Abstract Choice Representations Generalize Between Task Contexts

Choice Representations Generalize Between Task Contexts Florian Sandhaeger (florian.sandhaeger@uni-tuebingen.de) Hertie-Institute for Clinical Brain Research, Centre for Integrative Neuroscience & MEG Center University of Tuebingen, Otfried-Mueller-Str. 25, 72076 Tuebingen, Germany Nina Omejc (nina.omejc@student.uni-tuebingen.de) Hertie-Institute for Clinical Brain Research, Centre for Integrative Neuroscience & MEG Center, University of Tuebingen, Otfried-Mueller-Str. 25, 72076 Tuebingen, Germany Anna-Antonia Pape (anna.antonia.pape@gmail.com) Centre for Integrative Neuroscience & MEG Center University of Tuebingen, Otfried-Mueller-Str. 25, 72076 Tuebingen, Germany Markus Siegel (markus.siegel@uni-tuebingen.de) Hertie-Institute for Clinical Brain Research, Centre for Integrative Neuroscience & MEG Center, University of Tuebingen, Otfried-Mueller-Str. 25, 72076 Tuebingen, Germany


Introduction
Sensory decisions are often linked to an appropriate motor action. This has led to a framework of choices being represented as action intentions (Cisek & Kalaska, 2010). In addition, humans and other primates also make abstract choices when a suitable action is not known in advance. Task designs in which the mapping of choices to responses is only revealed after stimulus presentation have revealed neuronal representations of abstract choices in non-human primates and humans (Bennur & Gold, 2011;Horwitz, Batista & Newsome, 2004;Merten & Nieder, 2012;Hebart, Donner & Haynes, 2012;Ludwig, Herding & Blankenburg, 2018). It is unclear if these abstract choice-representations generalize to choices that are linked to actions. While some recent evidence suggests choice representations in response-modality specific motor and premotor areas only when motor actions can be planned (Herding, Ludwig & Blankenburg, 2017;Ludwig, Herding & Blankenburg, 2018;Wang et al., 2019), other studies have shown choice signals in oculomotor areas before the choice-response mapping was specified (Bennur & Gold, 2011;Horwitz, Batista & Newsome, 2004).
We used magnetoencephalography (MEG) in human participants and employed multivariate decoding analyses to address two questions: First, are there motor-and stimulus-independent choice signatures in human MEG during both, presence and absence of a known motor mapping? Second, do abstract choice representations generalize between these two decision-making contexts?

Methods
We measured MEG in 33 healthy participants while they performed variants of a sensorimotor decision-making task (Fig. 1). Participants viewed random dot kinematograms and decided whether these contained coherent motion or not. In each trial, a cue indicated the mapping of a particular choice to a left-or right-hand button press either before (pre-cue -condition) or after the stimulus (post-cue -condition). Stimuli were presented near psychophysical threshold, allowing us to disentangle physical stimulation from perceptual choice and motor response.
Using multivariate decoding (cross-validated MANOVA; Allefeld & Haynes, 2014), we assessed the time-courses of stimulus-, choice-, mapping-, and response information. To estimate the cortical sources of information, we performed the decoding analysis in a searchlight-fashion. We employed cross-condition decoding to investigate whether choice-representations generalized between contexts. We assessed statistical significance using non-parametric cluster permutation tests. We tested the magnitude of cross-decoding against an estimate of cross-decoding under the assumption of identical patterns (similar to Spaak et al., 2017).

Results
We found that all relevant task variables were decodable from MEG (Fig. 2). Crucially, we found information about participants' behavioral choices in both conditions, independent of stimulus identity and motor response. Source level analysis revealed that the cortical distribution of stimulus information was centered on early visual areas, while response information was present most strongly in motor areas. Choice information showed no distinct spatial peaks but was available throughout cortex in both conditions. We then directly compared choice representations with known and unknown motor mappings using crossdecoding. There was significant cross-decoding with a time-course similar to the within-condition choice information. Importantly, empirical cross-decoding values were statistically indistinguishable from the expected cross-decoding for identical choice representations in pre-and post-cue conditions.

Conclusions
Our results suggest the presence of abstract choice representations in the human brain also for choices linked to actions. Within the limits of MEG, these representations seem identical in both contexts, raising the question whether the formation of abstract choices may be a general process in sensorimotor decision making.