Gender Perception From Gait: A Comparison Between Biological, Biomimetic and Non-biomimetic Learning Paradigms

Area of Science:

• * Cognitive Neuroscience
• * Computational Neuroscience
• * Biomechanical Engineering

Background:

• * Understanding human perception of biological motion is crucial for developing advanced artificial systems.
• * Existing models for biological motion perception present conflicting learning system theories.
• * Automatic gait classification requires effective computational models that mimic biological processes.

Purpose of the Study:

• * To investigate human mechanisms for biological motion perception, specifically gait.
• * To evaluate biological, biomimetic, and non-biomimetic computational models for gait-based gender identification.
• * To compare the performance of computational models with human observers in gait perception tasks.

Main Methods:

• * Conducted psychophysical experiments with 21 human observers observing gait stimuli.
• * Developed and tested computational models: biological, biomimetic artificial neural networks (ANNs), and non-biomimetic models.
• * Performed computational experiments without gender-specific modifications to models or stimuli.

Main Results:

• * Human gait perception utilizes a generic memory-based learning system, resolving ambiguities in existing theories.
• * Memory-based ANNs demonstrated biomimetic capabilities, effectively emulating biological neural networks.
• * The biomimetic ANN model achieved 83% accuracy in gait classification, outperforming human observers (66%) and non-biomimetic models (83%).

Conclusions:

• * A memory-based biomimetic model is optimal for generic artificial gait classification.
• * This model exhibits human-like sensitivity for gender identification while offering superhuman performance potential.
• * The findings support the application of biomimetic ANNs in diverse gait perception objectives.