What behaviourism can (and cannot) tell us about brain imaging

Robert Shulman belongs to a rare breed. On the one hand, he is a true pioneer of MRI, a technology that has galvanised cognitive neuroscience by allowing researchers to trace the flux of metabolic activity in the living brain under diverse experimental conditions. On the other, he is a staunch behaviourist who believes that terms like ‘attention’ and ‘working memory’ should be expunged from science – though these are the very terms held precious in the vernacular of cognitive neuroscience. Shulman has even published important work on the neuroimaging of consciousness, the one topic guaranteed to annoy any cardcarrying behaviourist. So it is not surprising that his new book, Brain Imaging: What it Can (and Cannot) Tell Us About Consciousness, is by turns fascinating, enlightening, provocative, and frustrating. Thankfully, it is also extremely well written, combining clarity, erudition, and authority across a range of topics in neuroimaging and the philosophy of science. At the same time, the narrative includes a beguiling personal perspective on Shulman’s own development as a scientist in love with the humanities, all in the brief compass of 160 pages. A central theme throughout the book is that popular applications of methods like functional MRI (fMRI) may be more limited than is commonly assumed. The strongest parts of this argument arrive in the book’s later chapters where Shulman describes his important work linking measures of brain metabolism (which underlie fMRI) to neuronal signalling and emphasises that fluctuations in these measures reflect only a small fraction of overall activity. Shulman argues, quite rightly, that this ongoing activity must be functionally relevant and, indeed, this view has been gaining substantial traction as neuroimaging studies of the default-mode network, resting-state networks, and intrinsic activity proliferate. From this worthwhile position, Shulman criticises the subtractive method common in many neuroimaging experiments, which involves identifying regional differences in metabolic signals between experimental conditions (e.g., low versus high memory load) while controlling for as many other factors as possible. Statistically significant regional differences (identified by the good-old general linear model) are then used to generate the multicoloured brain maps we both love and hate. Shulman’s main critique