Accuracy optimized neural networks do not effectively model optic flow tuning in brain area MSTd.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• Accuracy-optimized convolutional neural networks (CNNs) excel at modeling neural responses in the primate ventral stream.
• Their effectiveness in modeling neurons in the complementary primate dorsal stream, specifically area MSTd, remains largely unexplored.

Purpose of the Study:

• To evaluate how well CNNs model the optic flow tuning properties of neurons in the dorsal area MSTd.
• To compare CNN performance with the Non-Negative Matrix Factorization (NNMF) model, known for modeling MSTd neurons.
• To investigate the computational properties of NNMF, such as non-negative weights and sparse coding, by creating CNN variants with these constraints.

Main Methods:

• Explored CNNs' ability to model optic flow tuning in primate dorsal area MSTd.
• Compared CNNs with the Non-Negative Matrix Factorization (NNMF) model.
• Developed CNN variants incorporating NNMF constraints (non-negative weights, sparse coding) to understand their impact on optic flow tuning.

Main Results:

• Both CNNs and NNMF accurately estimated self-motion from simple (translational or rotational) optic flow.
• NNMF and CNNs with non-negative weights showed significantly lower accuracy on complex optic flow combining translation and rotation.
• NNMF exhibited tuning properties more aligned with primate MSTd neurons than accuracy-optimized CNNs.

Conclusions:

• While accuracy-optimized CNNs are effective for the ventral stream, they do not fully capture the optic flow tuning of dorsal area MSTd neurons.
• NNMF, despite lower accuracy on complex stimuli, better reflects the observed tuning properties in MSTd.
• This study advances understanding of the computational properties and constraints underlying optic flow tuning in primate area MSTd.