Choice Dynamics and Geometry in Premotor Cortex

In a groundbreaking study published in Nature, researchers from multiple institutions have unveiled new insights into how the premotor cortex (PMd) encodes the formation of decisions in dynamic, single-trial neural activity. This work shines a light on the intricate geometry underlying neural population codes, offering a fresh perspective on the mechanisms by which the brain transforms sensory inputs into choices.

The exploration focused on neural activity recorded from the PMd, an area well-known for its involvement in the planning and execution of movements. What sets this research apart is its use of unsupervised modeling methods that infer latent neural dynamics without relying on information about the choices animals ultimately made during task performance. The researchers aimed to determine whether these latent trajectories, extracted solely from neural spiking data, correspond to the behavioral outcomes on a trial-by-trial basis.

To achieve this, the team decoded latent trajectories representing the evolving neural decision variable, denoted as x(t), from spike trains recorded while monkeys performed a choice task. Crucially, the prediction of the animals’ choices was based on the boundary that the latent trajectory approached at the measured reaction time for each trial. This approach allowed the researchers to test whether the model-derived latent decision variables genuinely reflected single-trial decision formation dynamics.

.adsslot_8FUaSNEQkn{ width:728px !important; height:90px !important; }
@media (max-width:1199px) { .adsslot_8FUaSNEQkn{ width:468px !important; height:60px !important; } }
@media (max-width:767px) { .adsslot_8FUaSNEQkn{ width:320px !important; height:50px !important; } }

ADVERTISEMENT

Remarkably, both single-neuron and population-level models were able to predict the animals’ choices with accuracies significantly exceeding chance. The single-neuron models achieved balanced accuracies approaching 70% in one monkey and near 60% in another, while population models pushed these figures even higher, reaching nearly 90% accuracy in some cases. These results strongly suggest that the latent variables inferred from PMd neural populations meaningfully capture the underlying decision-making processes on a fine temporal scale.

The team further compared their unsupervised model predictions against a conventional logistic regression decoder trained in a supervised manner to predict choices from binned spike counts. Despite this decoder having direct access to choice labels during training, the unsupervised models outperformed it significantly across multiple metrics. This outcome challenges conventional notions about the necessary supervision for effective decoding of cognitive variables and highlights the power of latent dynamics models in revealing embedded cognitive signals.

Statistical analyses, including Wilcoxon signed-rank tests and Mann–Whitney U-tests, confirmed the robustness and significance of these findings across different animals and modeling approaches. Notably, population-level dynamics consistently yielded higher choice prediction accuracies than single-neuron models. This disparity underscores how decision variables are not restricted to isolated neurons but are encoded across ensembles, reflecting a higher-dimensional neural representation.

Beyond choice prediction, the research delves into the architecture of the population code, revealing that PMd neurons exhibit heterogeneous tuning functions for the same dynamic decision variable. This heterogeneity is not mere noise or redundancy. Instead, it sculpts a geometric configuration of neural activity in a latent space that faithfully represents cognitive variables like choice. Such a unifying geometric principle bridges the encoding of both sensory and higher-level cognitive dynamics within a consistent neural framework.

The implications of this study extend far beyond the premotor cortex or the specific decision-making tasks investigated. By exemplifying an unsupervised approach to infer behaviorally relevant latent variables, the findings pave the way for broader applications in brain-machine interfaces, clinical diagnostics, and fundamental cognitive neuroscience research. Decoding neural representations on a single-trial basis captures the real-time computations underlying behavior, which is crucial for understanding how the brain navigates complex environments.

Moreover, the discovery that population dynamics encode decisions more effectively than single neurons fits neatly within emerging paradigms emphasizing population codes. Neural ensembles, through collective dynamics and diversity of tuning, are better suited to capture the richness and variability of cognitive processes than any individual neuron could. This principle highlights the importance of advanced dimensionality reduction and latent variable methods in contemporary neuroscience.

The methodological rigor demonstrated in this work is noteworthy. The researchers carefully cross-validated their models, applied stringent statistical tests, and incorporated comparison analyses with established decoders. These efforts ensure that conclusions about the nature of decision variables and their encoding geometry rest on solid empirical ground, not artifacts of modeling or data interpretation.

Conceptually, framing neural computation in terms of latent dynamic variables and their geometry provides an elegant mathematical lens to view cognitive function. Such frameworks enable characterization of neural trajectories as they evolve toward decision boundaries, offering mechanistic insights into how choices emerge from sensory evidence accumulation and internal deliberation.

As our understanding deepens, this line of research promises to reveal not only the encoding but also the transformations and computations that neural populations undertake during decision-making. Future studies may explore how such latent dynamics relate to other brain regions’ activities, neuromodulatory influences, or learning-induced changes in neural geometry.

In summary, this research presents compelling evidence that heterogeneous populations in the PMd encode a shared dynamic decision variable through diverse tuning functions that construct a geometric population code. This discovery unites sensory and cognitive neural encoding within a common dynamical systems framework, advancing our comprehension of brain function and offering new avenues for decoding neural signals in naturalistic behaviors.

Subject of Research:
The neural dynamics and population encoding of decision variables in the premotor cortex during behavior.

Article Title:
The dynamics and geometry of choice in the premotor cortex.

Article References:
Genkin, M., Shenoy, K.V., Chandrasekaran, C. et al. The dynamics and geometry of choice in the premotor cortex. Nature (2025). https://doi.org/10.1038/s41586-025-09199-1

Image Credits: AI Generated

Tags: behavioral outcomes and neural dynamicschoice prediction in primate studiesdynamic neural activity in decision-makingencoding sensory inputs in the braininsights from Nature neuroscience studieslatent trajectories in neural spiking dataneural decision variables in choice tasksneural population coding geometrypremotor cortex decision-makingreaction time and neural activitytrial-by-trial decision variablesunsupervised modeling in neuroscience