Perceptual Learning in Autonomous Systems

Yuri Ivanov of Honda Research Labs and MIT. He develops machine learning for the Honda humanoid robot, ASIMO. He believes that an adaptive user interface requires some way for a human to direct the adaptation in some naive way. He wants a solution that is on-line and practical so that it works in real-time.

Some of the things that he has investigated are different ways of keeping data points to best represent clusters. This is motivated by the fact that new data points are constantly arriving, and given resource constraints you must choose what to keep and what to throw out. So he came up with a bunch of ways of defining the important data points in a cluster and evaluated how each idea did at reconstructing the cluster.

He also showed a couple of videos of a synthetic dog learning commands from a human giving commands. I couldn't follow his formalism for how the dog learned the command that the trainer wanted. It sounded to me like he was claiming that he developed POMDP's, but he didn't use the standard formulation of a POMDP although the problem solved seemed to be the same: hidden state, distribution over rewards in a state and reward feedback.

After asking a question I found out that he is compactly representing a state space and then applying a POMDP-like techniques to the reduced state space, but he is not modelling a state transition model. So his technique could be considered solution to a POMDP with a known-uniform state transition model.

Conclusion: This could be cool if there was some efficiency achieved by virtue of the fact that the state-transition model is static - not sure if this is the case.

Second part of the talk is a technique of recognizing a person from ASIMO. It uses a combination of features: a face recognizer, a speech classifier, a speaker classifier, a clothing classifier, and a height classifier. These are fed into a Naive(?) Bayesian classifier which has some external knowledge of how good the features are at person classification.

I think I missed a key point because he suggested that each person can have a different weight on each feature so that someone who has a distinct voice can be recognized by voice, but someone with distinct clothes can be recognized by that. I'm not sure how the feature weights are figure out unless there is a separate classifier for each person. That would mean classifier one says it is Don or it is not Don with some confidence level. Then you choose the classifier which is the most confident.

Conclusion: Interesting problem with robust feature set, but I'm not convinced that the approaches are to solving the problem are that deep.