Key Arguments

• Eleanor Rosch’s experiments show that 'subjects consistently classify objects not in terms of necessary and sufficient features, but as more or less distant from a typical example or prototype,' which contradicts the classical, feature‑list model presupposed by much symbolic AI.
• Roger Shepard and Jacqueline Metzler demonstrate that 'subjects rotate mental images at constant speeds,' suggesting continuous, quasi‑spatial operations over images rather than discrete symbol manipulations.
• Workers in AI themselves admit that 'insofar as images are different from symbolic descriptions they cannot be accounted for in an information-processing model,' so robust experimental evidence for image‑like processes is embarrassing for the information‑processing conception of cognition.

Source Quotes

Alternatives to the Traditional Assumptions Introduction 1. There is, however, a new interest in what might be called "experimental phenomenology." Eleanor Rosch, for example, has Shown that subjects consistently classify objects not in terms of necessary and sufficient features, but as more or less distant from a typical example or prototype. See "Principles of Categorization" in Cognition & Categorization, E.

Erlbaum Press, 1977). There is also much interest in the work of R. N. Shepard showing that subjects rotate mental images at constant speeds. See R.

701 703. Such experimental work with images is an embarrassment to workers in AI, since all agree that insofar as images are different from symbolic descriptions they cannot be accounted for in an information-processing model. 2.

Key Concepts

• There is, however, a new interest in what might be called "experimental phenomenology." Eleanor Rosch, for example, has Shown that subjects consistently classify objects not in terms of necessary and sufficient features, but as more or less distant from a typical example or prototype.
• There is also much interest in the work of R. N. Shepard showing that subjects rotate mental images at constant speeds.
• Such experimental work with images is an embarrassment to workers in AI, since all agree that insofar as images are different from symbolic descriptions they cannot be accounted for in an information-processing model.

Context

Introductory note to Part III ('Alternatives to the Traditional Assumptions'), where Dreyfus points to contemporary cognitive psychology experiments on categorization and mental imagery as empirical support for phenomenological critiques of information‑processing AI.