Predicting proprioceptive cortical anatomy and neural coding with topographic autoencoders

Max Grogan, Kyle P. Blum, Yufei Wu, John Alexander Harston, Lee E. Miller, A. Aldo FaisalPLOS Computational Biology 2024 · DOI

TL;DR

This paper introduces a novel "topographic variational autoencoder" (topo-VAE) that models how the brain represents limb position and movement. By training on natural human movement data, the model surprisingly predicts key features of neural activity and spatial organization in the primate somatosensory cortex, offering a new computational framework for understanding proprioception.

What Problem Does This Paper Address?

Proprioception, our sense of limb position and movement, is crucial for motor control but remains poorly understood compared to senses like vision or touch. A major challenge is mapping how the brain represents this complex sensory information. While we know neurons in the somatosensory cortex are organized topographically (neighboring neurons represent similar body parts), the specific computational principles driving this organization for proprioception are unclear. Existing experimental techniques, like microelectrode arrays, have limited spatial resolution to fully capture these fine-grained cortical maps. This paper aims to bridge this gap by developing a computational model that can predict both the neural coding properties and the spatial organization of proprioceptive neurons.

Methods

The core of this work is the topo-VAE, a novel deep neural network architecture. It's a type of variational autoencoder (VAE), which is an unsupervised learning method that learns to compress data into a lower-dimensional representation (the "latent space") and then reconstruct the original data from that representation.

What makes the topo-VAE unique is its topographic latent space. Instead of a standard VAE where the latent space is just a collection of independent variables, the topo-VAE arranges its latent neurons in a 2D grid, mimicking the spatial layout of cortical neurons. Crucially, it incorporates a "Mexican hat" lateral interaction term between these latent neurons. This term models biological neural interactions: short-range excitation (neighboring neurons excite each other) and longer-range inhibition (neurons further apart inhibit each other). This topographic organization and lateral interaction are added as a regularizer to the standard VAE loss function.

The model is trained to reconstruct proprioceptive stimuli, effectively learning efficient representations of movement data. The authors then compare the properties of the simulated neurons (their tuning preferences, receptive fields) and their spatial organization to actual neural recordings.

Datasets

  • Training Data:
    • Natural Human Movement Data: Full-body movement data from 18 participants performing unconstrained daily tasks (e.g., preparing food, eating) in a kitchen environment. This data captures the natural covariance and complexity of human arm movements.
  • Stereotyped Planar Centre-Out Reaching Data: Data from participants performing a simplified, planar centre-out reaching task. This is used to compare the model's performance when trained on more constrained movements.
  • Evaluation Data:
    • Nonhuman Primate Behaviour and Data: Previously recorded neural data from area 2 of the somatosensory cortex of rhesus macaques performing a planar centre-out reaching task. This dataset is used to validate the model's predictions about neural coding and spatial organization.

Key Insights and Ideas

  • Natural Movement Statistics are Key: Training the model on natural, complex human movements, rather than simplified tasks, leads to significantly better predictions of neural activity and organization. This supports the "efficient coding hypothesis," suggesting that sensory systems are optimized to represent the statistics of their natural input.
  • Lateral Interactions Shape Cortical Maps: The "Mexican hat" lateral interaction between neurons in the model's latent space is crucial for reproducing the observed topographic organization and tuning properties of real neurons. Without it, the model's predictions are much poorer.
  • Blob-and-Pinwheel Structure: The model predicts a "blob-and-pinwheel" type geometry for preferred direction (PD) maps in the cortex, where PDs are clustered in blobs and organized around pinwheel centers. This is a novel prediction for proprioception.
  • Multi-Joint Representation: The model shows that neurons don't just encode single joints but often represent combinations of multiple joints, reflecting the natural covariance of arm movements. This suggests a more integrated representation of limb state.
  • Beyond Tactile Homunculus: The paper challenges the idea of a simple, point-to-point "homunculus" for proprioception, suggesting a more complex, potentially "pinwheel-like" organization related to movement direction.

Key Results

  • The topo-VAE model, trained on natural movement data, successfully reproduces two key observations from macaque proprioceptive neurons:
  • The shape and velocity dependence of proprioceptive receptive fields.
  • The distribution of neuronal preferred directions (PDs).
  • The model's PD distribution was statistically similar to recorded neurons, showing a bimodal distribution with a bias towards a specific reach axis (120°/300°), consistent with musculoskeletal constraints.
  • When comparing tuning curve geometry features (half-peak width, velocity gradients), the model trained with lateral effects on natural data showed the highest similarity to recorded neurons.
  • The model predicts a clustered, "blob-and-pinwheel" organization of preferred directions, which is consistent with the observation that neurons on the same electrode (closely spaced) have more similar PDs than neurons on different electrodes.
  • The model predicts that few neurons will encode just a single joint; instead, most neurons represent combinations of multiple joints.

Limitations and Future Work

  • Qualitative Similarity: While promising, the comparisons between modelled and recorded neurons are largely qualitative, especially at the population statistics level.
  • Model Assumptions: The model assumes that proprioceptive organizational principles derived from human kinematics extend to monkeys. Future work could use musculoskeletal models to identify kinematic differences between species.
  • Input Data: The current model uses only joint kinematics. Future work should incorporate more biologically realistic inputs like muscle spindle and Golgi tendon organ signals.
  • Experimental Validation: The authors acknowledge that some predictions, like the precise receptive field structure, are difficult to test with current data. Future experiments could aim to validate these predictions.
  • Alternative Lateral Effects: While the Mexican hat function was effective, other forms of lateral interaction could be explored.

Summary updated

BibTeX
@article{grogan2024predicting,
  author = {Max Grogan and Kyle P. Blum and Yufei Wu and John Alexander Harston and Lee E. Miller and A. Aldo Faisal},
  title = {{Predicting proprioceptive cortical anatomy and neural coding with topographic autoencoders}},
  journal = {PLOS Computational Biology},
  year = {2024},
  doi = {10.1371/journal.pcbi.1012614},
}
Figure