Purchase this article with an account.
Nicola van Rijsbergen, Guillaume Rousselet, Philippe Schyns; Reconstructing a representational space of learned faces. Journal of Vision 2015;15(12):700. doi: 10.1167/15.12.700.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
When we learn to discriminate between new faces, we memorize the information that best identifies and discriminates them. To study these face memories, we first parameterized the information of 97 face identities using a recursive sinusoidal basis (12 orientations, 2 polarities and 5 spatial frequencies), and computed the principle components of this space (excluding 4 face identities). We trained participants to discriminate these 4 identities by naming them. Participants trained until they reached 100% accuracy (< 40 trials). By construction, the memorized diagnostic information enabling 100% accuracy must be expressible as specific parametric values in the reduced sinusoidal basis. To understand these face memories, for each participant we performed reverse correlation using randomized parameters in the sinusoidal basis (Gosselin & Schyns, 2003, Mangini & Biedermann, 2004, van Rijsbergen et al. 2014). Trials were presented in blocks of 10, for a total of 600. Before each block a memory screen was shown containing the four test identities labeled with their names. On each of the 10 subsequent trials participants selected one of the named identities from an array of three faces randomly generated with the sinuisoidal basis. Crucially, participants did not know in advance of these trials which of the four memorized identities they would be asked to select until after the memory screen disappeared. Every 120 trials, participant memory performance was retested. Ideal observers provide the limit of identity representation. After 15,000 trials an Ideal Observer achieved a parameter correlation of 0.2 with the test identity. A PCA observer that randomizes combinations of principle components reached maximum performance in 200 trials, but failed to learn one identity. A hybrid observer that switched learning strategies from principle components to individual parameters over time, performed better across all 4 identities, suggesting that human performance could be sped up by combining the two
Meeting abstract presented at VSS 2015
This PDF is available to Subscribers Only