Context Dependent Encoding Using Convolutional Dynamic Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Computational Neuroscience

Background:

• Sensory perception is heavily influenced by spatial and temporal context.
• Hierarchical models are crucial for understanding complex visual inputs.
• Predictive coding frameworks offer a biologically plausible approach to perception.

Purpose of the Study:

• To propose a novel hierarchical model, convolutional dynamic networks, for visual input representation.
• To effectively utilize spatial and temporal contextual information in visual perception.
• To develop a robust object recognition system for video sequences.

Main Methods:

• Developed a hierarchical model based on a predictive coding framework.
• Incorporated recurrent and top-down connections using empirical priors.
• Employed a smoothing proximal gradient method for efficient inference.
• Trained the model on unlabeled video sequences.

Main Results:

• The model learned a hierarchy of stable attractors representing object parts.
• Demonstrated effective utilization of contextual information for robust representations.
• Achieved stable object recognition in video sequences, even with corrupted inputs.

Conclusions:

• Convolutional dynamic networks effectively leverage contextual information for visual perception.
• The proposed model offers a robust solution for object recognition in challenging video conditions.
• This hierarchical approach provides stable and meaningful representations of visual data.