Dynamic Construction of Reduced Representations in the Brain for Perceptual Decision Behavior

Abstract

Current models propose that the brain uses a multi-layered architecture to reduce the high dimensional visual input to lower dimensional representations that support face, object and scene categorizations. However, understanding the brain mechanisms that support such information reduction for behavior remains challenging. We addressed the challenge using a novel information theoretic framework that quantifies the relationships between three key variables: single-trial information randomly sampled from an ambiguous scene, source-space MEG responses and perceptual decision behaviors. In each observer, behavioral analysis revealed the scene features that subtend their decisions. Independent source space analyses revealed the flow of these and other features in cortical activity. We show where (at the junction between occipital cortex and ventral regions), when (up until 170 ms post stimulus) and how (by separating task-relevant and irrelevant features) brain regions reduce the high-dimensional scene to construct task-relevant feature representations in the right fusiform gyrus that support decisions. Our results inform the occipito-temporal pathway mechanisms that reduce and select information to produce behavior.. CC-BY-NC-ND 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.The copyright holder for this preprint . http://dx.doi.org/10.1101/284158 doi: bioRxiv preprint first posted online Mar. 18, 2018; 3 Over the past decade, there has been extensive studies of the regions of the brain that support face, object and scene recognition using different methodologies and modalities of brain measurements (e.g. electrophysiology, E/MEG and fMRI), including across species. A converging set of results now suggests a hierarchically organized architecture of brain regions, spanning the occipital and temporal lobes (1-17), where categorizations unfold over the first few hundred milliseconds of post stimulus processing (18)(19)(20)(21)(22). This same architecture is flexibly involved in multiple categorization tasks that require multiple representational bases. With extensive knowledge of the where and when, the next challenge is to unravel the thorny how.That is, how detailed information processing mechanisms in the occipitoventral pathway dynamically implement flexible visual categorization by selecting, from the high-dimensional input, the low-dimensional information basis required for behavior?In other words, how does our brain extract, from the features of the visual scene that it initially represents, those that are actually useful for the categorization task?To start addressing this considerable challenge, we used Dali's ambiguous painting Slave Market with Disappearing Bust of Voltaire (see Figure1, Stimulus) as a case-study stimulus because it is a complex visual scene that affords two distinct perceptions. We applied the Bubbles technique (Gosselin & Schyns, 20001) to character...

Modelling face memory reveals task-generalizable representations

“…As shown by Chang et al [36], neurons selectively respond along a single axis of the face space, not to other, orthogonal axes. An interesting direction of research is to determine whether our reduced diagnostic features, as defined by our 'abstract information goal' (see also [37]), provide a superior fit to the neural data than the full feature sets used in the axis model used by Chang et al [36] .…”

“…as representation in multi-scale, multi-orientation Gabor-like, retinotopically mapped receptive fields [38,39]; at intermediate levels of processing, as the sort of local surface patches [40,41] that we reveal, and at the top level as the combinations of surface patches that enable identification and resemblance responses. Under a framework of top-down prediction [42,43], the abstract information goal of a familiar face identity should trim, in a top-down manner, the fully-mapped but redundant information on the retina into the task-relevant features that are transferred along the occipital to ventral/dorsal visual hierarchy [37]. Tracing the construction of such a reduced memory representation of face identity in the brain should enable an accurate and detailed modelling of the processing mechanism along the visual hierarchy (see also [12,[44][45][46]).…”

Disentangling presentation and processing times in the brain

“…Despite the same experimental stimuli, this gating phenomenon is only seen in the gender task, suggesting that it is specific to lateralized task-relevant features (the eyes being used almost exclusively for the gender task). In a recent study, the N170 also appeared to filter out task-irrelevant features: while both task-relevant and task-irrelevant features were processed prior to 170 ms, only taskrelevant features were processed afterwards 37 . Of note, the cause of this gating cannot be repetition priming because it is also visible in trials where only one feature is revealed once.…”

Making thebrain-activity-to-informationleap using a novel framework: Stimulus Information Representation (SIR)

“…Applying SIR to understand visual information processing SIR has been recently applied to analyze brain imaging data collected during the performance of a visual categorization task 22 . This study illustrates how the SIR framework can be used to disentangle brain activity into the processing of task-relevant stimulus information, the processing of task-irrelevant stimulus information, and other brain processes, to enrich the interpretation of brain imaging data.…”

“…Zhan et al 22 used these three SIR components to trace the dynamic flow of taskrelevant features that were processed between 50 and 220 ms post stimulus, from their early representation in the visual cortex, through the ventral pathway. In the ventral pathway, we found that task-relevant features converge onto a few MEG voxels at the top of the right fusiform gyrus, ~200 ms post-stimulus, where they agglomerate into distinct representations that support each behavioral decision (see "task-relevant" images in Figure 2C).…”