Statistical inference on representational geometries.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neuroscience has advanced significantly, generating complex neural data and sophisticated brain-computational models.
• A gap exists in robust methods for evaluating these large-scale models against extensive experimental data.
• Connecting computational theory with empirical findings is crucial for progress.

Purpose of the Study:

• To introduce novel inference methods for evaluating and comparing computational models of the brain.
• To assess model performance based on the accuracy of predicting representational geometries in neural populations.
• To provide tools for robustly linking neuroscience theory with experimental big data.

Main Methods:

• Developed new inference techniques combining 2-factor cross-validation and bootstrapping.
• Cross-validation prevents overfitting to subjects or conditions, enhancing accuracy estimates.
• Bootstrapping allows inferential model comparison with simultaneous generalization to new subjects and conditions.

Main Results:

• Validated the inference methods using simulated data from deep neural networks.
• Confirmed method validity and generalizability with calcium-imaging and functional MRI data.
• Demonstrated that the methods accurately predict representational geometry distances.

Conclusions:

• The novel inference methods provide a robust framework for evaluating computational neuroscience models.
• These techniques facilitate the comparison of models based on their predictive accuracy of neural representations.
• An open-source Python toolbox (rsatoolbox) is available for applying these data analysis methods.