Improved Visual Recognition Memory Model Based on Grid Cells for Face Recognition.

Area of Science:

• Computational Neuroscience
• Computer Vision
• Machine Learning

Background:

• Traditional facial recognition requires extensive training data for feature extraction.
• Prior brain-inspired models using grid cells were limited to familiar object recognition.
• Existing models struggled with unfamiliar facial expressions and required many training samples.

Purpose of the Study:

• To develop an improved computational model for facial recognition using grid cells.
• To enhance recognition accuracy and efficiency with minimal training data.
• To address limitations of previous models in recognizing unfamiliar faces and expressions.

Main Methods:

• Proposed an improved computational model incorporating grid cells for facial recognition.
• Utilized the histograms of oriented gradients (HOG) algorithm for initial stimulus identity hypothesis.
• Employed Hebbian learning for representing feature positions using grid and feature label cells.

Main Results:

• The improved model achieved high recognition accuracy with only one training sample per face.
• Demonstrated robustness against image occlusion and variations in size or scaling.
• Reduced the probability of false hypotheses and the average number of fixations for successful recognition.

Conclusions:

• The proposed grid cell-based model offers efficient facial recognition with minimal training samples.
• This approach surpasses conventional neural network models like CNNs and DBNs in performance.
• The findings provide insights for developing efficient, small-sample-based recognition systems using neural networks.