Beyond Divisive Normalization: Scalable Feedforward Networks for Multisensory Integration Across Reference Frames.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Multisensory integration is crucial for perception and action, involving reference frame transformations.
• Bayesian inference models multisensory integration, but neural mechanisms remain debated.
• Divisive normalization is a proposed mechanism, yet its brain implementation is unclear and models struggle with scalability.

Purpose of the Study:

• To propose an alternative model for multisensory integration that approximates Bayesian inference.
• To investigate if feed-forward neural networks can achieve multisensory integration without explicit divisive operations.
• To challenge the necessity of divisive normalization in neural computations for multisensory integration.

Main Methods:

• Developed a multilayer-feedforward neural network model for multisensory integration (MSI).
• Trained the network on the analytical Bayesian solution for multisensory integration across different reference frames.
• Evaluated the model's ability to replicate empirical principles of multisensory integration and neural activity in Ventral Intraparietal (VIP) neurons.

Main Results:

• The proposed feed-forward network successfully approximates Bayesian inference for multisensory integration.
• The model exhibits empirical principles of multisensory integration and mimics behavior observed in VIP neurons.
• Achieved multisensory integration across reference frames without requiring explicit divisive normalization or specific connectivity structures.

Conclusions:

• Simple feed-forward networks with additive units can approximate optimal Bayesian inference for multisensory integration.
• Explicit divisive normalization may not be necessary for the brain to perform multisensory integration.
• This work provides insights into neural computations underlying multisensory processing and challenges existing models.