Invariant recognition drives neural representations of action sequences.

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Computational Neuroscience

Background:

• Human action recognition involves discriminating actions despite visual transformations.
• Understanding the neural basis of action recognition requires computational models.
• Convolutional Neural Networks (CNNs) excel at object recognition but their role in action recognition is less understood.

Purpose of the Study:

• To test if invariant action discrimination explains neural representations of action sequences.
• To investigate if CNNs can model human action recognition across transformations.
• To determine if CNNs' internal representations align with human neural data.

Main Methods:

• Utilized spatiotemporal Convolutional Neural Networks (CNNs) for video action categorization.
• Modified CNN architectures to enhance invariant action recognition performance.
• Compared CNN representational similarity with human neural recordings.

Main Results:

• Spatiotemporal CNNs accurately categorized video action stimuli.
• CNN modifications improving invariant recognition yielded representations matching human neural data.
• Performance on invariant discrimination correlated with neural representation similarity.

Conclusions:

• Invariant action discrimination is key to neural representations of actions in the brain.
• CNNs provide a framework for understanding visual intelligence in action perception.
• The invariant recognition framework extends from static images to dynamic action sequences.