Predicting proprioceptive cortical anatomy and neural coding with topographic autoencoders

Max Grogan, Kyle P. Blum, Yufei Wu, J. Alex Harston, Lee E. Miller, A. Aldo Faisal

Introduction

Proprioception—the "sixth sense" that allows us to perceive the position and movement of our bodies—is fundamental to motor control. Yet, compared to vision or touch, our understanding of how the brain represents limb pose in the somatosensory cortex remains limited. Proprioceptive neurons are located in relatively inaccessible areas like the sulcal regions of the primary somatosensory cortex (Area 2 and 3a), making high-density recordings challenging. Furthermore, unlike the 2D surface of the skin, proprioception involves high-dimensional inputs from numerous muscles and joints, raising the question of how the brain maps this complexity onto a 2D cortical sheet.

This research introduces a "topo-VAE"—a topographic variational autoencoder with biologically inspired lateral connectivity—to predict cortical maps from natural movement data. Remarkably, without being trained on neural data, the model replicates key features of proprioceptive neurons in macaques, suggesting that cortical organization is a direct consequence of the statistical structure of natural behavior.

Methods: The Topographic Variational Autoencoder

The researchers developed the topo-VAE to perform unsupervised feature learning on sensory stimuli while adhering to biological constraints. The model consists of an encoder that compresses high-dimensional movement data into a latent space and a decoder that reconstructs the original input.

Two critical biological features distinguish the topo-VAE from standard models:

  1. Topographic Latent Space: Latent neurons are arranged on an $80 \times 80$ grid representing the 2D cortical surface.
  2. Mexican-Hat Lateral Connectivity: A lateral interaction term mimics cortical synaptic architecture, where nearby neurons excite each other, intermediate-range neurons inhibit each other, and distant neurons have no effect.
The topo-VAE framework and experimental pipeline.

The model was trained using the Infomax principle on two primary datasets: full-body kinematics from humans performing daily tasks (natural movement) and constrained 2D "center-out" reaching tasks. The resulting representations were then compared against single-unit recordings from the somatosensory cortex of rhesus macaques.

Results: Replicating Biological Neural Coding

The topo-VAE successfully reproduced several key observations from neural recordings without explicit fitting to those data.

Preferred Direction (PD) Matching

One of the most significant findings was the model's ability to match the distribution of neuronal "preferred directions" (PDs). When trained on natural 3D movement data, the topo-VAE's PD distribution closely mirrored the entropy and shape found in real monkey data. Interestingly, models trained only on simplified planar reaching data failed to capture this distribution, highlighting the importance of "natural statistics" in shaping neural representations.

Comparison of modelled and recorded neural properties.

Emergent Multi-Joint Coding and Topography

The model naturally developed neurons that respond to combinations of joints (e.g., shoulder and elbow), a known feature of Area 2. Furthermore, the "Mexican-hat" constraint led to the emergence of "blob-and-pinwheel" structures, where neurons with similar tuning cluster together, interrupted by singular points where all directions meet. This suggests that the physical layout of the cortex is a functional solution to mapping high-dimensional data onto a 2D surface.

Discussion and Future Directions

The success of the topo-VAE demonstrates that the structure of the proprioceptive cortex can be predicted from the statistics of natural movement and basic biological constraints. While the current model is purely kinematic and does not account for muscle force or spindle dynamics, it provides a principled basis for understanding sensorimotor representations.

Future work aims to integrate biomechanical models to explore how force-sensing affects the cortical map. Additionally, the predicted "pinwheel" geometry offers a testable hypothesis for future high-density recording studies. These insights have significant implications for the development of more effective sensory feedback in brain-computer interfaces and the design of humanoid robot controllers that mimic biological efficiency.

Summary updated

Abstract

Abstract Proprioception is one of the least understood senses, yet fundamental for the control of movement. Even basic questions of how limb pose is represented in the somatosensory cortex are unclear. We developed a topographic variational autoencoder with lateral connectivity (topo-VAE) to compute a putative cortical map from a large set of natural movement data. Although not fitted to neural data, our model reproduces two sets of observations from monkey centre-out reaching: 1. The shape and velocity dependence of proprioceptive receptive fields in hand-centered coordinates despite the model having no knowledge of arm kinematics or hand coordinate systems. 2. The distribution of neuronal preferred directions (PDs) recorded from multi-electrode arrays. The model makes several testable predictions: 1. Encoding across the cortex has a blob-and-pinwheel-type geometry of PDs. 2. Few neurons will encode just a single joint. Our model provides a principled basis for understanding of sensorimotor representations, and the theoretical basis of neural manifolds, with applications to the restoration of sensory feedback in brain-computer interfaces and the control of humanoid robots. Author Summary It is well established that proprioception is essential for effective for motor control, yet our understanding of proprioceptive coding in somatosensory cortex is far behind that of more established sensory modalities such as vision and touch. Here, we use unsupervised learning of a deep neural network imposed with biological constraints to reproduce coding properties of proprioceptive neurons in area 2 of primary somatosensory cortex. With this model, we demonstrate that the tendency for area 2 neurons in close physical proximity to share similar directional tuning can be explained by local topographic organisation driven by Mexican hat lateral effects, a phenomenon that is otherwise unobservable with the spatial resolution of available electrode microarray recordings. This provides, to the best of our knowledge, the first evidence of local topographic organisation in area 2 proprioceptive neurons. We also predict that the structure of proprioceptive receptive fields are typically multi-joint rather than single-joint in nature and provide, to the best of our knowledge, and find that our model best reproduces the tuning properties of neural data when training on natural kinematic recordings, rather than those from a stereotyped task, underlining the importance of training on data that reflect the natural distribution of sensory stimuli.
BibTeX
@inproceedings{grogan2022predicting,
  author = {Max Grogan and Kyle P. Blum and Yufei Wu and J. Alex Harston and Lee E. Miller and A. Aldo Faisal},
  title = {{Predicting proprioceptive cortical anatomy and neural coding with topographic autoencoders}},
  booktitle = {2022},
  year = {2022},
  doi = {10.32470/ccn.2022.1178-0},
}
Figure