Mechanisms of human dynamic object recognition revealed by sequential deep neural networks

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Human visual system excels at rapid object recognition in dynamic environments.
• Mechanisms underlying fast, dynamic object recognition are not fully understood.
• Current computational models often process images independently, limiting dynamic understanding.

Purpose of the Study:

• To develop and compare deep learning models for dynamic object recognition.
• To investigate computational mechanisms underlying human performance in rapid visual recognition.
• To identify key factors contributing to efficient and fast object recognition in changing visual scenes.

Main Methods:

• Developed deep learning models contrasting feedforward vs. recurrent, single-image vs. sequential processing.
• Compared model performance against human recognition data (N=36) across various image durations (13-80 ms/image).
• Incorporated adaptation mechanisms into recurrent models to assess impact on performance and dynamics.

Main Results:

• Models integrating images sequentially via lateral recurrence best matched human performance and trial-by-trial responses.
• Model performance correlated with image presentation duration, mirroring human capabilities.
• Adaptation significantly enhanced dynamic recognition and accelerated representational changes, reducing computational needs.

Conclusions:

• Sequential image integration through lateral recurrence is crucial for rapid, human-like dynamic object recognition.
• Adaptation mechanisms play a vital role in improving efficiency and speed of visual recognition.
• Findings offer insights into the neural computations enabling effective object recognition in dynamic visual worlds.