Predictive Coding Theory

(Information Science Master) Lecture on Cognitive Robotics

• “Computational Psychiatry”

• A computational model of the brain and cognition based on prediction error minimization#constructive approach

  • Calculates the error (prediction error) between prediction (output) and sensation (input)
  • Weighs the prediction error based on the uncertainty (reliability) of the sensory input at that time
    • The more certain the sensory input at that time, the more weight is given to the prediction error
  • Learning, perception, and action are assumed to minimize the weighted prediction error
    • To reduce the error, the “intention” is changed to modify the action
  • (Aiming to understand cognition through a constructive approach)

• Cognitive Robotics aims to test and verify such models of human cognition in an embodied manner

  • When this theory of cognition is implemented in a robot, it works well
  • By changing the input that represents “intention” based on prediction error minimization
  • Key point: Complex movements such as cooperation between robots can be achieved by the simple principle of prediction error minimization

• Furthermore, aiming to reproduce ASD (Autism Spectrum Disorder) using this model

  • Introduce damage to the part of the Neural Network that estimates uncertainty
  • As a result, an unnatural weighting is applied to the prediction error
  • Actually move the robot using this method
    • In the case of insufficient uncertainty
      • The error is evaluated larger than it should be
      • Therefore, the “intention” that should not actually change is modified, leading to abnormal behavior
    • In the case of excessive uncertainty
      • The error is evaluated smaller than it should be
      • Therefore, even if the environment changes, the error does not increase and there is no switch in intention, leading to abnormal behavior
  • As a result, psychiatric-like laws observed in the real world were also observed in the movement of the robot
    • The gap between the layer of the computational model and the layer of behavior in the real world was filled by Cognitive Robotics!
      • (blu3mo) Can this also be considered#embodied knowledge?