Les machines y voient-elles quelque chose ? - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Astérion Année : 2021

Can machines really see? Vision and representation in the light of deep learning

Les machines y voient-elles quelque chose ?

Résumé

Computer vision is one of AI’s most successful fields. In the last twenty years, machines have become increasingly good at extracting information from images and at identifying objects. But does this mean that machines really can see, or is computer vision just a fancy metaphor for object detection? This paper aims to provide a reasoned answer to the question. First, three criteria for vision attribution are reviewed and it is argued that a functionalist criterion, in terms of exploitable internal representations based on visual stimuli, fares better than behaviourist or phenomenological ones. The functionalist criterion is then applied to vision algorithms based on neural networks and it is argued that such machines can indeed see, insofar as neural networks are precisely trained to generate representations of the visual data they are fed with. Those representations present the specific traits associated with successful training: They are hierarchical, robust, and versatile. We argue that these properties may be used as further constraints on representational devices in general, helping to solve some of the classical issues faced by teleosemantic theories of representation such as that expounded by Dretske.

Dates et versions

hal-04437997 , version 1 (05-02-2024)

Identifiants

Citer

Denis Bonnay. Les machines y voient-elles quelque chose ?. Astérion, 2021, 25, ⟨10.4000/asterion.7611⟩. ⟨hal-04437997⟩
10 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More