Information Systems Lab (ISL) Colloquium

Nonparametric generative modeling via optimal transport and diffusions with provable guarantees

Topic: 
Nonparametric generative modeling via optimal transport and diffusions with provable guarantees
Abstract / Description: 

By building upon the recent theory that established the connection between implicit generative modeling (IGM) and optimal transport, in this study, we propose a novel parameter-free algorithm for learning the underlying distributions of complicated datasets and sampling from them. The proposed algorithm is based on a functional optimization problem, which aims at finding a measure that is close to the data distribution as much as possible and also expressive enough for generative modeling purposes. We formulate the problem as a gradient flow in the space of probability measures. The connections between gradient flows and stochastic differential equations let us develop a computationally efficient algorithm for solving the optimization problem. We provide formal theoretical analysis where we prove finite-time error guarantees for the proposed algorithm. Our experimental results support our theory and show that our algorithm is able to successfully capture the structure of different types of data distributions.

The talk will be based on the following paper:
"Sliced-Wasserstein Flows: Nonparametric Generative Modeling via Optimal Transport and Diffusions", ICML 2019

Date and Time: 
Friday, November 8, 2019 - 1:00pm
Venue: 
Allen 101X

IT Forum & ISL Colloquium presents "The Case Against Reality"

Topic: 
The Case Against Reality
Abstract / Description: 

If I have a visual experience that I describe as a red tomato a meter away, then I am inclined to believe that there is, in fact, a red tomato a meter away, even if I close my eyes. I believe that my perceptions are, in the normal case, veridical—that they accurately depict aspects of the real world. But is my belief supported by our best science? In particular: Does evolution by natural selection favor veridical perceptions? Many scientists and philosophers claim that it does. But this claim, though plausible, has not been properly tested. In this talk I present a new theorem: Veridical perceptual systems are never more fit than non-veridical perceptual systems of equal complexity that are simply tuned to the relevant fitness functions. This entails that perception is not a window on reality; it is more like a windows interface on your laptop. Spacetime is your desktop, and physical objects, like apples and neurons, are simply icons on this desktop. I discuss this interface theory of perception and its implications for one of the most puzzling unsolved problems in science: the relationship between brain activity and conscious experiences.

Date and Time: 
Friday, November 8, 2019 - 1:15pm
Venue: 
Packard 202

ISL & IT Forum present " Dictionary Compression and its Applications"

Topic: 
Dictionary Compression and its Applications
Abstract / Description: 

Dictionary compression is a known technique, promising to solve the problem of compressing small inputs. However, it has been only available to implementers since relatively recently, as newer compression algorithms shipped dictionary builders alongside their main codec. Due to this recent timeframe, complexities around deploying this solution at larger scales only start to be appreciated. Understanding these difficulties, and finding ways to harness them, is key to target system's performance and reliability. Yet, the price is big, as dictionary compression not only improves compression, it also offers the potential to redesign systems around their capabilities.

We'll cover the benefits, trade-off and operational difficulties of dictionary compression, as well as their important second-order impacts for systems adopting it.

Date and Time: 
Friday, November 1, 2019 - 1:15pm
Venue: 
Packard 202

ISL Colloquium & IT Forum present "Generalized Resilience and Robust Statistics"

Topic: 
Generalized Resilience and Robust Statistics
Abstract / Description: 

Robust statistics traditionally focuses on outliers, or perturbations in total variation distance. However, a dataset could be corrupted in many other ways, such as systematic measurement errors and missing covariates. We generalize the robust statistics approach to consider perturbations under any Wasserstein distance, and show that robust estimation is possible whenever a distribution's population statistics are robust under a certain family of friendly perturbations. This generalizes a property called resilience previously employed in the special case of mean estimation with outliers. We justify the generalized resilience property by showing that it holds under moment or hypercontractive conditions. Even in the total variation case, these subsume conditions in the literature for mean estimation, regression, and covariance estimation; the resulting analysis simplifies and sometimes improves these known results in both population limit and finite-sample rate. Our robust estimators are based on minimum distance (MD) functionals (Donoho and Liu, 1988), which project onto a set of distributions under a discrepancy related to the perturbation. We present two approaches for designing MD estimators with good finite- sample rates: weakening the discrepancy and expanding the set of distributions. We also present connections to Gao et al. (2019)'s recent analysis of generative adversarial networks for robust estimation.

Joint work with Banghua Zhu and Jacob Steinhardt

Date and Time: 
Friday, October 11, 2019 - 1:15pm
Venue: 
Packard 202

ISL Colloquium welcomes Fanny Yang

Topic: 
TBA
Abstract / Description: 

TBA


The Information Systems Laboratory Colloquium (ISLC)

is typically held in Packard 101 every Thursday at 4:30 pm during the academic year. Coffee and refreshments are served at 4pm in the second floor kitchen of Packard Bldg.

The Colloquium is organized by graduate students Joachim Neu, Tavor Baharav and Kabir Chandrasekher. To suggest speakers, please contact any of the students.

To receive email notifications of seminars you can join the ISL mailing list.

Date and Time: 
Thursday, December 5, 2019 - 4:30pm
Venue: 
Packard 101

ISL Colloquium welcomes Christopher Metzler

Topic: 
TBA
Abstract / Description: 

TBA


The Information Systems Laboratory Colloquium (ISLC)

is typically held in Packard 101 every Thursday at 4:30 pm during the academic year. Coffee and refreshments are served at 4pm in the second floor kitchen of Packard Bldg.

The Colloquium is organized by graduate students Joachim Neu, Tavor Baharav and Kabir Chandrasekher. To suggest speakers, please contact any of the students.

To receive email notifications of seminars you can join the ISL mailing list.

Date and Time: 
Thursday, November 21, 2019 - 4:30pm
Venue: 
Packard 101

IT-Forum and ISL Colloquium SPECIAL SEMINAR: "Towards an Average-case Complexity of High-dimensional Statistics"

Topic: 
Towards an Average-case Complexity of High-dimensional Statistics
Abstract / Description: 

The prototypical high-dimensional statistical estimation problem
entails finding a structured signal in noise. These problems have
traditionally been studied in isolation, with researchers aiming to
develop statistically and computationally efficient algorithms, as well
as to try to understand the fundamental limits governing the interplay
between statistical and computational cost. In this talk I will
describe a line of work that yields average-case reductions directly
between a number of central high-dimensional statistics problems,
relating two problems by transforming one into the other. It turns out
that several problems described by robust formulations can be addressed
by one set of techniques, and we will focus on these in the talk. In
this direction, we obtain the following average-case lower bounds based
on the planted clique conjecture: a statistical-computational gap in
robust sparse mean estimation, a detection-recovery gap in community
detection, and a universality principle for computational-statistical
gaps in sparse mixture estimation. In addition to showing strong
computational lower bounds tight against what is achievable by
efficient algorithms, the methodology gives insight into the common
features shared by different high-dimensional statistics problems with
similar computational behavior. Joint work with Matthew Brennan.


Special joint seminar:  IT-Forum and The Information Systems Laboratory Colloquium (ISLC)

Date and Time: 
Friday, November 15, 2019 - 1:15pm
Venue: 
Packard 202

ISL Colloquium presents "Spectral graph matching and regularized quadratic relaxations"

Topic: 
Spectral graph matching and regularized quadratic relaxations
Abstract / Description: 

Given two unlabeled, edge-correlated graphs on the same set of vertices, we study the "graph matching" problem of identifying the unknown mapping from vertices of the first graph to those of the second. This amounts to solving a computationally intractable quadratic assignment problem. We propose a new spectral method, which computes the eigendecomposition of the two graph adjacency matrices and returns a matching based on the pairwise alignments between all eigenvectors of the first graph with all eigenvectors of the second. Each alignment is inversely weighted by the distance between the corresponding eigenvalues. This spectral method can be equivalently viewed as solving a regularized quadratic programming relaxation of the quadratic assignment problem. We show that for a correlated Erdos-Renyi model, this method can return the exact matching with high probability if the two graphs differ by at most a 1/polylog(n) fraction of edges, both for dense graphs and for sparse graphs with at least polylog(n) average degree. Our analysis exploits local laws for the resolvents of sparse Wigner matrices. Based on joint work with Zhou Fan, Cheng Mao, Yihong Wu, all at Yale.


The Information Systems Laboratory Colloquium (ISLC)

is typically held in Packard 101 every Thursday at 4:30 pm during the academic year. Coffee and refreshments are served at 4pm in the second floor kitchen of Packard Bldg.

The Colloquium is organized by graduate students Joachim Neu, Tavor Baharav and Kabir Chandrasekher. To suggest speakers, please contact any of the students.

To receive email notifications of seminars you can join the ISL mailing list.

Date and Time: 
Thursday, November 14, 2019 - 4:30pm
Venue: 
Packard 101

ISL Colloquium Presents "Beyond Supervised Learning for Biomedical Imaging"

Topic: 
Beyond Supervised Learning for Biomedical Imaging
Abstract / Description: 

Today, many biomedical imaging tasks, such as 3D reconstruction, denoising, detection, registration, and segmentation, are solved with machine learning techniques. In this talk, I will present a flexible learning-based framework that has allowed us to derive efficient solutions for a variety of such problems, without relying on heavy supervision. I will primarily employ image registration as a concrete application and present the details of VoxelMorph, our unsupervised learning-based image registration tool. I will show empirical results obtained by co-registering thousands of brain MRI scans where VoxelMorph has yielded state-of-the-art accuracy with runtimes that are orders of magnitude faster than conventional tools. Finally, I will present some recent results where we used VoxelMorph to learn conditional deformable templates that can reveal population variation as a function of factors of interest, such as aging or genetics. Our code is freely available at https://github.com/voxelmorph/voxelmorph.


The Information Systems Laboratory Colloquium (ISLC)

is typically held in Packard 101 every Thursday at 4:30 pm during the academic year. Coffee and refreshments are served at 4pm in the second floor kitchen of Packard Bldg.

The Colloquium is organized by graduate students Joachim Neu, Tavor Baharav and Kabir Chandrasekher. To suggest speakers, please contact any of the students.

To receive email notifications of seminars you can join the ISL mailing list.

Date and Time: 
Thursday, November 7, 2019 - 4:30pm
Venue: 
Packard 101

ISL Colloquium presents Iterative Collaborative Filtering for Sparse Noisy Tensor Estimation

Topic: 
Iterative Collaborative Filtering for Sparse Noisy Tensor Estimation
Abstract / Description: 

We present a generalization of the collaborative filtering algorithm for the task of tensor estimation, i.e. estimating a low-rank 3-order n-by-n-by-n tensor from noisy observations of randomly chosen entries in the sparse regime. Not only does the algorithm have desirable computational properties, it also provably achieves sample complexity that (nearly) matches the conjectured lower bound on the sample complexity. Furthermore, our analysis results in high probability bounds on the infinity norm of the error, as opposed to the weaker MSE bounds achieved by previous approaches. Our proposed algorithm uses the matrix obtained from the ''flattened'' tensor to compute similarity with respect to a corresponding observation graph. The algorithm recovers the tensor with max entrywise error decaying to 0 with high probability as long as the entries are sampled uniformly at a density of Omega(n^{-3/2 + epsilon}) for any arbitrarily small epsilon > 0. This sample complexity threshold (nearly) matches a conjectured lower bound as well as the ''connectivity threshold'' of the corresponding observation graph used in our algorithm, providing a different angle to explain the conjectured lower bound.


The Information Systems Laboratory Colloquium (ISLC)

is typically held in Packard 101 every Thursday at 4:30 pm during the academic year. Coffee and refreshments are served at 4pm in the second floor kitchen of Packard Bldg.

The Colloquium is organized by graduate students Joachim Neu, Tavor Baharav and Kabir Chandrasekher. To suggest speakers, please contact any of the students.

To receive email notifications of seminars you can join the ISL mailing list.

Date and Time: 
Thursday, October 31, 2019 - 4:30pm
Venue: 
Packard 101

Pages

Subscribe to RSS - Information Systems Lab (ISL) Colloquium