Performance-optimized hierarchical models predict neural responses in higher visual cortex.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• The ventral visual stream is essential for human visual object recognition.
• Neural encoding in higher ventral stream areas, like the inferior temporal (IT) cortex, is not well understood.
• Hierarchical neural network models are increasingly used to study visual processing.

Purpose of the Study:

• To develop a quantitatively accurate computational model of the inferior temporal (IT) cortex.
• To investigate the relationship between categorization performance and neural encoding in IT cortex.
• To predict neural responses in the ventral visual stream using artificial neural networks.

Main Methods:

• Utilized high-throughput computational techniques to model the IT cortex.
• Developed and evaluated biologically plausible hierarchical neural network models.
• Correlated model categorization performance with the ability to predict IT neural unit responses.
• Identified a high-performing neural network matching human recognition task performance.

Main Results:

• A strong correlation was found between model categorization performance and prediction of IT neural unit responses.
• The top output layer of a high-performing neural network accurately predicted IT spiking responses to naturalistic images.
• Intermediate layers of the model predicted neural responses in the V4 cortex.
• Performance optimization in biologically appropriate models can yield quantitative predictive models of neural processing.

Conclusions:

• Computational models, when optimized for performance within a biologically plausible framework, can accurately predict neural processing in the ventral visual stream.
• This approach offers a powerful tool for understanding neural encoding in higher visual areas.
• The study provides quantitative insights into the functional organization of the IT cortex and its relationship with earlier visual areas like V4.