10.18.2009 - Moving in an uncertain world - Daniel Wolpert

    • Movement is the only we have of interacting with the world
      • "The only reason we have a brain is to be able to move"
    • Level of dexterity is much different between "robots" and humans.
    • There is noise and variability in:
      • Sensors
      • Motor output
      • Task
    • Decoding motor uncertainty
      • Bayesian learning
        • Beliefs are updated based on combing "data" with stored information
        • Bayes rule
          • P(A|B) = P(B|A) * P(A) / P(B)
        • P(state) = probability based on prior knowledge
        • P(sensory input| state) = likelihood of the sensory state matching a prior state
        • P(state|sensory input) = motivation or future prediction of knowledge
      • Predicting the consequences of action
        • Required to account for:
          • Control for delays
          • Mental simulation
          • Likelihood estimation
            • Do many predictions in parallel
          • Sensory filtering
            • Two types of information
              • Changes in the outside world
              • Changes that we cause - we can get that from the efference copy
            • We ignore changes that we cause
              • Showed throw a "tickle" experiment
                • The coupling between spatial and temporal sensory information is tightly coupled to whether the information is considered "self-produced"
              • Tit-for-tat experiment
                • Trying to match forces
                • The sensed force appears stronger than the force applied, which suggests that we subtract off our own actions.
              • Schizophrenia subjects may have difficulty in predicting what their own level of determining if forces are generated by themselves or from external sources.
      • Loss functions in movement
        • Possible loss functions
          • Only a hit matters
          • Error to some power
        • Turns out that the loss function appears to be mostly quadratic for small errors but linear for large errors
      • Optimal movements
        • Stereotypical motions are recognizable
        • How do we figure out which movements are "best"?
          • Signal dependent noise leads to a distribution of possible movements
          • The optimal movement is then the one that has the least variability in the end-point.
      • Decisions and changes of mind
        • Why do you change your mind once you have begun a particular motion?
          • This can be explained by estimating the integration of information as a random-walk
            • The extension that can explain for the change in movement by changing the boundary of decision during the delay of the motor command.