Talk:Audio Genre - Revision history

139.222.3.147: /* Database size */

2005-02-25T13:09:13Z

‎Database size

139.222.3.147: Database size

2005-02-25T13:08:52Z

Database size

Xiaohu at 22:15, 20 February 2005

2005-02-20T22:15:21Z

130.126.146.79 at 21:25, 20 February 2005

2005-02-20T21:25:53Z

New page

I like the idea of hierarchical genre categorization, and multiple labels. We should not be content with good performance on simple questions (since there are quite good systems on single genre classification problem.) I would like to see how "bad" the systems will be in harder problems.

Genre problems are of great importance to music retrieval, because queries on genre are quite popular. Progresses in genre classification will enhance approaches in real retrieval problems.

← Older revision		Revision as of 13:09, 25 February 2005
Line 7:		Line 7:
	I also want to see algorithms evaluated on hard problems. E.g. which algorithms work well on 100 examples, 1000 examples and 10,000 examples (scalability). This applies to classes as well, stick to 1000 examples but increase number of classes from 10 to 100 (essentially what will happen with artist id task). Some algorithms will fail or take a lot longer to run under these conditions.		I also want to see algorithms evaluated on hard problems. E.g. which algorithms work well on 100 examples, 1000 examples and 10,000 examples (scalability). This applies to classes as well, stick to 1000 examples but increase number of classes from 10 to 100 (essentially what will happen with artist id task). Some algorithms will fail or take a lot longer to run under these conditions.

−	~~Kriss~~	+	Kris

← Older revision		Revision as of 13:08, 25 February 2005
Line 2:		Line 2:

	I like the idea of hierarchical genre categorization, and multiple labels. We should not be content with good performance on simple questions. I would like to see how "bad" the systems will be in harder problems. However, to evaluate harder problems, a good set of data is the key. Besides copyright issues, we need people (probably music librarians) to annotate and organize a reliable corpus.		I like the idea of hierarchical genre categorization, and multiple labels. We should not be content with good performance on simple questions. I would like to see how "bad" the systems will be in harder problems. However, to evaluate harder problems, a good set of data is the key. Besides copyright issues, we need people (probably music librarians) to annotate and organize a reliable corpus.
		+
		+	== Database size ==
		+
		+	I also want to see algorithms evaluated on hard problems. E.g. which algorithms work well on 100 examples, 1000 examples and 10,000 examples (scalability). This applies to classes as well, stick to 1000 examples but increase number of classes from 10 to 100 (essentially what will happen with artist id task). Some algorithms will fail or take a lot longer to run under these conditions.
		+
		+	Kriss

← Older revision		Revision as of 22:15, 20 February 2005
Line 1:		Line 1:
−	~~I like the idea~~ of ~~hierarchical genre categorization~~, ~~and multiple labels. We should not be content with good performance~~ on ~~simple questions (since there~~ are quite ~~good systems on single~~ genre classification ~~problem.) I would like to see how "bad" the systems~~ will be in ~~harder~~ problems.	+	Genre problems are of great importance to music retrieval, because queries on genre are quite popular. Progresses in genre classification will enhance approaches in real retrieval problems.

−	~~Genre problems are~~ of ~~great importance to music retrieval~~, ~~because queries~~ on ~~genre are quite popular~~. ~~Progresses in genre classification~~ will ~~enhance approaches~~ in ~~real retrieval~~ problems.	+	I like the idea of hierarchical genre categorization, and multiple labels. We should not be content with good performance on simple questions. I would like to see how "bad" the systems will be in harder problems. However, to evaluate harder problems, a good set of data is the key. Besides copyright issues, we need people (probably music librarians) to annotate and organize a reliable corpus.