Generic fitting models learn edge representations from prenatal retinal waves

Area of Science:

• Computational Neuroscience
• Developmental Biology
• Vision Science

Background:

• Orientation selectivity is a fundamental feature of biological vision across diverse species.
• The developmental origins of orientation selectivity, whether genetic or experience-dependent, remain largely unknown.

Purpose of the Study:

• To investigate whether orientation selectivity emerges from experience-dependent fitting processes during prenatal development.
• To test if generic computational models can spontaneously develop orientation selectivity without innate priors.

Main Methods:

• Utilized generic image-computable fitting models (transformers) with no pre-existing orientation selectivity.
• Trained models using unsupervised temporal learning with biologically plausible prenatal data (retinal waves).
• Validated results across various model architectures, training conditions, and species-specific retinal waves.

Main Results:

• Models spontaneously developed robust orientation selectivity from scratch.
• The emergence of orientation selectivity was dependent on adaptation to prenatal visual input (retinal waves).
• Results were consistent across different model sizes and training parameters.

Conclusions:

• Orientation selectivity arises from domain-general fitting mechanisms adapting to prenatal experiences, not genetic predetermination.
• Supports experience-dependent theories of learning and neural development in the visual system.
• Suggests a unified mechanism for the development of edge representations across species.