Séminaire ayant lieu dans le cadre du doctorat en informatique cognitive, en collaboration avec le centre de recherche CRIA et l'ISC    

Herbert ROITBLAT

Jeudi le 20 février 2025 à 10h30

Local: PK-5115   (Il est possible d'y assister en virtuel en vous inscrivant ici)   

TITRE :   Theories of Artificial Intelligence

RÉSUMÉ

GenAI models are computationally complex, but conceptually simple.  They are trained to fill in the blanks.  Model semantics is limited to word distribution patterns (Harris, 1956), yet many claim that GenAI models are capable of deep cognitive processes (such as reasoning and understanding). These assertions imply that cognitive processes can spontaneously emerge from these behavioral patterns—that a theory of cognition can be constructed at the purely behavioral level of word use patterns.  We have seen that movie before, but Chomsky (1959/1967) and a good deal of research with humans and animals, peaking in the 1980s, have demonstrated that a purely behavioral theory of cognition is not viable.  Those same research methods could be applied to the analysis of the latest forms of artificial intelligence, but their relevance is rarely recognized.  Instead, much of what passes for theoretical analysis of GenAI models is based on the logical fallacy of “affirming the consequent.” The models behave as if they had underlying cognitive processes, but their proponents fail to consider whether other explanations (e.g., stochastically parroting training data) could also explain the observations.  I will discuss the internal structure of GenAI models and how to understand them.  I will also offer some theoretical suggestions for approaching understanding and artificial general intelligence.

BIOGRAPHIE

Herbert ROITBLAT, is lead data scientist for Egnyte’s research and development in artificial intelligence. Formerly professor of Psychology, Marine Biology, and Second Language Acquisition, University of Hawaii, Roitblat’s work on how dolphins recognize targets underwater with biosonar led to significant contributions to early neural network research and a patent on a binaural sonar. His more recent work is on artificial intelligence: “Algorithms Are Not Enough: Creating General Artificial Intelligence” (MIT Press, 2020) argues that algorithms and neural network models cannot fully capture the complexity of human cognition or animal intelligence and suggests what is needed to achieve artificial general intelligence.  

RÉFÉRENCES

Bender, E. M., et al. (2021) On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM. 610-623.

Chomsky, N. (1967) A Review of B. F. Skinner’s Verbal Behavior. 142-143.

Harris, Z. (1954). Distributional structure. Word, 10(23): 146-162.

Huang K., & Chang, J.-C. (2023) Towards Reasoning in Large Language Models: A Survey. Findings of the Association for Computational Linguistics: ACL 2023, pages 1049–1065.

Roitblat, H. L. (2024). An Essay concerning machine understanding. arXiv preprint arXiv:2405.01840.

Roitblat, H. L. (2020). Algorithms are not enough: Creating general artificial intelligence. Mit Press.

Roitblat, H. L. (2017). Animal cognition. In: Bechtel & Graham, eds, A Companion to Cognitive Science, Wiley.