Abstract

The practice of reverse inference--inferring that a subject is engaged in a cognitive process given a particular pattern of brain activity--is widely viewed as unprincipled,  inaccurate, and ineffective. Here, we show it can be principled, accurate, and effective. Given an principled model of brain activity and an ontology of cognitive functions, reverse inference can be accurate and can progress our understanding of the brain. Here, we develop a method for reverse inference--jointly and probabilistically--based on the intuition that task evoked brain activity is the result of the activity of multiple cognitive functions, each of which is subserved by a network of brain regions, some of which overlap across cognitive functions. We show that individual frames of brain activity during tasks can be captured with this model, and the task the subject is in can be decoded very accurately. Moreover, we generated two new insights into the brain by using this reverse inference method. First, we replicate a previous meta-data finding at a high temporal and spatial resolution in individual subjects. Second, we find that subjects' task evoked activity can be clustered into states, each of which correspond to a single resting-state network or a combination of resting-state networks. While states are rather similar across tasks, each task has one to two unique states.

Introduction

Two influential articles by Mark Desposito \cite{D_Esposito_1998} and Russ Poldrack \cite{POLDRACK_2006} cautioned against the use of reverse inference — the practice of inferring that a cognitive function is engaged based on observing an activated brain region previously evoked by tasks that presumably probe that cognitive function. This inference is only deductively valid if the brain region is involved in only that cognitive function, and the original task (contrast) engaged only that cognitive function. For example, consider the language example from Russ Poldrack's influential paper \cite{POLDRACK_2006}. If we observe activation in the inferior frontal gyrus, does this indicate language processing? In other words, we want to know the probability of language processing given activation in the inferior frontal gyrus. To compute this probability, Poldrack used Bayes rule (Figure 1) and analyzed brain activity maps in the BrainMap database. There are 166 language studies that activate the inferior frontal gyrus and 703 language studies that do not activate the inferior frontal gyrus. Therefore, under this method, the probability that language studies activate the gyrus is 0.19 (166 divided by 166 plus 703. However, there are 199 non-language studies that activate the inferior frontal gyrus and 2154 studies that do not activate the gyrus. Therefore the probability of non-language studies activating the gyrus is 0.08 (199 divided by 199 plus 2154). Considering these non-language studies is critical. If non-language studies consistently activate the inferior frontal gyrus, then the gyrus is not specific to language processing, but perhaps some other cognitive function. Finally, we also have to set the prior probabilist of langue processing being engaged. If we assume a neutral probability of 0.5 \cite{Yarkoni_2011}, this gives us a 22% percent probability of language processing given inferior frontal gyrus activity.