Recognition vs. cognition

Recognition

Neural net AIs are very good at recognizing and classifying, there are many successful examples.

Recognition means classifying inputs into groups. The input can be images, forms (a form filled out by someone applying for a job or a loan), data (a stream of data from a machine), etc.

Applications include:

  • face recognition - Picking a face out of a crowd and identifying its owner
  • repair prediction - Recognizing when a machine is likely to need repair
  • sales prediction - Recognizing when a customer is likely to buy a product
  • stock trading - Recognizing when it’s time to buy or sell a stock
  • handwriting recognition - Turning handwritten letters into text
  • steering and braking - Recognizing when to change the direction or speed of an autonomous vehicle
  • translation - Translating a phrase from one language into another

Cognition

There are no examples of a neural net AI that can perform general cognition. There are examples of AIs which perform cognitive tasks, they just don’t use neural nets to do so. When neural nets are not a good tool for the job, semantic models are used by AIs instead.

Human-like intelligence covers a broad range of cognitive skills that go beyond recognition and classification. Some examples; this is only a very small partial list:

  • Being taught means you can directly add an individual fact to your knowledge after hearing it only once.
  • Reasoning means you can combine facts to discover new facts without direct observation.
  • Reflection and introspection means you can reason about your experiences and learn from them.
  • Using analogies make it possible to reason about one situation by selectively using knowledge from another.
  • Storytelling means you can share a rich mental model with another intelligent being. This is done by unwrapping it into a linear stream of words, which the listener reassembles in their mind.
  • Asking questions means you can recognize a gap in your own knowledge and seek to remedy it.
  • Criticism makes it possible to identify a specific flaw in your knowledge.
  • Correction makes it possible to fix flaws in your knowledge if you can find them.
  • Attribution makes it possible to remember different sources for your knowledge, so you can see how they hold up in practice and adjust your trust in those sources accordingly.
  • Proof makes it possible to rely on knowledge with high confidence.
  • Managing counterfactuals makes it possible to imagine things that you’re not observing at the moment. This in turn makes it possible to have an imagination, keep fact and fiction separate, recall the past, make plans for the future, compare alternatives, make a wish, or test hypotheses
  • Generate counterexamples makes it possible to validate your recognizer by testing it without a training set.
  • Abstraction means you can generalize specific ideas into abstract ideas which capture their essential truth.
See also