Cognitive neuroscience assumes a correspondence between specific spatio-temporal patterns of neural activity and the states of a mechanism that processes cognitive information. Mechanistic explanations of cognition should therefore translate patterns of neural activity into the components of a formal mechanism: a set of information processing states and their transitions. For the first time, we carried out this research programme with four naive observers instructed to categorise randomly presented face information. With classification image techniques, we revealed the diagnostic features that the brain requires to produce correct behaviour (i.e., two eyes for gender categorisation in one session; the mouth for expression in the other session). With the same techniques applied to brain signals, we revealed the features processing states associated with modulations of oscillatory EEG energy (measured on occipito-temporal face-sensitive electrodes). Here we show how transitions between distinct feature processing states in the theta/alpha [4-12 Hz] oscillatory bands implement two face categorisations. On the left and right occipito-temporal electrodes of each observer, processing of the contra-lateral eye precedes bilateral integration of the features required for behaviour. For the first time, we relate stimulus information to behaviour via sequences of categorisation-specific feature processing states in the brain. © 2007 Elsevier Inc. All rights reserved.To study face categorisation mechanisms in an information system such as the brain, automata theory (Hopcroft and Ullman, 1979) specifies three generic questions that must be addressed to relate brain activity to the states of an information processing mechanism (i.e., a formal automaton). The first question is that of form: what is the nature of the brain activity supporting face processing-i.e., the states of the brain correlated with face processing? The second question is that of content: what is the information content processed in these brain states? The third question is that of transition: how does information flow from one brain state to the next between stimulus onset and behavioural response? So far, cognitive neuroimaging studies have not simultaneously addressed these three questions. Until this is resolved, inferences about cognitive information processing mechanisms will be severely restricted.To illustrate the problem, researchers have proposed that slow brain rhythms could form distinct states of information processing, on the basis of considerable evidence of a consistent relationship between low frequency, theta (4-8 Hz) and alpha (8-12 Hz), oscillatory EEG activity of the neural substrate and behavioural variables (Van Rullen and Koch, 2003;Ward, 2003, for reviews). However, definitive evidence is lacking in part because it has been proven difficult to identify precisely the information content of these states from the parametric modulations (i.e., phase and amplitude) of slow brain rhythms. Consequently, little is known about the detailed f...