ISMIR 2009 Day 1

Tutorial MIR at the Scale of the Web

On the first day of the 10th International Society for Music Information Retrieval Conference (ISMIR) Malcolm Slaney and I gave a tutorial on music information retrieval methods that scale to millions, or even billions of items. Topics covered included random hashing, reliable tagging, machine learning methods, hybrid machine-learning and hashing algorithms, and software systems for scaling up MIR research and applications.

Software we discussed:
AudioDB Code (Feature Indexing via Locality Sensitive Hashing)
Hadoop Code
Vowpai Rabbit Algorithm and Code

ISMIR 2009 Day 2

KEYNOTE TALK

9AM Great talk by Don Byrd, Tim Crawford, J. Stephen Downie (presented by latter two) on history of last 10 years of the International Society for Music Information Retrieval.

Conclusions:

MIR field should ensure diverse participation from other fields

MIR needs to change its focus to explore complete systems

TUTORIAL FALLOUT
10:30AM Malcolm Slaney and I gave a tutorial on large data methods on Day 1 (Monday).

Today Geoffroy Peeters asked me some very good questions about Locality Sensitive Hashing versus M-Tree approaches.

1. Can LSH be used on non-metric feature spaces such as Kullback-Leibler. Geoffroy’s point was that KL is not a metric because it doesn’t obey triangle inequality (even symmetric KL).

A. it is possible to embed KL into simple metric space with an L1 norm. LSH can be used with metric spaces with a simple norm, e.g. L1 and L2. But how does this deal with triangle inequality ?

2. M-Trees can have their dimensions re-scaled, weighted, and will still work. LSH cannot have the dimensions re-scaled.

A. This seems like an advantage for M-Trees. LSH depends critically on the scaling of the dimensions. Rescaling dimensions requires re-indexing the data set.