A deep convolutional visual encoding model of neuronal responses in the LGN

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• The Lateral Geniculate Nucleus (LGN) is a critical hub in the visual pathway, yet it remains less studied than the retina or visual cortex.
• Understanding LGN function is vital for developing visual prostheses targeting this area.

Purpose of the Study:

• To introduce a deep learning encoder for predicting neuronal firing in the LGN.
• To assess the model's performance in response to diverse visual stimuli and varying temporal windows.

Main Methods:

• A deep Convolutional Neural Network (CNN) was designed to integrate visual stimulus spatiotemporal information and LGN neuronal firing history.
• In vivo extracellular recordings from 150 single units in the rat LGN were performed using multi-electrode arrays.
• The model was trained using recorded neural activity and corresponding visual stimulation patterns (single-pixel, checkerboard, geometric shapes).

Main Results:

• The deep learning model achieved high prediction accuracy, with mean correlation coefficients of 0.57 (10 ms window) and 0.7 (50 ms window) between actual and predicted firing rates.
• The CNN model demonstrated robustness to neuronal spatiotemporal variability.
• Performance surpassed that of the Generalized Linear Model (GLM).

Conclusions:

• Deep convolutional neural networks show significant potential as accurate models for predicting LGN neuronal firing.
• This study highlights the efficacy of deep learning in deciphering complex neural processing in the visual system.