![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslDaphne Koller's 2001 talk: Representation, Reasoning and Learning: slides and MP3 (about 90 minutes total)
She begins talking about different kinds of representation, from concrete to abstract: Atomic Worlds, Propositional Worlds, and Object-Relational Worlds. At each step, the search space is dramatically reduced.
Then she talks about her problem of unifying logic and probability, and the solution: Probabilistic Relational Models, which bridges relational logic and probabilistic models, and is much more expressive Bayesian networks and Probable Worlds.
But expressivity just makes reasoning harder (in the computational complexity sense), right? Wrong! First of all, non-expressive languages often illusion of being easy to reason with: if the number of worlds is exponential.
Secondly, expressive languages may allow one to find shortcuts, which makes reasoning easier.
Web of influence: everything is correlated with everything. How can we deal with such complexity? Use approximations of locality and hierarchy. Herbert Simon: we can only make sense of the world to the extent that it obeys such constraints (so if our goal is human-level intelligence, we should be happy with this)
Then she concludes talking about systems that discover representations, systems that bootstrap their learning iteratively, and how the divide-and-conquer approach to AI can greatly benefit from some occasional bridge-building.
It's quite nice.
She begins talking about different kinds of representation, from concrete to abstract: Atomic Worlds, Propositional Worlds, and Object-Relational Worlds. At each step, the search space is dramatically reduced.
Then she talks about her problem of unifying logic and probability, and the solution: Probabilistic Relational Models, which bridges relational logic and probabilistic models, and is much more expressive Bayesian networks and Probable Worlds.
But expressivity just makes reasoning harder (in the computational complexity sense), right? Wrong! First of all, non-expressive languages often illusion of being easy to reason with: if the number of worlds is exponential.
Polynomial in the size of a representation is great, but not if the size is itself exponential. Specifically, many tasks are linear time in an atomic representation --- we simply enumerate the worlds. But the number of worlds is huge. Thus, more expressive languages are not necessarily harder.
Secondly, expressive languages may allow one to find shortcuts, which makes reasoning easier.
Web of influence: everything is correlated with everything. How can we deal with such complexity? Use approximations of locality and hierarchy. Herbert Simon: we can only make sense of the world to the extent that it obeys such constraints (so if our goal is human-level intelligence, we should be happy with this)
Then she concludes talking about systems that discover representations, systems that bootstrap their learning iteratively, and how the divide-and-conquer approach to AI can greatly benefit from some occasional bridge-building.
It's quite nice.