Yoshiyuki Ueda¹; ¹Kyoto University

Face memory is highly susceptible to distortion due to various factors, and recognition tests are often used to investigate it. Although these tests can reveal whether memories are preserved, they cannot clarify how they are distorted. The reverse correlation image classification method (Dotsch & Todorov, 2012) offers a potential solution. In this paradigm, participants memorize a face and then select the face most similar to the one they memorized from a pair of faces generated from the base face image, with inverted polarity noise added. By aggregating the noise patterns from the selected faces, it is possible to construct classification images (CIs) that reflect the features participants used in their judgments in a bottom-up manner, and these CIs are considered to represent participants’ mental images. However, it is currently up to the subjective judgment of a third person to decide whether the facial features of a CI correspond to those of the actual memorized face. This study applied a deep learning-based face recognition model (Serengil & Özpınar, 2024) to CIs to assess their similarity to memorized faces. The results showed that the similarity between CIs and memorized faces became stable after completing more than 200 trials, and CIs included features of the memorized faces. However, significant individual differences and variations due to the specific faces used in the experiment were observed. Furthermore, the group CI for one group was more similar to the memorized face compared to the other group that memorized a different face, but the face model was not always able to correctly identify which face had been memorized. These findings highlight the limitations of the current method and suggest the need for more elaborate techniques to visualize our representations of faces.