Draft: What Humans Can Learn From Machine Learning

Human Learning differs from Machine Learning in important ways.

However, there are useful things we can take from ML

  • the alpha parameter
  • momentum: https://en.wikipedia.org/wiki/Stochastic_gradient_descent#Momentum
  • optimality in the face of an impossibly large search space <- needed this source
  • generalize > memorize by using a testing set
  • epsilon exhaustion and Probably Approximately Correct learning
  • naive bayes - how can it be right
    • So you just need to be directionally correct.
  • random restart hill climbing
    • if theres a v specific point that is superbly better than others, you're in a bad world
  • simulated annealing - hot and cold

  • convert definites to maximum likelihood/probability densitiy estimation?

  • transfer learning as a bootstrap -> genetic evolution?
  • eager learning vs lazy learning

RL

  • policy CAN change depending on whether youre playing with infinite horizon
  • temporal attribution problem
  • MDP => Bellman Equation => Value iteration strategy (rewards vs utiltiy vs always adding truth)
  • PLANNER: model -> policy
  • LEARNER: transitions -> policy
  • MODELER: transition -> model
  • SIMULATOR: model -> transition
  • exploration/exploitation/explain dilemma

https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcRpiXLm24rgkZJbENxamD6f3ZDJfK7viU5gbhoGrwj1jp-AMBXVVg https://swizec.com/blog/only-self-help-business-advice-you-need/swizec/7190

linear digressions

  • conv nets - move stuff around - translational invariance
  • rnn - languages
  • micromanagement
    • starcraft etc alphazero were total untrained
    • given infinite time and data ofc we let total untrain
    • but dont have infinite data so helps to bootstrap

human in the loop


⚠️ You are reading an unpublished, incomplete draft. Questions are welcome but feedback/criticism may be premature.