Personal tools
You are here: Home Events Workshop: Colin Mclean and Jim Bednar

Workshop: Colin Mclean and Jim Bednar

— filed under:

What
  • ANC Workshop Talk
When Mar 19, 2013
from 11:00 AM to 12:00 PM
Where IF 4.31/4.33
Add event to calendar vCal
iCal

Colin Mclean

Using Probabilistic Clustering to help Identify Multiple Protein Instances in Static PPI Datasets

Functioning of synapses is governed by macromolecular complexes held together by protein-protein, protein-lipid and lipid-lipid interactions. In the Post-Synaptic Density (PSD) multi-protein complexes contain multiple core proteins, such as the scaffolding proteins: PSD-95, SHANK, SAPAP. Commonly used clustering techniques fail to capture the role of these important proteins, by placing each protein in only a single community. Probabilistic clustering techniques circumnavigate this problem by assigning a probability to each protein belonging to a community, allowing overlapping structure of the data. We have implemented a principled statistical approach, based on clustering edges between nodes, for finding overlapping communities in our static Protein-Protein Interaction (PPI) networks. 
We extend this implementation by including a natural partitioning density function, to search for the optimal community structure in the data. We apply this clustering technique to the previously studied MASC complex and investigate which proteins belong to overlapping communities, and if this can be related back to isolate those core proteins playing a role in multiple complexes.

 


Jim Bednar

Categorizing visual scenes using edge co-occurences

Humans can rapidly make a category judgment about a visual scene, e.g. whether it contains an animal.  Previous explanations (e.g. Serre et al. 2007) use a series of processing steps in a multi-level hierarchy, successively interpreting the scene at different levels of abstraction, from contour extraction to low-level object recognition and finally to object categorization.  We explore here an alternative hypothesis that second-order statistics of edges alone are sufficient to perform a rough yet robust (translation-, scale-, and
rotation-invariant) scene categorization.

We first build a sparse representation of a natural image as a set of oriented edges, using an extension of the method from Geisler et al. (2001).  We then compute the second-order statistics of edge co-occurrences in a given dataset, giving an estimate of the ``association field'' for those images.  Finally, the second-order statistics are fed to a naive classifier, which gives image classification performance (natural vs. man-made, or animal vs. non-animal) similar to that of hierarchical models. 

These results suggest that image categorization could occur at relatively low levels of the visual system, prior to object recognition processes, perhaps by utilizing lateral connections in the primary visual cortex.