Biologically Inspired Model for Visual Cognition Achieving Unsupervised Episodic and Semantic Feature Learning.

Area of Science:

• Computational Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• Biologically inspired computational models offer novel solutions for visual recognition.
• Existing models often lack unsupervised learning capabilities for complex feature extraction.

Purpose of the Study:

• To propose a framework mimicking the primate visual cortex for active and dynamic learning.
• To achieve unsupervised learning of episodic and semantic features for object cognition.

Main Methods:

• Utilizing a deep neural network (DNN) for unsupervised learning of object key components and spatial relations.
• Employing contour detection to compute semantic geometrical values from learned episodic features.
• Developing a method for forming general class knowledge including key components, spatial relations, and semantic values.
• Implementing dynamic model updating with high-confidence test samples.

Main Results:

• Successful unsupervised learning of episodic features (key components, spatial relations) without prior knowledge.
• Effective learning of semantic features based on episodic features via contour detection.
• Formation of concise general knowledge representations for object classes.
• Achieved high performance in classification and semantic description, with dynamic updating capabilities.

Conclusions:

• The proposed framework successfully mimics primate visual cortex functions for object recognition.
• The model demonstrates robust unsupervised learning and dynamic updating for visual cognition tasks.
• The framework shows potential for generalization to various object recognition applications.