Hans op de Beeck, KU Leuven, Belgium, has given a talk on Zoom on September 12th.

https://ppw.kuleuven.be/lbp/lbpMembers/00029058

Short abstract:

The adult human brain contains very rich, multidimensionalrepresentations of objects and animals. It has proven difficult to understandhow and why these representations and their properties come about. Luckily, manyof these representational properties also emerge in deep artificial neuralnetworks (DNNs) trained in image classification. Here I will illustrate howthese DNNs help us understand how some properties of human representationsmight develop. A general phenomenon in human perception is the perception ofanimal- or person-like features in inanimate objects, such as clouds andsilhouettes in your bedroom at night. In human culture, it is also common to find objects thatresemble animals on purpose (e.g., toys), and the perception of such objectsas animal-like seems obvious to humans. We have shown that this “Animal bias” forzoomorphic objects is absent in DNNs, despite the ability of thesenetworks to categorize animals from objects. Yet, we successfully induced this bias in DNNs trained explicitly withzoomorphic objects. Alternative training schedules, focusing on previouslyidentified differences between the brain and DNNs, failed to cause an Animalbias. Specifically, we considered explicit training in the superordinate distinctionbetween animate and inanimate classes, the sensitivity for faces andbodies, the bias for shape over texture, and the role of ecologically valid categories. These findings providecomputational support that the Animal bias for zoomorphic objects is a uniqueproperty of human perception yet can be explained by human learning history during development.