Testing the generalization of neural representations.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Neuroscience

Background:

• Multivariate analysis is key for understanding neural representations.
• Pattern generalization assesses representational similarity across contexts but its interpretation in mass signals is debated.

Purpose of the Study:

• To investigate the validity of pattern generalization findings in mass neuroimaging signals.
• To develop a method for accurately assessing neural representations despite signal confounds.

Main Methods:

• Utilized simulations to model signal mixing and measurement dependencies in neural data.
• Developed a method to estimate expected pattern generalization for orthogonal representations.
• Applied the method to assess similarity of neural representations across time and contexts.

Main Results:

• Signal mixing and measurement dependencies can spuriously inflate pattern generalization.
• Identified conditions where orthogonal representations yield significant generalization.
• The proposed method provides a reliable estimate of expected pattern generalization.

Conclusions:

• Standard pattern generalization in mass signals (LFP, EEG, MEG, fMRI) may not reflect true representational similarity.
• The developed method allows for robust hypothesis testing on neural representation generalization.
• Accurate estimation of expected generalization is crucial for validly interpreting representational similarity across contexts.