Sunday, April 1, 2007

Why Google's AI Vision Is Wrong (Not)

Why Google's AI Vision Is Wrong:
The central theme of Google's AI is that massive scale, vast data sets and planet-sized computers will, eventually - almost naturally - result in AI. This is a weak 'vision'. The reason it upsets me is that driving for scale of this type sidesteps the fundamental power to generalize. [...] To be honest, when I talk about AI, I really mean: systems that exhibit human-like intelligence (which could be far more powerful in some dimension than a human, but ultimately with a capacity to reason, conjecture, plan and execute). AI, as used by Eric Schmidt, clearly means something more like: a useful tool.
The obvious question to ask Matt is "How do you know?" Where does the power to generalize come from? The hypothesis that it is based on a very large associative memory is at least as credible as any of the alternatives. It's certainly the case that all of the advances in information retrieval, speech recognition, machine translation, image understanding, and genomics, of the last twenty years are basically advances in extracting statistical regularities for very large data sets. No other approach has had even a teeny fraction of that impact.
Let me talk a bit more about genomics, because it touches on what we all recognize as the most impressive examples of the “capacity to reason, conjecture”: scientific discovery. Evolution predicts that the genomes of organisms will have similarities and dissimilarities governed by descent with modification and by conservation of genes for valuable traits. Computational biologists developed a variety of statistical pattern-matching methods that discover the specific conserved and mutated elements by comparing the genomes of related species. These methods are able to find generalizations that lead to experimentally testable predictions of, for instance, important conserved regulatory modules. That's just one among many discoveries that would be totally impossible without statistical generalization from huge data sets. These statistical methods are discovering new facts about biological evolution and function. Is that intelligence?
Matt's criticism appears to me archaically essentialist. For him, there must be something in “intelligence” beyond mere statistical pattern discovery. In the same way as for some there must be something to life beyond mere blind mutation and natural selection.
I prefer empirical evidence to metaphysical claims. While it is obvious that current statistical generalization methods are very coarse compared with the most refined products of the human intellect, they are already able to amplify that intellect in ways that were unimaginable just a couple of decades ago, and we see no evidence of fundamental roadblocks to further progress. Just, sometimes, failures of imagination.

5 comments:

Natalie said...

Beautifully written, Fernando!

steve said...

applauds

Matthew Hurst said...

Fernando,

I'll find time to write a follow up to your post. However, for now I'd like to highlight a basic weakness in your post. The pattern matching algorithms haven't discovered new facts about biological evolution - the scientists have. The same scientists that wrote the 'find me bits of these sequences that are the same' tools.

Don't be so sure that tools that can amplify a human capability are themselves intelligent. A telescope, perhaps the purest example of such a tool, is not an eye. Nor is it capable of discovering a new planet or galaxy.

vladmirlambda said...
This comment has been removed by the author.
Mitra said...

You are forgetting that google indirectly uses human intelligence.