High-level visual processing in the lateral geniculate nucleus revealed using goal-driven deep learning.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Computer Vision

Background:

• The Lateral Geniculate Nucleus (LGN) is crucial for visual processing, yet its role in high-level vision is under-explored by current models.
• Existing computational models of the LGN primarily focus on basic visual properties, neglecting its contribution to complex visual perception.

Purpose of the Study:

• To develop a high-level computational approach for encoding mouse LGN neural responses to natural scenes.
• To utilize deep neural networks (DNNs) as goal-driven models to investigate visual features represented in the LGN.

Main Methods:

• Employed VGG16 and ResNet50 deep neural networks (DNNs) as goal-driven models.
• Analyzed the predictive power of different DNN layers for LGN neural activity.
• Investigated the encoding of numerosity, a high-level visual feature, within the LGN.

Main Results:

• Early DNN layers effectively model basic LGN responses, while intermediate layers better capture complex features like numerosity.
• LGN neural activity encodes numerosity alongside other visual features.
• An ensemble model combining early and intermediate DNN layers achieved high accuracy in neural prediction and numerosity representation.

Conclusions:

• Deep neural networks (DNNs) offer a powerful tool for understanding high-level feature representation in the LGN.
• The findings challenge the traditional view of the LGN as a simple visual relay, underscoring its role in sophisticated visual processing.
• Goal-driven DNNs serve as effective high-level vision models for neural prediction and exploration of the LGN's function.