Large-scale parameters framework with large convolutional kernel for encoding visual fMRI activity information

Area of Science:

• Neuroscience
• Computer Vision
• Machine Learning

Background:

• Deep neural networks are used in visual encoding models to interpret brain responses to stimuli.
• Large receptive fields, created with large convolutional kernels, have shown to boost the performance of convolutional encoding models.

Purpose of the Study:

• To investigate the performance of large convolutional kernel encoding models at larger parameter scales.
• To propose a large-scale parameter framework utilizing sizeable convolutional kernels for encoding visual functional magnetic resonance imaging (fMRI) activity.

Main Methods:

• A large-kernel convolutional network was employed for stimulus image feature extraction, with increased channel numbers to expand parameter size.
• A multi-subject fusion module was used to enlarge input data during training, accommodating the increased parameters.
• A voxel mapping module was developed to translate stimulus image features into fMRI signals.

Main Results:

• The proposed framework demonstrated an approximate 7% improvement in performance on the Natural Scenes Dataset compared to base-scale models.
• Analysis revealed a trade-off between encoding performance and trainability within the framework.
• Expanding parameters in visual coding was confirmed to yield performance enhancements.

Conclusions:

• The study validates that increasing parameters in visual encoding models, particularly with large convolutional kernels, leads to improved performance.
• The proposed large-scale framework offers a promising approach for encoding visual fMRI data.
• Further research can explore optimizing the balance between performance and trainability in these models.