1Selective attention is fundamental to cognitive activity and can be deployed in 2 different ways. Non-human primate data suggests that spatial and feature-based 3 visual attention have qualitatively different effects on neural tuning, but this has 4 been challenging to assess in humans. Using multivariate decoding of MEG data, 5we tracked the effects of spatial and feature-selective attention on population-level 6 coding of novel objects. We found that spatial and feature-selective attention 7 interacted multiplicatively to enhance object representation. Moreover, the two 8 types of attention induced qualitatively different patterns of enhancement in 9 occipital cortex, and these differences were accounted for by the principles of 10 response-gain and tuning curve sharpening derived from single-unit work. A novel 11 information flow analysis further showed that stimulus representations in occipital 12 cortex were Granger-caused by coding in frontal cortices earlier in time. We find 13 that human spatial and feature-selective attention rely on qualitatively different, 14interacting, neural mechanisms. 15At any moment, there is far more information available from our senses than we can 16 possibly process at once. Accordingly, only a subset of the available information is 17 processed to a high level, making it crucial that brain can dynamically devote greatest 18 processing resources to the most relevant information. Our ability to selectively attend 19 to relevant information is remarkably flexible. For instance, we can adapt our 20 attentional state by directing our attention in space (spatial attention, e.g. attend left), 21 to a specific feature dimension (feature-selective attention, e.g. detect changes in color 22 across a scene) or based on a particular feature value along that feature dimension 23 (feature-based attention, e.g. find all the red objects), using the definitions of Chen et 24 al. (2012). Each of these types of attention can change behavior, improving 25 performance related to the attended location or feature-dimension, while decreasing 26 performance on the ignored dimension/location (Pestilli and Carrasco, 2005; Rossi and

21 22Multivariate decoding methods applied to neuroimaging data have become the standard in 23 cognitive neuroscience for unravelling statistical dependencies between brain activation patterns 24 and experimental conditions. The current challenge is to demonstrate that information decoded 25 as such by the experimenter is in fact used by the brain itself to guide behaviour. Here we 26 demonstrate a promising approach to do so in the context of neural activation during object 27 perception and categorisation behaviour. We first localised decodable information about visual 28 objects in the human brain using a spatially-unbiased multivariate decoding analysis. We then 29 related brain activation patterns to behaviour using a machine-learning based extension of signal 30 detection theory. We show that while there is decodable information about visual category 31 throughout the visual brain, only a subset of those representations predicted categorisation 32 behaviour, located mainly in anterior ventral temporal cortex. Our results have important 33implications for the interpretation of neuroimaging studies, highlight the importance of relating 34 decoding results to behaviour, and suggest a suitable methodology towards this aim. 35 36 not peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission.The copyright holder for this preprint (which was . http://dx.doi.org/10.1101/248583 doi: bioRxiv preprint first posted online Jan. 16, 2018; 3 Significance statement 37 38Brain decoding methods are a powerful way to analyse neuroimaging data. An implicit assumption 39 in many decoding studies is that when information can be decoded, then the brain is using this 40 information for behaviour. However, this assumption must be explicitly tested. Here, using visual 41 object categorisation as an example, we separately localised decodable information and 42 information that can be used to predict behaviour. Our findings showed that only from a subset of 43 areas that had decodable category information, we could predict observer categorisation reaction 44 times. Our results highlight the distinction between decodable information and information that is 45 suitably formatted for read-out by the brain in behaviour. Our results have critical implications for 46 the interpretation of decoding studies in general. 47 48 not peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission. 1-7). An implicit assumption often made by 52 the experimenters conducting such analyses is that if information can be decoded, then this 53 information is available to the brain to use in behaviour (8, 9). However, this does not have to be 54 the case, as it could be that the decoded information is different from the signal that is relevant 55 for the brain (8, 10). For example, the decodable information could be epiphenomenal, or the 56 information could be used for a different behaviour than stipulated (cf. 11). Thus, a current 57 challenge in cognitive neuroscience is to ...