Download: pdf (4 pages; 1.2M)

Abstract: Human expertise for recognizing unfamiliar faces has recently been called into question, highlighting a deficit when compared to familiar face recognition. We present simulations of a fixed-architecture deep convolutional neural network (DCNN) with different training regimens, highlighting the extent to which learning to recognize many "familiar" faces allows for robust, but incomplete, generalization to new "unfamiliar" faces as compared to performance after familiarization. With some training, verification performance for previously unfamiliar faces improves modestly, but the performance difference between unfamiliar and familiar faces is much smaller than the performance boost from pre-training on faces as compared to objects in the ImageNet 1000-way image classification database. We also assess the generalization performance of our networks to other fine-grained visual tasks such as bird species and car model verification. We find that expert face recognition does not improve generalization to birds or cars compared to a network trained on a subset of ImageNet with all vehicles and birds removed. We conclude that the specific learned statistics within a domain of visual expertise determine its generalization to other domains, in contrast with domain-general accounts which highlight level of processing over domain-specific statistics.

Copyright Notice: The documents distributed here have been provided as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.