Neural dynamics outside task-coding dimensions drive decision trajectories through transient amplification

Ulises Pereira-Obilinovic, Kayvon Daie, Susu Chen, Karel Svoboda, Ran DarshanbioRxiv 2025 · DOI

TL;DR

This paper challenges the conventional view that neural decision-making is solely driven by "coding dimensions" that directly represent task variables. Instead, it demonstrates that "residual dimensions," previously considered noise, play a crucial causal role in shaping decisions through a mechanism of transient amplification. This finding suggests that weakly selective neural populations, operating in these residual dimensions, are critical for flexible and nuanced behavior.

What Problem Does This Paper Address?

For decades, neuroscience research has focused on identifying "coding dimensions" within neural activity – low-dimensional subspaces that directly correlate with task variables like stimuli or choices. The prevailing assumption has been that these coding dimensions are the primary drivers of behavior. However, it remains unclear whether these dimensions actively cause decisions or merely reflect underlying computations. Furthermore, neural activity outside these task-aligned subspaces, termed "residual dimensions," has largely been ignored, despite the possibility that it could also causally influence behavior. This paper tackles the question of whether these seemingly irrelevant residual dimensions play a critical role in shaping decision-making dynamics.

Methods

The authors developed a novel recurrent neural network (RNN) framework designed to explicitly model neural activity by distinguishing between "coding" and "residual" subspaces. This model was constrained to capture all possible recurrent interactions between these two types of dimensions. Coding dimensions were optimized to represent task variables (stimulus, choice, response), while residual dimensions captured the remaining variance and were orthogonal to the coding subspace. This approach allowed for the direct modeling of dynamic interactions between coding and residual subspaces at the single-trial level.

They applied this framework to electrophysiological recordings from the anterior lateral motor cortex (ALM) and motor thalamus in mice performing a delayed-response decision-making task. Key analytical techniques included:

  • Dimensionality Reduction: Linear Discriminant Analysis (LDA) was used to identify task-relevant coding dimensions (sample, choice, response).
  • Network Perturbation: The trained RNN was perturbed along coding and residual dimensions to assess their causal influence on decision outcomes.
  • Pseudospectra Analysis: This technique was used to analyze the non-normal interactions within the network's dynamics, revealing how transient amplification occurs.
  • Schur Decomposition: This method was employed to decompose the interaction matrix into stable and transiently amplifying components, providing insight into the dynamics.
  • Single-Trial Modeling: A framework was developed to model trial-to-trial variability as low-dimensional input biases acting on both coding and residual dimensions.

Datasets

The study utilized electrophysiological recordings from mice performing a memory-guided decision-making task. Specifically, data were obtained from:

  • Anterior Lateral Motor Cortex (ALM): Recordings from 7,884 cortical neurons (3,640 left hemisphere, 4,244 right hemisphere) across 176 behavioral sessions.
  • Ventromedial (VM) and Ventrolateral (VL) Thalamic Nuclei: Recordings from 1,932 thalamic neurons (1,004 left hemisphere, 928 right hemisphere).

For single-trial fits, a session with 102 simultaneously recorded ALM neurons was used. The paper also provides links to GitHub repositories containing the code for training and analyzing the models, which could be used to reproduce results on similar datasets.

Key Insights and Ideas

  • Residual Dimensions are Causally Important: The most striking finding is that perturbations to residual dimensions, which have weak task encoding, significantly altered behavioral choices, while perturbations to the choice dimension had little effect. This directly challenges the notion that only task-aligned dimensions drive decisions.
  • Transient Amplification is Key: The mechanism by which residual dimensions influence decisions is through "non-normal amplification." This involves transient, high-dimensional dynamics that temporarily amplify small deviations before collapsing into choice-selective attractors.
  • Intermediate Dimensionality is Crucial: The dynamics that drive decisions unfold in intermediate-dimensional subspaces (around 10 dimensions), which are a combination of coding and residual dimensions, rather than just the low-dimensional coding subspaces.
  • Thalamus Influences Cortex via Residual Dimensions: The study suggests that weakly selective thalamic populations influence cortical decision-making by projecting onto residual dimensions in ALM, rather than through direct choice-selective pathways.
  • Errors Arise from Residual Variability: Trial-to-trial variability, particularly fluctuations along residual dimensions, is a primary driver of decision errors.

Key Results

  • Perturbation Effects: Perturbing residual dimensions reliably flipped behavioral choices, whereas perturbing the choice dimension was largely ineffective. This effect was robust across hemispheres and mirrored findings from experimental optogenetic perturbations.
  • Dimensionality: The model found that an intermediate dimensionality of approximately 10 dimensions (a mix of coding and residual) provided the best fit and alignment with empirical data.
  • Error Trials: Analysis of error trials revealed that they primarily arise from shifts in trajectories along residual dimensions, biasing decisions.
  • Thalamocortical Influence: In a thalamocortical model, perturbations of residual dimensions in the thalamus disrupted cortical choice selectivity, while choice-dimension perturbations had less impact.
  • Model Fit: The trained RNN models successfully captured both condition-averaged and single-trial neural activity, demonstrating the importance of residual dimensions in explaining observed dynamics.

Limitations and Future Work

  • Model Simplifications: While the RNN model captures key aspects of neural dynamics, it is a simplification of biological complexity. For instance, the model uses rate-based neurons, not spiking neurons, and the connectivity is constrained to be low-rank.
  • Generalizability: While the study used data from ALM and motor thalamus, future work could explore whether these findings generalize to other brain regions involved in decision-making.
  • Biological Mechanisms: Further research is needed to fully elucidate the specific biological mechanisms underlying non-normal amplification and the role of specific neuronal populations in residual dimensions.
  • Task Complexity: The study focused on a specific delayed-response task. Investigating how these principles apply to more complex decision-making scenarios with varying sensory evidence or cognitive demands would be valuable.
  • Thalamic Specificity: While the study suggests thalamic residual dimensions influence cortical selectivity, further work could explore the specific types of thalamic neurons and their connectivity patterns that mediate this effect.

Summary updated

Abstract

Most behaviors involve neural dynamics in high-dimensional activity spaces. A common approach is to extract dimensions that capture task-related variability, such as those separating stimuli or choices, yielding low-dimensional, task-aligned neural activity subspaces ("coding dimensions"). However, whether these dimensions actively drive decisions or merely reflect underlying computations remains unclear. Moreover, neural activity outside these coding subspaces ("residual dimensions") is often ignored, though it could also causally shape neural dynamics driving behavior. We developed a recurrent neural network model that fits population activity and uncovers the dynamic interactions between coding and residual subspaces on single trials. Applied to electrophysiological recordings from the anterior lateral motor cortex (ALM) and motor thalamus in mice performing a delayed response task, our model demonstrates that perturbations of residual dimensions reliably alter behavioral choices, whereas perturbations of the choice dimension, which strongly encodes the animal's upcoming decision, are largely ineffective. These perturbation effects arise because residual dimensions drive transient amplification across an intermediate number of coding and residual dimensions (~10), before the dynamics collapse into discrete attractor states corresponding to the animal's choice. By dissecting the low-dimensional variability underlying error trials, we find that it primarily shifts trajectories along residual dimensions, biasing single decisions. Residual activity in thalamus shapes cortical decision dynamics, implicating weakly selective thalamic populations in the emergence of cortical selectivity. Our findings challenge the conventional focus on low-dimensional coding subspaces as sufficient framework for understanding neural computations, demonstrating that dimensions previously considered task-irrelevant and accounting for little variance can have a critical role in driving behavior.
BibTeX
@misc{pereiraobilinovic2025neural,
  author = {Ulises Pereira-Obilinovic and Kayvon Daie and Susu Chen and Karel Svoboda and Ran Darshan},
  title = {{Neural dynamics outside task-coding dimensions drive decision trajectories through transient amplification}},
  year = {2025},
  doi = {10.1101/2025.11.20.689599},
  howpublished = {bioRxiv},
}
Figure