The feature-weighted receptive field: an interpretable encoding model for complex feature spaces.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Encoding models are crucial for understanding neural representations of sensory information.
• Classical receptive field models have limitations in expressiveness and scalability with complex stimuli.
• Feature maps offer a richer representation of visual stimuli than raw pixel data.

Purpose of the Study:

• Introduce the feature-weighted receptive field (fwRF) model.
• Develop a scalable and interpretable encoding model for visual neuroscience.
• Demonstrate fwRF's ability to analyze complex features, including those from deep neural networks.

Main Methods:

• The fwRF model separates spatial pooling ('where') from feature tuning ('what') parameters.
• Model complexity is independent of feature map resolution, allowing high-resolution feature analysis.
• An optimization algorithm estimates fwRF models from neuroimaging data.

Main Results:

• fwRF successfully recovers known visual cortex principles using Gabor features.
• fwRF achieves state-of-the-art prediction accuracy when regressing deep convolutional neural networks against brain activity.
• The model allows direct read-off of spatial pooling and feature tuning without extensive post-processing.

Conclusions:

• The fwRF model offers a powerful and flexible tool for visual encoding and decoding.
• It bridges the gap between traditional neuroscience models and modern deep learning approaches.
• fwRF has broad applications, from retinotopic mapping to analyzing complex neural network representations.