A Bayesian method for reducing bias in neural representational similarity analysis