Personal tools
You are here: Home Events ANC Workshop Talks: Jyri Kivinen and Peter Orchard, Chair: David Reichert

ANC Workshop Talks: Jyri Kivinen and Peter Orchard, Chair: David Reichert

What
When Nov 15, 2011
from 11:00 AM to 12:00 PM
Where IF 4.31/33
Add event to calendar vCal
iCal

PETER ORCHARD:

Sparse structure in high-dimensional financial data

Finding structure in data is a fundamental problem in machine learning that has applications across multiple fields.

Standard approaches are often computationally expensive, and may not handle latent variables or high-dimensional data. In this talk, I will describe a method for learning a sparse, high-dimensional latent variable model, and illustrate its potential in financial stress-testing and visualising dependencies between asset prices.

JYRI KIVINEN:

Statistical Modelling of Natural Images: A Structured Approach

Natural images are effortlessly analyzed and parsed into semantically high-level descriptions by most humans. Producing such descriptions in image analysis systems is often complicated by the large and complex variability exhibited in the appearance of these high-dimensional signals, obtained via a noisy imaging process.

In this talk, I will discuss statistical modeling of such data using Boltzmann machines. Such Markov random field models with hidden units have shown significant promise for various unsupervised learning problems, including as models for statistical structure occurring in natural images.

We will begin by identifying critical issues with the current approaches, and argue that they are over-optimistic in terms of their goals. One evidence of this is the inability of the current generative models to produce image samples containing textured regions, a necessary subcomponent of any credible model for visual scenes. It is difficult to come up with quantitative assessment methods for unnormalized generative models; this can lead to rather distant proxies being used, such as discriminative performance of a classifier based on the generative model inferences.

Motivated by these observations we take a step back, and ask whether current methods are able to model even the more structured class of visual textures, and dissect them in the generative tasks of texture synthesis and inpainting, for which effective quantitative assessment methods are available. Based on the findings, we will also consider structured extensions to model more complicated visual data, starting by Boltzmann machines capable of generating multiple textures, and demonstrate state-of-the-art performance with them in texture modelling.