“Not conscious machines, omnipotent machines or even reasoning machines (yet), but statistical machines that automate and increasingly can outperform humans at certain pattern-recognition tasks. Computer vision, language understanding, anomaly detection and other fields have made immense advances in the past few years.” — Derrick Harris (Source)

Types of AI

  • Supervised Learning: Given a large set of inputs and corresponding outputs learn to come up with outputs for any set of inputs
  • Reinforcement Learning: Learn to control a task by executing it over and over in an environment and getting better at it each time (like a baby learns to walk)
  • Unsupervised Learning: try to make sense of unlabeled data — sort of find an explanation or pattern to something observed. The issue here is that the outcome can’t be evaluated
  • Expert Systems: Based on interviews of human experts be able to apply and combine that knowledge to reason better than human experts

What is AI good at:

  • Optimize: solve tasks more optimal than humans do, especially if many inputs need to be considered
  • Automize: Execute tasks that take little effort for a human (rule of thumb: takes human 1 second to do), but do it on a massive, parallel scale
  • Predict: Be able to extrapolate the near term future based on historic data and other signals
  • Reason: Given a lot of constraints and relations find the optimal solution
  • Search: not finding a solution by reasoning about it but being able to search vast amounts of data and find a solution that someone had found before / existed long before

Where AI is not good at:

  • Work on problems where only few samples exist
  • Problems where it is hard or impossible to formulate a reasonable cost function
  • Systems that need to generalize to a wider range of tasks
  • Systems where memory and computing power are a constraint

This is joint work with Sebastian Schaal.