From macglashan at tti-c.org Tue Dec 2 09:04:05 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Dec 2 09:03:19 2008 Subject: [TTIC Colloquium] U of C Guest Speaker: Kilian Pohl Message-ID: <8274BC0686444AF4BC11A212AAA4B6F4@jmacglDPLFYD1> When: Thursday, December 4 @ 1:30pm Where: Ryerson 277 Who: Kilian Pohl Title: Representing Objects via the Logarithm of Odds In this talk, we describe a new representation for capturing the uncertainty of objects in images based on the logarithm of odds. This representation addresses several problems in imaging as it provides an intrinsic, probabilistic representation for combining and deforming objects. We use this technology in order to solve the Mean Field approximation in the level set framework. Conventional likelihood models are combined with a curve length prior on boundaries, and an approximate posterior distribution on labels is sought via the mean field approach. Optimizing the resulting estimator by gradient descent leads to a level set style algorithm where the level set functions are the logarithm of odds encoding of the posterior label probabilities. Applications with more than two labels are easily accommodated.The label assignment is accomplished by the Maximum A Posteriori rule, so there are no problems of "overlap'' or "vacuum". Kilian Pohl received his doctorate in computer science from the Medical Vision Lab at MIT, was a junior faculty member at Harvard Medical School, and is currently a research staff member at IBM Almaden. His main research area is computational image analysis with an emphasis on studying statistical models from a Bayesian perspective. Kilian has been the recipient of several awards such as the Medical Image Analysis - MICCAI'06 Best Paper Prize for his work on logarithm of odds. For more details about his research, please visit his website at http://people.csail.mit.edu/pohl. Contact: Pedro Felzenszwalb pff@cs.uchicago.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081202/6150b134/attachment.htm From macglashan at tti-c.org Fri Dec 12 10:55:29 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Dec 12 10:54:05 2008 Subject: [TTIC Colloquium] TTI-C Colloquium: Jason Riggle, UoC References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <7B5C6F21CF6B475EAF200CA12A7F3B39@jmacglDPLFYD1> When: Monday, December 15 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Jason Riggle, University of Chicago Title: Complexity, Learnability, and Constraint-Based Grammars In this talk, I present results on the learnability of phonological grammars for two constraint-based models, Harmonic Grammar (HG; Legendre, Miyata, and Smolensky 1990) and Optimality Theory (OT; Prince and Smolensky 1993). I first establish that grammars in these models are learnable from reasonably sized samples of data and then present a learning algorithm for OT that is guaranteed to make no more than k log2 k mistakes when learning grammars with k constraints. I demonstrate that this mistake bound is within a logarithmic factor of the best possible mistake bound for any OT/HG learning algorithm. The proposed learning algorithm calculates the number of rankings that are consistent with a set of data. This makes possible a simple and effective Bayesian heuristic to guide learning - all else equal, choose candidates that are preferred by the highest number of rankings consistent with previous observations. This general strategy can be applied to OT, HG, or any parameterized model of grammar, and it associates with each language generated by the theory an abstract quantity, the p-volume, that measures the fraction of the parameter space corresponding to grammars that generate that language. The p-volume seems to encode 'restrictiveness' in a way similar to Tesar and Prince's (1999) r-measure. Preliminary investigations indicate that p-volume is significantly correlated with typological frequency (cf. Bane and Riggle 2008). This fact is neatly explained if language learners use a strategy that is sometimes called a Gibbs leaner wherein they keep track of the region of the parameter space consistent with previous observations but make guesses according to a single hypothesis grammar randomly selected from that region. Upon making an error the Gibbs leaner updates the parameter region and randomly selects a new hypothesis grammar from that region. Following this strategy, learners will be predisposed towards grammars with large p-volume in cases where the hypotheses are underdetermined by the data. Moreover, priors other than the 'flat' distribution over rankings can be included to implement models of ranking bias. One of the primary assets of this strategy is that it allows linguistic theory to be informed by the relative frequencies of patterns in linguistic typologies rather than only by the boolean distinction of whether or not a pattern is attested. Though some of the frequency asymmetries surely come from non-linguistic historical accidents, a model of learning that is able to account for some of the frequency variance is clearly of interest and makes a range of predictions that can be tested in experimental settings. Contact: Karen Livescu, TTI-C klivescu@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081212/b7e0df53/attachment.htm From macglashan at tti-c.org Mon Dec 15 15:52:53 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Dec 15 15:51:12 2008 Subject: [TTIC Colloquium] TTI-C Talk: Payman Yadollahpour, Brown University References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <5E90CFB3EFDF4A0FBF4694C563B00DA7@jmacglDPLFYD1> When: Thursday, December 18 @ 11:00am Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Payman Yadollahpour, Brown University Title: Subspace learning for optimal neural control of movement Neural decoding of motor control of hand and arm movements in primates requires developing statistical models that explain how the recorded neural population activity relates to motor behavior. Until recently, much of the work in this area has focused on learning linear models of decoding for low-dimensional motor control, such as 2D control of a computer cursor. Capturing a richer set of motor behaviors such as hand and arm posture during object grasping and manipulation tasks introduces much higher dimensional representations of motor control. In this context we focus on understanding the underlying degrees of freedom in complex kinematics that are explained by the neural activity. One way of learning these "effective" degrees of freedom has been to employ dimensionality reduction techniques, such as Principal Component Analysis (PCA), to find a linear kinematic subspace that accounts for the observed motor behavior, separate from the observed neural activity. The orthonormal bases that span this subspace are then considered as the underlying latent variables, or "motor primitives" that describe behavior. However, these motor primitives are not guaranteed to be optimally correlated with the observed neural activity. In this work we compare such a naive PCA based low-dimensional representation of motor behavior against an optimized linear subspace learned under an objective that explicitly includes optimization of decoding. Maximizing this objective function produces a linear subspace that explains the motor behavior with reasonable fidelity while increasing the correlation between the neural activity and the latent variables of the subspace. This is joint work with G. Shakhnarovich and M. Black. Contact: Greg Shakhnarovich, TTI-C greg@tti-c.org 834-2572 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081215/cf21cbf6/attachment-0001.htm