Using machine learning to understand cognition from brain images

Brain imaging provides increasingly rich and complex measures of neural activity. Making full use of this data within the framework of conventional statistics is challenging as it requires formulating clear and simple hypotheses to test. Progress in machine learning and data mining enables fitting detailed models. However these are geared towards prediction and our inference framework must be revisited to use them to draw conclusions on brain mechanisms [Varoquaux & Thirion 2014]. I will present a few progresses in this endeavor.

First, I will show how "encoding" models, that predict brain activity from the perception and action of the subject can uncover the features of our world that are naturally reflected in the brain. I will explore a detailed example of such encoding models that outline a high-resolved decomposition of the human visual system from natural stimuli [Eickenberg, submitted].

What is the cognitive function of brain regions outlined in such an encoding model, or in any standard cognitive neuroimaging experiment? These brain regions are defined as activating during a certain task or for certain features of the stimuli. Cognitive studies would like draw the converse conclusion: that activity in the brain regions implies the behavior. Without careful analysis, such conclusion is a fallacy from a logical standpoint: a reverse inference.

A first methodological tool the ground reverse inferences is decoding models: predictive models deducing the task from the brain observed brain images. Here, machine learning is not only used for its prediction, but also in an inverse problem settings, to recover discriminant regions. Given the scarcity of data and the dimensionality of the brain images, the problem is very ill-posed and requires injecting various priors. I will summarize five years of research in developing and validating regularizations that encode efficiently priors well suited for neuroimaging, in particular to segment and outline brain regions. Specifically, these draw ideas from total-variation [Michel TMI 2011] or randomized-clustering [Varoquaux ICML 2012] approaches.

To come to general conclusions on the function of a brain region, it is also necessary to explore a vast repertoire of cognitive function. Doing so requires joint analysis of many studies. Such an analysis is very valuable, as it enables going beyond the idiosyncrasies of a study, but it requires describing a huge variety tasks with a common vocabulary. On 30 different fMRI studies, we describe experiment conditions with 20 different experimental-psychology notions [Schwartz NIPS 2013]. Using a predictive model that leverages contrasts between conditions with similar stimuli, we can describe a completely new experiment with these notions. Importantly, prediction is perform on unseen studies, thus validating that the decoder has learned maps specific to the notion of interest, and not to the experimental paradigm. A precious outcome is a "reverse inference" atlas of the predictive regions for these notions.

[Varoquaux & Thirion 2014]: How machine learning is shaping cognitive neuroimaging http://www.gigasciencejournal.com/content/3/1/28

[Michel TMI 2011] Total variation regularization for fMRI-based prediction of behavior https://hal.inria.fr/inria-00563468

[Varoquaux ICML 2012] Small-sample brain mapping: sparse recovery on spatially correlated designs with randomization and clustering https://hal.inria.fr/hal-00705192

[Schwartz NIPS 2013] Mapping cognitive ontologies to and from the brain https://hal.inria.fr/hal-00904763v2