Current Research





Below are brief descriptions of some of the projects I have been involved and interested in during my PhD at Caltech. I have worked on some of these projects more extensively than others. In general my work has focused on using different machine learning algorithms to effectively learn representation of sensory data, mostly visual sensory data.


Generative vs. Discriminative Learning of Object Categories





From CVPR2005

How can we learn to distinguish between visually similar object categories? A good model must encompass enough information to accurately describe the variability within a category while simultaneously having enough discriminative power to discriminate between different categories. Generative classifiers model the joint distribution of the input and class while discriminative classifiers model the conditional distribution of the class given then data. We explore discriminative methods in the form of (1) Conditional Likelihood, and (2) Fisher Kernels and compared them to their generative counterparts. We find that discriminative methods are often useful in creating distinct representations for very similar object categories. In addition, we found that Fisher Kernels in conjunction with generative Constellation Models (hybrid generative-discriminative models) can provide a powerful framework which allows one to conveniently combine multiple object models and transfer knowledge between object classes. Joint work with Max Welling (UC Irvine). (References: CVPR2005, ICCV2005, NIPS2006, IJCV in review).



Creating a Perceptual Facial Similarity Space


Consider the task of finding similar-looking faces. In general it is unclear what features humans use to make perceptual judgments of facial similarity. We proceed by obtaining perceptual similarity information from subjects and using this information to generate a functional mapping from visual feature space to a perceptual space. We can use this mapping to better understand what features subjects use in order to assess facial similarity. (References: VSS2005).



Automatically Learning to Detect Lung Nodules in Thoracic Images




Thoracic CT scan

Lung cancer is the primary cancer killer in the United States with victims rarely displaying symptoms until later stages where the cancer is untreatable. Recent reports (see 1,2 for popular press coverage) have shown the potential benefits of early screening and detection. The deluge of data which would result from early screening systems necessitates improvements in automatic malignant nodule detections systems. Most Computer Aided Detection (CAD) systems rely on hand-crafted models to detect nodules. We take a different approach by designing machine learning algorithms which learn the statistics of malignant nodules automatically from a training set of labeled data. Joint work with Pierre Moreels and Andre Michelin. (Project Website)



Active Learning of Object Categories








Consider an image search in Google which would allow users to refine their initial search based on image content. Which images should the user label in order to achieve the best subsequent results? Active learning decides which data-points are most useful to have labeled in order to obtain the best future performance. We provide a general framework for active learning based on entropy and apply it to computer vision. (Submitted). Joint work with Greg Griffin.



Multi-Class Discrimination






Discriminating between many object classes simultaneously is an important milestone for visual object recognition. We are interested in multi-class boosting, hierarchical, and other methods which provide efficient methods for discriminating between multiple object categories. One data-set of interest is The Caltech 256 Object Category Data-Set. Greg Griffin has designed the website for this data-set and there is an accompanying technical note. (References: Technote).


Modeling Olfactory Processing in the Locust Antennal Lobe


Figure From Laurent Lab website, available here.

Together with Gilles Laurent, Professor of Biology at Caltech, we created a recurrent binary neural network model which mimics aspects of early olfactory processing in the Locust Antennal Lobe. In particular we model how the chaotic dynamics of the recurrent system can be used to discriminate between odors by using the temporal domain. In particular, the system is useful for discriminating between odors which activate similar sets of olfactory receptors. (References: NIPS 2003).