Feature-Specific Denoising of Neural Activity for Natural Image Identification

Decoding the content in neural activity through voxelwise encoding plays an important role in investigating cognitive functions of the human brain. However, unlike multivoxel pattern analysis (MVPA), voxelwise encoding builds a model for each individual voxel; therefore, ignores the interactions between voxels and is sensitive to noise. In this work, we propose the feature-specific denoise (FSdenoise), a noise reduction method for encoding-based models to improve their decoding performance. FSdenoise considers the response of a voxel to a stimulus as a combination of two components: 1) feature-relevant component, which can be predicted from stimulus features and 2) feature-irrelevant component, which shows no direct relation to the concerned features. Exploiting the correlations between voxels, FSdenoise reduces the feature-irrelevant component in voxels that exhibit more feature-relevant component, enhancing their predictive power from stimulus features. Decoding performance with the denoised voxels would be improved in consequence. We validate the FSdenoise on two functional magnetic resonance imaging data sets and the results demonstrate that FSdenoise can efficiently improve the decoding accuracy for encoding-based approaches. Moreover, the encoding-based approaches combined with FSdenoise can even outperform the MVPA-based approach in brain decoding.