Complexity scale in AI

There is a significant amount of effort being devoted in the research space to map machine learning and AI, but it has been challenging to categorise them according to their ‘intelligence’. Thus far, attempts at categorisation have been limited to looking at their generic ability to solve new problems, and at the speed with which they adapt to these problems. A more straightforward way of understanding AI is to classify AI systems by their complexity.

High complexity AI applications

  • Autonomous vehicle
  • Machine translation tool
  • Care companion robot
  • Chat bot
  • Surgical or pharmacy robot
  • Mammogram interpretation system
  • ECG interpreter
  • Diagnostic decision support system
  • Speech driven radiology report tool with SNOMED coded output

Middle complexity AI modules or components

  • Natural language to SNOMED code processing module
  • Image processing module
  • Text to speech module
  • Knowledge based or expert system module
  • Signal processing & classification module
  • Recommender module

Low complexity AI reasoning methods

  • Deep learning module
  • Ensemble methods (e.g. Random Forest Models)
  • Neural networks
  • Object segmentation algorithm
  • Signal processing algorithm / filter
  • Generative adversarial networks
  • Time series analysis
  • Graphical models
  • Decision trees, rule induction e.g. CART
  • Clustering algorithm
  • Classification algorithm
  • Regression – linear, multiple, logistic
  • Inference engine for rules or frames
  • Argumentation, temporal or spatial reasoner e.g. QSIM
  • Text generator using DCGs
  • Case-based reasoning algorithm

The lowest level of the complexity scale comprises single specific reasoning methods (e.g. neural networks, pattern recognition algorithms). When these reasoning methods are combined with other functions (e.g. a database or user interface), we get ‘modules’, which sit at the next level of complexity and are the problem-solving components of a system. At the top level of complexity, we have applications or packaged systems comprising two or more of these modules (e.g. an autonomous robot).

Algorithms in healthcare are not a new phenomenon and have been deployed for decades. What we have attempted to show here is how technology utilising intelligence within its algorithms can fall under many different subsections and with varying degrees of complexity. We encourage developers and industry to be transparent as to what complexity or methodology they are utilising when they state that they use ‘AI’. We would also like to encourage those investing in these technologies to understand what type of AI is being developed, how complex it is, and indeed question what the ‘A’ in AI truly represents.