Learning Online from 25 Million Images Bill Freeman	We want to teach machines to see. If we could recognize objects and locations, this would have large impact on robotics, assistive care, and public safety, to name just a few areas. Presently, machine vision systems can recognize a small number of object categories, and can localize objects within an image on moderately well. The next big task in computer vision is to scale-up object recognition: to reliably detect thousands of object classes. An important component of any long term solution to the vision problem is an online, unsupervised training from a massive dataset of images.
	Natural Language Interface for Big Data Boris Katz	As we develop storage and compute platforms for scaling to big data and practical algorithms for efficiently processing it, we will need to create new ways to access and interact with massive scale data. A comprehensive solution to the problem of dealing with large amounts of Web and sensor data involves not only analysis strategies, but also access strategies. It is entirely possible that for a given, large dataset, there will be hundreds if not thousands of distinct types of queries that may be applied to the data.
	Cloud-scale, Flexible Scaling and Factoring of Machine Learning Algorithms: The FlexGP Project Una May O'Reilly	Scalable machine learning systems will underpin the complex data analytics afforded by BigData. The goal of the FlexGP project is scaled machine learning which exploits cloud-scale parallelism and resource elasticity. FlexGP launches diverse learning engines onto the cloud which are different along training data partitions and explanatory dimensions as well as model expression and objective function choices.
	Sublinear time algorithms Ronitt Rubinfeld	The goal of this project is to develop powerful algorithmic sampling techniques which allow one to estimate parameters of the data by viewing only a miniscule portion of it. Such parameters may be combinatorial, such as whether a large network has the "six degrees of separation property", algebraic, such as whether the data is well-approximated by a linear function, or even distributional, such as whether the data comes from a distribution over a large number of distinct elements.
	Execution Migration Machine (EM2) Srini Devadas	Big Data needs Big Processors, and Big Processors need Big Caches. Increasingly, however, power and thermal considerations dictate that many small processors and many small caches supplant the paradigm of few big processors and caches. The Execution Migration Machine (EM²) project aims to find the best way of using these resources.
Christine Daniloff, MIT News Office	Maintaining Coherent Memory in Dynamic Distributed Systems Nancy Lynch	Modern data services manage data by storing it on distributed servers. The data should appear to its users to be consistent, as if it were maintained on a single centralized server, even though it is actually distributed and replicated. It should be efficient to access, and available in the face of unpredictable failures and other network changes. We design and analyze algorithms to implement such data services.