Comments on the book "Architects of Intelligence" by Martin Ford in the light of the SP Theory of Intelligence

J Gerard Wolff (1)

1 CognitionResearch.org

The book "Architects of Intelligence" by Martin Ford presents conversations about AI between the author and people who are influential in the field. This paper discusses issues described in the book in relation to features of the "SP System", meaning the "SP Theory of Intelligence" and its realisation in the "SP Computer Model", both outlined in an appendix. The SP System has the potential to solve most of the problems in AI described in the book, and some others. Strengths and potential of the SP System, which in many cases contrast with weaknesses of deep neural networks (DNNs), include the following: the system exhibits a more favourable combination of simplicity and expressive or descriptive power than, arguably, any alternatives; the system has strengths and long-term potential in pattern recognition; the system appears to be free of the tendency of DNNs to make large and unexpected errors in recognition; the system has strengths and potential in unsupervised learning, including grammatical inference; the SP Theory of Intelligence provides a theoretically coherent basis for generalisation and the avoidance of under-or over-generalisations; that theory of generalisation may help driverless cars avoid accidents; the system, unlike DNNs, can achieve learning from a single occurrence or experience; the system, unlike DNNs, has relatively tiny demands for computational resources and volumes of data, with potentially much higher speeds in learning; the system, unlike most DNNs, has strengths in transfer learning; the system, unlike DNNs, provides transparency in the representation of knowledge and an audit trail for all its processing; the system has strengths and potential in the processing of natural language; as a by-product of its design, the system exhibits several different kinds of probabilistic reasoning; the system has strengths and potential in commonsense reasoning and the representation of commonsense knowledge; other strengths of the SP System are in information compression, biological validity, scope for adaptation, and freedom from catastrophic forgetting. Despite the importance of motivations and emotions, no attempt has been made in the SP research to investigate these areas.

Mots clés: artificial intelligence; SP Theory of Intelligence; deep neural networks; unsupervised learning; generalisation; transfer learning; probabilistic and commonsense reasoning; information compression

Domaines: Intelligence artificielle [cs.AI]

J Gerard Wolff. Comments on the book "Architects of Intelligence" by Martin Ford in the light of the SP Theory of Intelligence. 2019. ⟨hal-02061171⟩