Machine Learning Meetings and Events

Group Meetings: Group meetings are held Mondays from 11am to Noon (talk starts 11:10am) in D.L. Pratt 290C unless otherwise noted. Meetings are coordinated by Hugo Larochelle.

Tea Talks: Tea talks are held every Wednesday at 4:00pm in D.L. Pratt 290C. Talks should be simple, accessible, and not exceed 15 minutes. Speakers bring snacks, make tea, and provide a copy of the presented paper.

Group Meeting Apr 20, 2009: Non-metric Neighbor Embedding

Speaker: Seunghak Lee
Abstract:
Non-metric multidimensional scaling (MDS) is a dimensionality reduction technique that seeks a low-dimensional embedding whose inter-point distances have the same rank order as the original dissimilarities. While metric MDS is a dissimilarity-preserving embedding, where neighbors relationships are of particular importance, non-metric MDS is an order-respecting embedding, and thus reconstructs global, rather than local features of the data. In this paper we present non-metric neighbor embedding (NNE). NNE preserves the local structure inherent in non-metric datasets by faithfully preserving neighbor relationships in low-dimensional Euclidean space. We find an embedding solution by optimizing an objective function that attracts neighbor data points while repelling non-neighbors. In our experiments, NNE has proven effective at handling missing distances (>95%) and preserving local structure in non-metric datasets. We show that our method outperforms non-metric MDS, Isomap, and Maximal Variance Unfolding in preserving neighbors, while the rank order of dissimilarities of neighbors is also retained. We apply NNE to the problem of reconstructing a 3-dimensional protein structures from contact maps, an important problem in structural bioinformatics, and show that it can faithfully reconstruct structures with no prior biological knowledge.