Structured connectivity in the cerebellum enables noise-resilient pattern separation

Tri Nguyen, Logan A. Thomas, Jeff L. Rhoades, Ilaria Ricchi, Xintong Cindy Yuan, Arlo Sheridan, David G. C. Hildebrand, Jan Funke, Wade G. Regehr, Wei-Chung Allen LeebioRxiv 2021 · DOI

TL;DR

This paper challenges the long-held assumption of random connectivity in the cerebellum's input layer. By mapping the actual synaptic connections using electron microscopy, they discovered structured, non-random motifs that surprisingly enhance pattern separation and noise resilience, offering valuable insights for designing more robust artificial neural networks.

What Problem Does This Paper Address?

The cerebellum is crucial for motor control, social behaviors, and cognition, requiring it to process information rapidly and accurately, even in the presence of noise. Traditional theories proposed that the cerebellum's input layer (mossy fiber to granule cell connections) uses random connectivity to maximize its encoding capacity. However, this randomness is also known to reduce noise resilience. This creates a fundamental tradeoff: how can a neural circuit be both highly capable of encoding information and robust to noisy inputs? This paper aims to resolve this by investigating the actual synaptic connectivity of the cerebellar cortex to understand how it balances these competing demands.

Methods

The authors employed a multi-pronged approach:

  • Large-Scale Electron Microscopy (EM): They generated a synapse-resolution EM dataset of a portion of the mouse cerebellar cortex. This involved serial sectioning, imaging at high resolution (4 nm/pixel), and stitching together a massive 3D volume (37.2 trillion voxels).
  • Automated Segmentation: They adapted and utilized a sophisticated pipeline involving convolutional neural networks (CNNs) for affinity prediction and watershed algorithms for over-segmentation, followed by agglomeration to reconstruct individual neurons.
  • Synapse Prediction: Another CNN was trained to identify synaptic clefts and predict pre- and post-synaptic partners, enabling the mapping of individual connections.
  • Connectivity Analysis: They quantified the precise synaptic connections between mossy fibers (MFs) and granule cells (GrCs), and between GrCs and Purkinje cells (PCs).
  • Numerical Simulations: They developed computational models to simulate network activity based on the reconstructed connectivity and compared it to models with random connectivity. This allowed them to analyze dimensionality, signal-to-noise ratio (SNR), and pattern separation capabilities.
  • Novel Metrics: They used Hamming similarity to quantify input similarity between neurons and spectral clustering to identify functional groups of PCs.

Datasets

  • Newly generated EM dataset: A synapse-resolution EM dataset of lobule V from the adult mouse vermis was created for this study. This dataset is described as a 37.2 trillion voxel volume.
  • Existing datasets for tool development: The authors mention using the CREMI network and datasets for training their segmentation and synapse prediction networks, which are standard benchmarks in the field.

Key Insights and Ideas

  • Challenging the Randomness Assumption: The most significant finding is that the MF→GrC connectivity is not random, as previously theorized. Instead, it exhibits structured, redundant, and selective motifs.
  • Oversharing and Overrepresentation: The input layer shows "oversharing" (GrC pairs sharing more MF inputs than expected) and "selective overrepresentation" (some MF boutons are sampled by many more GrCs). These are key redundant motifs.
  • Balancing Act: These non-random motifs appear to strike a crucial balance. They increase the discriminability of similar input patterns (improving pattern separation) with only a minimal cost to the network's overall encoding capacity.
  • Noise Resilience: The structured connectivity enhances noise resilience, a critical feature for real-time motor control and other cerebellar functions.
  • Predictive Power of Connectivity: The study demonstrates that the specific wiring patterns can predict functional subnetworks of Purkinje cells, suggesting a link between structure and function.
  • Implications for AI: The findings offer principles for designing more robust and efficient artificial neural networks, particularly in areas requiring noise tolerance and fine-grained pattern discrimination.

Key Results

  • MF→GrC Connectivity: GrC pairs shared significantly more MF bouton inputs than predicted by random models (p = 3.9 × 10⁻¹²). MF boutons were also not uniformly sampled, with the most connected third having 4 times as many postsynaptic GrCs as the least connected third. Kurtosis analysis confirmed over- and under-sampling (p = 0.0146).
  • GrC→PC Connectivity: While the overall connection rate was around 49%, the study found evidence of selective connectivity patterns, with local GrC axons showing higher input similarity to PCs than non-local ones.
  • Dimensionality and SNR: Simulations showed that the reconstructed network had higher dimensionality for similar input patterns but lower dimensionality for different ones. Crucially, the structured connectivity increased the signal-to-noise ratio (SNR), making it easier to discriminate similar patterns, especially when using selective subsampling in PCs.
  • Computational Efficiency: The authors developed and utilized advanced computational tools for large-scale EM data processing, segmentation, and reconstruction, enabling them to analyze over 150,000 synapses and >5,500 neurons.

Limitations and Future Work

  • Limited Volume: The EM dataset, while large, represents a specific portion of the cerebellar cortex (lobule V). Future work could explore connectivity across larger regions and different cerebellar areas.
  • Static Connectivity: The study maps static connectivity. Future research could investigate how this connectivity changes with learning and experience, particularly in light of activity-dependent plasticity.
  • In Vivo Validation: While simulations provide strong evidence, direct experimental validation of these connectivity motifs and their functional consequences in vivo would be valuable.
  • Long-Range Connectivity: The paper focuses on feedforward connectivity. Future studies could incorporate long-range connections to gain a more complete understanding of cerebellar circuit function.
  • Artificial Neural Network Design: The authors explicitly suggest that these findings have implications for improving artificial intelligence algorithms, particularly in designing more robust and efficient neural network architectures.

Summary updated

Abstract

Summary The cerebellum is thought to detect and correct errors between intended and executed commands 1–3 and is critical for social behaviors, cognition and emotion 4–6 . Computations for motor control must be performed quickly to correct errors in real time and should be sensitive to small differences between patterns for fine error correction while being resilient to noise 7 . Influential theories of cerebellar information processing have largely assumed random network connectivity, which increases the encoding capacity of the network’s first layer 8–13 . However, maximizing encoding capacity reduces resiliency to noise 7 . To understand how neuronal circuits address this fundamental tradeoff, we mapped the feedforward connectivity in the mouse cerebellar cortex using automated large-scale transmission electron microscopy (EM) and convolutional neural network-based image segmentation. We found that both the input and output layers of the circuit exhibit redundant and selective connectivity motifs, which contrast with prevailing models. Numerical simulations suggest these redundant, non-random connectivity motifs increase discriminability of similar input patterns at a minimal cost to the network’s overall encoding capacity. This work reveals how neuronal network structure can balance encoding capacity and redundancy, unveiling principles of biological network architecture with implications for artificial neural network design.
BibTeX
@misc{nguyen2021structured,
  author = {Tri Nguyen and Logan A. Thomas and Jeff L. Rhoades and Ilaria Ricchi and Xintong Cindy Yuan and Arlo Sheridan and David G. C. Hildebrand and Jan Funke and Wade G. Regehr and Wei-Chung Allen Lee},
  title = {{Structured connectivity in the cerebellum enables noise-resilient pattern separation}},
  year = {2021},
  doi = {10.1101/2021.11.29.470455},
  howpublished = {bioRxiv},
}

Loading related content...

Figure