Predictive learning

(Information Science Master) Lecture

• Normally, we tend to create a map from sensation to action.

  • For example, in the case of a robot that grasps objects, we tend to learn what kind of movement should be made based on the visual image given.

• However, it is often the case that our own state and sensations change as a result of our actions (this is also the essence).

  • It is effective to predict everything that we perceive (including not only the environment but also our own sensations [bodily experience]).
    • Understanding what kind of control leads to what kind of movement is also part of the prediction.
    • Understanding how objects react when we move is also part of the prediction.
  • Let’s try to do it end-to-end, all at once.
  • It is also possible to calculate them separately based on theory, but it is very complex and difficult.

• How do we learn?

  • Provide sequence data of the entire environment (including ourselves).
    • Can we collect the data through human operations?
  • Learn from that sequence data and enable the robot to predict the next moment’s environment.

• The scope of (blu3mo) Self also connects to philosophical discussions.

  • A robot that has learned predictive learning will have a very narrow range of self.
  • The self is only its own “intention” (= predictive data), everything else is non-self (environment).

• Predictive learning is difficult with a normal Neural Network (FNN)?

  • It is difficult to predict based on only one frame of the previous state.
  • For example, in predicting the motion of an object that is moving back and forth,
    • We don’t know if a certain frame is for the forward or backward motion.
  • Therefore, RNN is good (because it can handle time series data).

• Application examples

  • Predictive Learning with AutoEncoder