Purchase this article with an account.
Ming Bo Cai, Nicolas Schuck, Michael Anderson, Jonathan Pillow, Yael Niv; Should you trust your RSA result? A Bayesian method for reducing bias in neural representational similarity analysis.. Journal of Vision 2017;17(10):571. doi: 10.1167/17.10.571.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Representational similarity analysis (RSA) has become a popular tool in fMRI studies. It has recently been realized that calculating similarity between two neural patterns that were estimated during the same scanning session can induce bias in the estimated similarity matrix. However, the severity of this bias has not been fully appreciated. We analytically derive the source of the bias: serial correlations in fMRI noise, together with temporal relationships between task events, introduce structured noise in the estimated neural patterns. Correlation analysis of the estimated patterns translates the structured noise into spurious bias structure in the similarity matrix. The bias is especially severe with low signal-to-noise ratio and if experimental conditions cannot be fully randomized in the task design. For example, in an experiment in which task conditions had a fixed Markovian transition structure, 84±12% of the variance of the similarity matrix estimated from the OFC could be accounted for by this bias. We propose an alternative Bayesian framework for computing representational similarity, an extension of Diedrichsen et al., 2011. We treat the covariance structure of neural activity patterns as a hyper-parameter in a generative model of the neural data, and directly estimate this covariance structure from imaging data while marginalizing over the unknown activity patterns and parameters of autocorrelated noise. Converting the estimated covariance structure into a correlation matrix offers a much less biased estimate of neural representational similarity. Our method also simultaneously estimates a signal-to-noise map that informs where the learned representational structure is supported more strongly. In addition, the learned covariance matrix together with the SNR map can be used as a structured prior for the posterior estimation of neural activity patterns. The method also allows for learning a shared representational similarity structure across participants. Code is freely available in Brain Imaging Analysis Kit (BrainIAK, https://github.com/IntelPNI/brainiak).
Meeting abstract presented at VSS 2017
This PDF is available to Subscribers Only