Testing the generalization of neural representations

Multivariate analysis methods are widely used in neuroscience to investigate the presence and structure of neural representations. Representational similarities across time or contexts are often investigated using pattern generalization, e.g. by training and testing multivariate decoders in different contexts, or by comparable pattern-based encoding methods. It is however unclear what conclusions can be validly drawn on the underlying neural representations when significant pattern generalization is found in mass signals such as LFP, EEG, MEG, or fMRI. Using simulations, we show how signal mixing and dependencies between measurements can drive significant pattern generalization even though the true underlying representations are orthogonal. We suggest that, using an accurate estimate of the expected pattern generalization given identical representations, it is nonetheless possible to test meaningful hypotheses about the generalization of neural representations. We offer such an estimate of the expected magnitude of pattern generalization and demonstrate how this measure can be used to assess the similarity and differences of neural representations across time and contexts.