ENS Paris, 28 May. Lecture on the limits of machine Intelligence

June 21st, 2018

Séminaire Postdigital
Entrée libre et sans inscription
Intervenants : Mario Klingeman et Matteo Pasquinelli
Date et heure : 28 mai 2018 – 14h-16h
École Normale Supérieure – salle Dussane
45, rue d’Ulm, Paris 75005

 

World Models and Biased Automata: On the Logical and Political Limits of Machine Intelligence

Abstract. Félix Guattari said that machines are « hyperdeveloped and hyperconcentrated forms of certain aspects of human subjectivity ». In which way then does machine learning « concentrate » social features and knowledge? My talk brings back the genealogy of the different techniques of machine learning (what is popularly termed as AI) to the paradigm of computational geometry that is shared by computer science pioneers such as Frank Rosenblatt, Marvin Minsky, John von Neumann and Konrad Zuse. If machine intelligence emerged with neural networks as a sophisticated technique of spatial pattern recognition, today’s computer scientists prefer the more general and abstract notion of input-ouput mapping. In this sense, (artificial) « intelligence » has become recently just the most effective solution to the problem of information compression. The problem of the definition of AI could stop here with this arid quantitative concept of « compression ». What information compression implies, however, is information loss and a new breed of errors, transformation of collective perception, apophenia (seeing things that are not actually there) and asymmetrical accumulation of knowledge yet to be discussed. If machine intelligence is distorted, however, this is not necessarily bad, as human knowledge and art are always a form of distortion, that is a hypothesis about the world.

What's this?

You are currently reading ENS Paris, 28 May. Lecture on the limits of machine Intelligence at Matteo Pasquinelli.

meta