April 2010 - talks - lists.cs.princeton.edu

Colloquium Speaker, Mon April 12- Jake Abernethy
by Nicole E. Wagenblast 06 Apr '10

06 Apr '10

Learning, Adversaries, and Limited Feedback Jake Abernethy, UC Berkeley Monday, April 12, 2010 4:30pm Small Auditorium, CS 105 A basic assumption often made in Machine Learning is that the data we observe are independent and identically distributed. But is this really necessary? Is it even realistic in typical scenarios? A number of recent results provide guarantees for arbitrary (or even adversarially-generated) data sequences. It turns out, for a wide class of problems, the learner's performance is no worse when the data is assumed to be arbitrary as opposed to i.i.d. This result involves a nice application of the Minimax theorem, which I'll briefly describe. I'll also delve into a more challenging problem: the "bandit" setting, in which the learner receives only very little feedback on each example. It was unknown for some time whether there exists an efficient algorithm that achieves the same guarantee for non-i.i.d. data for a general class of learning/decision problems. I'll present the first known such bandit algorithm, and I'll sketch the central ideas behind the technique, borrowing several tricks from the Interior Point optimization literature.

1 0

2pm Fri Apr 9 talk on Content-Centric Opportunistic Communication for Mobile Phones
by Jennifer Rexford 05 Apr '10

05 Apr '10

TALK ANNOUNCEMENT Speaker: Erik Nordstrom, Princeton University Title: Haggle: Content-Centric Opportunistic Communication for Mobile Phones Date/time: 2:00-3:00pm on Friday April 9 Location: room 402 in the Computer Science building Abstract: Today, mobile phones are increasingly used for content-centric communication, whereby users access and share content with each other (e.g., pictures on Facebook). However, content sharing is limited to infrastructure networks and legacy Internet technologies are ill suited for content-centric communication. Moreover, the Internet protocols and APIs do not work well in highly mobile and disruptive environments, where intermittent connectivity is the norm rather than the exception. These deficiencies make it is cumbersome, or even impossible, to share content directly between mobile phones, leading to high stress on existing infrastructure and high costs for the consumer, while putting limitations on the possibilities for developing new and innovative applications. In this talk, I will present Haggle, a novel network architecture built from the ground up to support opportunistic content-centric communication directly between mobile phones. Haggle addresses content rather than hosts, which makes it irrelevant from where a user receives a content item. The content-centric communication is driven by a novel distributed search-based resolution scheme, which draws inspiration from online search engines. The search scheme resolves the mappings between the content and its interest group, i.e., the nodes that have an interest in the content. Search has the benefit of ranking, and therefore provides a natural limitation and ordering mechanism for content dissemination. This means that content, in general, has a higher likelihood of being disseminated if there is a high interest in it in the network. After the talk, I will demonstrate Haggle using a visualization tool that shows how photos taken with a phone's camera are shared with other mobile phones or devices. See http://www.haggleproject.org/ for more details.

1 0

Colloquium Speaker, Tues April 6- Rebecca Fiebrink
by Nicole E. Wagenblast 02 Apr '10

02 Apr '10

Real-time Human-Computer Interaction with Supervised Learning Algorithms for Music Composition and Performance Rebecca Fiebrink, Princeton University, Computer Science Department Tuesday, April 6, 2010- 4:30pm Small Auditorium CS105 Supervised learning offers a useful set of algorithmic tools for many problems in computer music composition and performance. Through the use of training examples, these algorithms offer human musicians a means to implicitly specify the relationship between low-level, human-generated control signals (such as gesturally-manipulated sensor outputs or audio captured by a microphone) and the desired computer response (such as a change in synthesis or structural parameters of dynamically-generated audio). In my work, I explore how to most effectively enable users to interact with supervised learning algorithms to compose and perform new music. I have built a general-purpose software system for applying standard supervised learning algorithms in real-time problem domains. This system, called the Wekinator, supports human interaction throughout the entire supervised learning process, including the generation of training examples and the application of trained models to real-time inputs. Already, the Wekinator has enabled the creation of several new compositions and instruments. Furthermore, this system has enabled me to study several aspects of human-computer interaction with supervised learning in computer music. I have used the Wekinator as a foundation for a participatory design process with practicing composers, ongoing work with non-expert users in a classroom context, and the design of a gesture recognition system for a sensor-augmented cello bow. This research has led to a clearer characterization of the requirements and goals of instrument builders and composers, a better understanding of how to design user interfaces for supervised learning in both real-time and creative application domains, and a greater insight into the roles that interaction (encompassing both human-computer control and computer-human feedback) can play in the development of systems containing supervised learning components. This work highlights how music and other creative endeavors differ from more traditional applications of supervised learning, and it contributes to a broader HCI perspective on machine learning practice.

1 1

C Richards FPO
by Melissa Lawson 01 Apr '10

01 Apr '10

Chris Richards will present his FPO on Thursday April 8 at 11AM in Room 402. The members of his committee are: Andrew Appel, advisor; David Walker and Paul-Andre Mellies (CNRS), readers; David August and Sanjeev Arora, nonreaders. A copy of his thesis, "The Approximation Modality in Models of Higher-Order Types" is on display in Room 310. Everyone is invited to attend his talk.

1 0