====Artificial Intelligence - Strong AI====

Strong AI (as opposed to [[weak_vs_strong_ai|weak]] AI) is also known as //machine learning//, 

While weak AIs are hand-made by people, machine learning AIs are //taught//. They learn by being fed sample data.

The key advantage of these types of AI is that they can make incredibly complex decisions based on huge amounts of data, arriving at insights and connections that humans would disregard.

===Advantages===

   * Can work on huge volumes of data,
   * Can make extremely complex decisions
   * Work without assumptions or preconceptions
   * Can learn using subtle patters that humans would ignore
   * Are extremely flexible

===Disadvantages===

   * Need a wide range of training data
   * Unable to explain their decisions
   * Can make incorrect connections with insufficient training
   * More power-intensive than weak AIs.

===Training===

{{ :cognition:ai_basic_layout.png?200|}}Learning AIs need //something// to learn with. The process of exposing an AI to data to teach it how it should behave is called [[training|training]].

The training data for our simple 'coffee temperature' problem (left) is fairly basic - we just need to provide a sample data for several different combinations of //input// and //output//.

^Temperature^Too Hot^
|60|No|
|70|No|
|80|Yes|
|90|Yes|

The AI is trained with this information. It learns that the transition between 'too hot' and 'ok' lies somewhere between 70 and 80 degrees. The more specific the training data, the more control you have over exactly where that transition happens.
 
In reality, the output of an AI such as this will often be //analogue//, showing how **confident** it is in its conclusions. 

{{:cognition:confidence.png?400|}}

In this example of analogue output, the confidence changes as the temperature increases, crossing the threshold between negative (good temperature) and positive (too hot) at 75 degrees.

At 40 degrees the confidence is below -1 - a very strong negative result. At 90 degrees it's above 1 - a very strong positive result. At 75, it's torn between the options and unsure. 

To reduce the amount of uncertainty, you could introduce more training data - adding values 
around 72, 74, 76 and 77 will result in a much more sudden change in confidence.

===AI Extrapolation===

The AI is able to continue operating even when your data goes outside its experience - such as the graph above where it continues to produce output for temperatures below 60 and above 90 degrees.

Note however that these results are effectively //educated guesses//, and accuracy usually falls off quite quickly. It's always best to ensure you provide a wide selection of training data to ensure that the AI knows how to behave at your extremes. 

===Adding Complexity===

Of course, this is far too simple an example of a strong AI. To understand where it comes in useful, we'll need to look at something [[complex example of machine learning|a more complex example]].