An AI Learning Hierarchy
Communication of the ACM December 2024 edition covered an opinion from Peter J. Denning and Ted G. Lewis on an AI Learning Hierarchy. The opinion categorizes learning machines at eight different levels.
- Level 0 (Basic Automation): No training required. Ex: classical control theory.
- Level 1 (Rule-Based Systems): Rules are learned from expert knowledge.
- Level 2 (Supervised Learning): System learns from examples.
- Level 3 (Unsupervised Learning): System learns automatically without any guidance on which category denotes what.
- Level 4 (Generative AI): Ability to generate information based on samples seen in the past.
- Level 5 (Reinforcement Learning): Learn by trying.
- Level 6 (Human-machine Interaction AI): Collaborating with human.
- Level 7 (Aspirational AI): Artificial General Intelligence.
This hierarchy classifies AI machine by their learning power. A machine more powerful at learning than another if, in a reasonable time, it can learn to perform some tasks that the other cannot. The opinion further delves into the details of each hierarchy and discuss examples of each level in the hierarchy.
While I agree with overall classification, I had concerns about the Generative AI where the authors limited the level to only text-based AI and natural language processing (NLP). The way I see Generative AI is ability to learn a distribution and then sample from that distribution. When we define Generative AI this way, the scope broaden to other data types as well, e.g. time series.
A key highlight for me from the article was, “The hierarchy shows that none of the machines built so far has any intelligence at all, leading to the intriguing possibility that human intelligence is not computable”. This question—whether machines can compute human intelligence—left me deeply intrigued about the future of AI.