This is an interesting video from Eric Horvitz at Microsoft Research that covers a little on the use of Inference Networks. Real world applications include traffic prediction and hospital admission. Eric also talks about how they constructed a model for the receptionist of the MS Research building. Video below the jump.
Inference Networks
Bayesian Belief Networks use Bayesian probability theory to infer a certain state given data inputs. Inference networks model expert data and use statistics from different sources to predict the future.
One real world application presented is the Seattle road traffic network. Traffic information comes down to a task of novelty detection - to predict when something new and unexpected happens. With maps integration, their Clear Flow system (or traffic sensitive routing) will help route you around the traffic. The data scales used are huge and now span the breadth of America.
The other real world application covers readmission of patients into hospital. The network can infer when a patient leaving hospital is likely to come back.
Top tip
Eric points out that when training your algorithm, only train it on 80% of the data available and leave 20% remaining to test it. That's a tip you'll find in lots of places and I heartily recommend it.
Decisions from learning
Eric mentions that there is a difficulty with making decisions from prediction. Once you've predicted that someone is going to be readmitted, deciding what to do is a second problem.
The Digital Assistant
This is an interesting piece of work where a system tracks the people talking to it and makes inferences to predict what they need. There are interesting human computer interfacing challenges here too, such as predicting when someone is talking to themselves. I find the voice and talking head spooky but then this is a research project and most AI research is spooky by nature.
0 comments:
Post a Comment