Tensor Analysis Reveals Distinct Population Structure that Parallels the Different Computational Roles of Areas M1 and V1

Jeffrey S. Seely, Matthew T. Kaufman, Stephen I. Ryu, Krishna V. Shenoy, John P. Cunningham, Mark M. ChurchlandPLOS Computational Biology 2016 · DOI

Introduction

Understanding how neural populations coordinate their activity to perform computations remains a central challenge in neuroscience. In the primary motor cortex (M1), a long-standing debate persists: does neural activity represent specific movement parameters, such as velocity or force (the "tuning" view), or does it reflect the internal evolution of a dynamical system (the "dynamics" view)? While both frameworks can generate the complex, oscillatory patterns observed in M1, distinguishing between them has proven difficult using traditional analytical methods.

This study introduces a novel approach by treating neural population activity as a third-order tensor—defined by neurons, experimental conditions, and time. By analyzing the mathematical structure of these tensors, researchers discovered a fundamental divergence between the primary visual cortex (V1) and M1, suggesting that their population structures are optimized for their respective roles in sensory processing and motor generation.

Preferred-Mode Analysis: A Tensor-Based Framework

Traditional neural data analysis often involves collapsing dimensions—for example, averaging across conditions or time—which can obscure the underlying structural organization. The authors propose Preferred-Mode Analysis, which preserves the three-dimensional nature of the data ($N$ neurons $ imes$ $C$ conditions $ imes$ $T$ time points).

The core of this method involves "unfolding" the neural tensor into two distinct matrices:

  1. The Neuron Mode: This matrix organizes data to highlight how individual neurons respond across all conditions and time points.
  2. The Condition Mode: This matrix organizes data to highlight how different experimental conditions appear across the entire population over time.

By applying Singular Value Decomposition (SVD) to these matrices, the researchers could determine which mode provides a more "parsimonious" description of the data. A population is considered neuron-preferred if its activity is best explained as a set of filters for external variables (consistent with tuning). Conversely, it is condition-preferred if the population follows a consistent internal rule where future states depend on current states (consistent with dynamics).

Distinct Structures in V1 and M1

The application of this framework to physiological datasets revealed a striking contrast. Data from V1 (monkeys viewing natural movies and cats viewing grating sequences) were consistently neuron-preferred. In these sensory regions, reconstruction error was approximately 33% lower when using the neuron mode compared to the condition mode. This indicates that V1 activity is primarily driven by external stimuli, with neurons acting as independent encoders of sensory features.

In contrast, M1 data (monkeys performing a 72-condition delayed-reach task) were overwhelmingly condition-preferred. The disparity was even more pronounced: neuron-mode reconstruction performed 68% worse than condition-mode reconstruction. This suggests that M1 operates as a dynamical system where the population activity is governed by internal recurrence rather than simple external parameter mapping.

Validating Models and the Muscle Paradox

The researchers further tested various computational models of M1. Most traditional "tuning" models, which assume neurons encode reach kinematics, failed the tensor test by appearing neuron-preferred. Only models incorporating strong internal dynamics, such as Recurrent Neural Networks (RNNs), successfully replicated the condition-preferred structure seen in biological M1 data.

Interestingly, the study also examined electromyographic (EMG) recordings from arm muscles. Despite being the direct output of the motor system, the EMG data were not condition-preferred. This implies a significant transformation: while M1 functions as an internal dynamical engine, its output is converted into a format that resembles a set of external parameters by the time it reaches the muscles.

Conclusion and Future Directions

The discovery that V1 and M1 occupy opposite ends of the tensor-structure spectrum provides a powerful new lens for viewing neural computation. These "deep" structural features are not apparent from simple firing rate plots but emerge only when considering the full dimensionality of the population activity.

Future research may extend this tensor framework to higher-order areas like the prefrontal cortex to determine if cognitive processes like decision-making align more with sensory filtering or internal dynamical generation. This methodology offers a robust tool for characterizing the global computational strategies employed by different regions of the brain.

Summary updated

BibTeX
@article{seely2016tensor,
  author = {Jeffrey S. Seely and Matthew T. Kaufman and Stephen I. Ryu and Krishna V. Shenoy and John P. Cunningham and Mark M. Churchland},
  title = {{Tensor Analysis Reveals Distinct Population Structure that Parallels the Different Computational Roles of Areas M1 and V1}},
  journal = {PLOS Computational Biology},
  year = {2016},
  doi = {10.1371/journal.pcbi.1005164},
}
Figure