Efficient coding of natural scenes improves neural system identification.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Computer Vision

Background:

• Neural system identification typically uses experimental data but often overlooks efficient coding principles.
• Visual systems are specialized for efficient processing of natural environmental stimuli.
• The efficient coding hypothesis suggests neural representations are shaped by information preservation under resource constraints.

Purpose of the Study:

• To develop and evaluate a normative network regularization for system identification models.
• To investigate if incorporating efficient coding via an autoencoder regularizer improves model performance and biological plausibility.
• To predict retinal neuron responses to stimuli using a hybrid model.

Main Methods:

• Developed a hybrid model combining system identification with an autoencoder regularizer based on efficient coding.
• Trained the model to predict retinal neuron responses to noise stimuli.
• Compared the hybrid model's performance and filter characteristics against a stand-alone system identification model.

Main Results:

• The hybrid model significantly outperformed the stand-alone system identification model in predicting neural responses.
• The regularized model generated more biologically plausible convolutional filters, resembling early visual system representations.
• Performance gains were observed across various artificial stimuli and model architectures, particularly for direction-of-motion sensitive neurons.

Conclusions:

• Normative regularization using efficient coding principles enhances neural system identification models, especially for noise stimuli.
• The findings support the benefit of efficient environmental input encoding for improving neural response prediction.
• Probing visual systems with naturalistic stimuli is recommended for a comprehensive understanding.