#038 - Prof. KENNETH STANLEY - Why Greatness Cannot Be Planned - 2021

Title : #038 - Prof. KENNETH STANLEY - Why Greatness Cannot Be Planned Author(s): Machine Learning Street Talk Link(s) : https://www.youtube.com/watch?v=lhYGXYeMq_E

• The gradient of interestingness:
  • Novelty does not imply interestingness.
  • Interestingness implies novelty.
• Neglected dimensions of intelligence:
  • Divergence.
  • Populations.
  • Diversity preservations.
  • Stepping stone collection.
  • Environment.
• Campbells law.
• Creativity is a search problem, and explicit objectives block this.
• Stepping stone to intelligence does not (#DOUBT or may not?) resemble intelligence.
• Exploration in Kenneth's work is completely different from exploration in the context of Reinforcement learning (RL).
• Novelty search can be a proxy for interestingness.
• It is not clear how to formalize curiosity. (#NOTE Maybe Schmidhuber's work is a step in the right direction?)
• Humans know great art and music when we see it, these latent capabilities we have maybe could be made more explicit if we have the right kind of assistance.
• Open endedness: Not how to learn something, but learn everything.
• Evolution created everything in 1 run.
• Coevolutionary algorithms still get stuck, GANs ultimately interpolate between samples (i.e. they do not introduce anything new).
• What explains open endedness and complexity explosions? We are beginning to understand: Divergence instead of convergence. The system must generate the problems and the solutions - we the organisms are both the problems (i.e. create opportunities) and solutions.
• (#NOTE See minimal criterion coevolution paper, and POET algorithm, gradient informed mutation operator)
• Interestingness implies accumulating information.
• The whole point is we don't know how to get to a solution from scratch, and it is often counter-intuitive. If we knew the stepping stone, we would just follow it.