I… yes ok. If we’re just making that narrow of a claim, I think I might be able to get on board.

I still have this nagging feeling though.. think about it this way: What is the different between (a) choosing the right input to your learning algorithm, and (b) choosing the right constraint on that that in the learning algorithm?

Everyone else seems to be expressing a preference for (a). But I think that they’re equivalent. And have different advantages and disadvantages.

I think I’ll blog about it tomorrow, using a music information retrieval example. Will let you know when it’s up.

In particular, I read his example of PageRank as arguing that Google’s improvement over the dominant approaches it displaced as being its introducing the use of links, rather than its simply assigning more weight to them relative to other factors.

Another way of looking at this is that it’s often more important to pick a good objective function (and particularly the right inputs) than to choose the best algorithm for optimizing relative to that objective function.

Just to clarify: We’re not talking about good feature selection, are we? (i.e. just picking the right attributes out of whatever data we currently have.)

And we’re also not talking about intelligently-initialized machine learning algorithms (i.e. what some might call “structured” learning, or initializing your learning algorithm with domain-dependent knowledge so as to guide the learning algorithm into the best task-specific models.)

Instead, we’re talking about simply identifying new sources of data? Is this correct?

Do you believe it?

Personally, I think statements like this are true, for certain limited domains or application types. Home page finding, factoid lookup. Those are areas where large data is very useful in web search.

But when your information need is exploratory, I have a difficult time seeing how large data will help. By definition, you have an information need or task that is orthogonal to the direction that the large data is pointing. Exploratory search runs against the large data grain, not with it.