Content area
Full Text
Reverse inference consists in inferring that a task recruits a psychological process (P) on the grounds that a brain structure (S) is activated during this task (as observed by, e.g., fMRI). It is often assumed that reverse inference is valid only if activation is selective, that is, if the ratio P(activation of S/P is recruited)/P(activation of S/P is not recruited) is high (Poldrack 2006). Because brain areas are typically multifunctional, cognitive neuroscientists have grown skeptical of area-based reverse inference. Anderson endorses this pessimistic conclusion – “It should go without saying that we must also curtail the common practice of reverse inference” (Anderson 2014, p. 113) – and the first two chapters of After Phrenology (Anderson 2014) extensively review the multifunctionality, hence low selectivity, of brain regions.
One can address the problem raised by multifunctionality in three different ways. First, reverse inference can be reformulated to depend on diagnosticity instead of selectivity (Machery 2014). In this approach, reverse inference is valid only if the activation discriminates between the recruitment of a first psychological process, P, and of a second psychological process, P′, that is, only if the ratio P(activation of S/P is recruited)/P(activation of S/P' is recruited) is high. Second, one can increase the selectivity of brain activation by revising cognitive neuroscientists' brain ontology: Instead of focusing on regional activation, one can reverse infer on the basis of activation in other brain structures (e.g., networks) that may be selectively associated with psychological processes (e.g., Glymour & Hanson, forthcoming). In chapter 4 of After Phrenology, Anderson rejects this second approach on the grounds that brain networks too can be multifunctional. Anderson's concern here is speculative, and more evidence is needed before discrediting brain ontology revision. Large-scale brain networks (e.g., effective connectivity networks), or activation patterns within those networks (e.g., as measured by MVPA), may be far more selective or diagnostic than individual regions. Third, one can increase the selectivity of brain activation by revising cognitive neuroscientists' cognitive ontology: On this approach, activation of brain structures is not selective because cognitive neuroscientists lack the right set of cognitive constructs...