CNS*2026 Halifax

Introduction

Many decisions require combining multiple evidence streams into a single action. In the human double-decision random-dot task, participants view moving colored dots and report a single choice among four spatial targets that jointly encode motion direction (left/right) and dominant color (blue/yellow) [1].  Motion and color coherences jointly determine the correct target. We show that a single LIP-inspired recurrent circuit reproduces the key error-rate (ER) and reaction-time (RT) signatures of both 2T trials (two targets; motion only) and 4T trials (four targets; motion+color), including similar ERs but longer RTs in 4T compared with 2T. We then ask how adding a second stream reshapes the decision-manifold geometry of population dynamics.



Methods
We used an E/I neural-field model of LIP with distance-dependent connectivity, extending the 2-target circuit (2T) of Monsalve-Mercado et al. [2] to four targets (4T) with four target-in (Tin) populations. Each target received a Gaussian input bump whose amplitude scaled with stimulus coherence. Tin activity bumps competed via shared broad inhibition (winner-take-all). Motion and color were independent noisy evidence streams with separate gains. Stimulus drive was maintained throughout the entire decision process. ER was the fraction of trials in which the correct Tin won. RT was defined as the first time the gap between the largest and second-largest Tin activities exceeded a fixed value. The decision manifold is reproduced via PCA.

Results

Our 4T network qualitatively captured the dependence of behavioral error rate (ER) and reaction time (RT) on motion and color coherence in the double-decision task [1]: low coherences yielded higher ERs and longer RTs, whereas high coherences produced lower ERs and shorter RTs. Motion outperformed color in the data and was captured by a higher motion-input gain. Model RTs required an additive 0.4 s offset consistent with non-decision time. Comparing matched 2T and 4T conditions, ERs were similar but RTs were consistently longer in 4T, consistent with behavioral results. Population activity showed a richer decision-manifold geometry in 4T, with participation ratio increasing nonlinearly from just above 2 in 2T to around 6 in 4T.



Discussion

A single LIP-inspired E/I neural field with local excitation and broad inhibition can account for core signatures of human double decisions [1]. In the model, four Tin activity bumps compete within a shared inhibitory pool, naturally producing a reaction-time cost in 4T without a comparable loss in accuracy, consistent with parallel evidence acquisition but a serial bottleneck in commitment. Adding a second evidence stream also reshapes state-space structure: effective dimensionality increases markedly from 2T to 4T, and low-coherence trials dwell longer near a quasi-indecision region before diverging toward the eventual choice state. This circuit provides a mechanistic bridge between multi-attribute behavior and decision-manifold geometry.


References

1 - Kang, Y. H., Löffler, A., Jeurissen, D., Zylberberg, A., Wolpert, D. M., & Shadlen, M. N. (2021). Multiple decisions about one object involve parallel sensory acquisition but time-multiplexed evidence incorporation. eLife, 10, e63721. https://doi.org/10.7554/eLife.63721
2 - Monsalve-Mercado, M. M., Stine, G. M., Shadlen, M. N., & Miller, K. D. (2025). The geometry of the neural state space of decisions. bioRxiv. https://doi.org/10.1101/2025.01.24.634806


Acknowledgement
I thank Prof. Kenneth D. Miller and his lab members for support and discussions. I was supported by the M.Sc. Neuroscience program at the Bernstein Center for Computational Neuroscience at the University of Freiburg, Germany and by an external research internship at Columbia University at the City of New York, USA.