The Difference Between AI and Machine Learning


The terms AI and machine learning are often used together, but they are quite different things. To a large extent machine learning can be considered as just one of many methods to build agents (things that do something) which exhibit intelligence. The word ‘intelligence’ does not bestow any mystical quality upon the agent, since an agent is intelligent when its actions meet some specified performance criteria. An automated floor cleaner is considered intelligent when it keeps the floor clean – for example. Intelligent agents are the desired outcome of any AI project, but the range of problems goes well beyond those which machine learning can address. Machine learning is a set of techniques which detect patterns of behavior and allow an agent to learn as its environment changes. Some AI problems have almost nothing in common with machine learning. Calculating the best route between London and Edinburgh (or any two cities) is a search problem, and this is not usually categorized as a machine learning problem.

AI embraces many techniques in addition to machine learning, including:

  • Search (as in the example given above)
  • Optimization – to allocate resources in the most efficient way for example.
  • Game theory – to solve adversarial and/or cooperation problems
  • Logic – e.g. business rules
  • Planning
  • Knowledge representation
  • Natural language processing – and at least a dozen other methods.

In essence there are many problems in AI which do not require an agent to learn. If it does need to learn, then machine learning is likely to be relevant. However it should be obvious that machine learning is becoming an important part of AI and may eventually come to dominate the way many intelligent agents are built. A classic application of machine learning is image recognition, and specifically the use of deep learning as a way of automatically recognizing objects in a picture (a cat or truck for example). Most of the exciting developments in AI are associated with machine learning, but hopefully the reader will see that the two are really quite different.