Decoding face identity: A reverse-correlation approach using deep learning.

Area of Science:

• Cognitive Science
• Computer Vision
• Neuroscience

Background:

• Traditional face recognition relies on subjective feature selection.
• Objective methods are needed to identify critical facial features.
• Deep convolutional neural networks (DCNNs) show promise in modeling human perception.

Purpose of the Study:

• To objectively identify facial features crucial for face recognition using a data-driven approach.
• To leverage DCNNs to understand the neural representations underlying face discrimination.
• To propose a new paradigm for exploring critical facial features in face recognition tasks.

Main Methods:

• Trained a DCNN (VGG-FD) for human-like facial identity discrimination.
• Employed reverse-correlation to link network responses with internal representations.
• Utilized representational similarity analysis (RSA) to analyze network performance and identify key features.

Main Results:

• Identified specific facial regions, including eyes, eyebrows, and the central facial area, as critical for identity discrimination.
• Found that significant representational differences in these areas increase the likelihood of perceiving faces as distinct identities.
• Demonstrated the importance of the eyes and central facial configuration in face recognition.

Conclusions:

• DCNNs can objectively identify critical facial features for face discrimination in a hypothesis-neutral manner.
• The eyes and central facial region play a significant role in face configuration and recognition.
• Advocates for using DCNNs and reverse-correlation as a novel paradigm for face feature research.