Binary and analog variation of synapses between cortical pyramidal neurons

Sven Dorkenwald, Nicholas L. Turner, Thomas Macrina, Kisuk Lee, Ran Lu, Jingpeng Wu, Ágnes L. Bodor, Adam Bleckert, Derrick Brittain, Nico Kemnitz, William Silversmith, Dodam Ih, Jonathan Zung, Aleksandar Zlateski, Ignacio Tartavull, Szi-chieh Yu, Sergiy Popovych, William Wong, Manuel Castro, Chris S. Jordan, Alyssa M. Wilson, Emmanouil Froudarakis, JoAnn Buchanan, Marc Takeno, Russel Torres, Gayathri Mahalingam, Forrest Collman, Casey M Schneider-Mizell, Daniel J. Bumbarger, Li Yang, Lynne A. Becker, Shelby Suckow, Jacob Reimer, Andreas S. Tolias, Nuno Maçarico da Costa, R. Clay Reid, H. Sebastian SeungbioRxiv 2019 · DOI

Abstract

Learning from experience depends at least in part on changes in neuronal connections. We present the largest map of connectivity to date between cortical neurons of a defined type (L2/3 pyramidal cells), which was enabled by automated analysis of serial section electron microscopy images with improved handling of image defects. We used the map to identify constraints on the learning algorithms employed by the cortex. Previous cortical studies modeled a continuum of synapse sizes by a log-normal distribution. A continuum is consistent with most neural network models of learning, in which synaptic strength is a continuously graded analog variable. Here we show that synapse size, when restricted to synapses between L2/3 pyramidal cells, is well-modeled by the sum of a binary variable and an analog variable drawn from a log-normal distribution. Two synapses sharing the same presynaptic and postsynaptic cells are known to be correlated in size. We show that the binary variables of the two synapses are highly correlated, while the analog variables are not. Binary variation could be the outcome of a Hebbian or other synaptic plasticity rule depending on activity signals that are relatively uniform across neuronal arbors, while analog variation may be dominated by other influences. We discuss the implications for the stability-plasticity dilemma.

Introduction

How does the brain store information while maintaining long-term stability? In both neuroscience and artificial intelligence, the standard model of learning assumes that information is encoded through the continuous adjustment of synaptic weights—an "analog" process. However, this model faces the stability-plasticity dilemma: synapses must be volatile enough to allow for new learning but stable enough to preserve memories over time.

By reconstructing the largest-ever connectivity map of Layer 2/3 (L2/3) pyramidal cells in the mouse primary visual cortex, this research reveals that synaptic strength is not merely a continuous spectrum. Instead, it is best modeled as a combination of a discrete binary state (Small vs. Large) and analog variation (continuous "noise"). This discovery suggests a structural backbone for stable memory within the cortex.

Large-Scale Connectomics of L2/3 Pyramidal Cells

The researchers utilized serial section electron microscopy (ssEM) to image a $250 \times 140 \times 90 \mu m^3$ volume of the mouse primary visual cortex. The primary dataset contained 364 pyramidal cells and 1,961 synapses.

Overview of the ssEM pipeline, including slice alignment, automated segmentation, and 3D reconstruction of the cortical volume and individual neurons.

A core technical challenge in large-scale connectomics is handling image defects, such as misalignments and missing sections. The authors addressed this by augmenting convolutional neural network (CNN) training with simulated defects and developing a "human-in-the-loop" alignment pipeline. Automated cleft detectors and spine head volume measurements were used as proxies for synaptic strength, allowing for a high-throughput analysis of connectivity patterns.

The Binary-Analog Hybrid Model

The study found that the discrete structure of synaptic strength only becomes apparent when focusing strictly on connections between defined cell types (L2/3-to-L2/3). When analysis included "orphan" axons from unknown sources, the discrete states were masked by a log-normal distribution.

Key findings include:

  • Model Superiority: A mixture of two log-normal distributions fit the synapse size data significantly better than a single continuous distribution ($p < 1e-39$).
  • Correlated Binary States: In "dual connections" (where two synapses connect the same pair of neurons), the binary states are highly correlated (Pearson’s phi = 0.637). If one synapse is "Large," the other is likely to be "Large" as well.
  • Independent Analog Jitter: While the binary states are linked, the remaining analog variation (the specific "size" within a state) between paired synapses is statistically independent.
Connectivity graph of the reconstructed volume and visualization of a dual connection between two pyramidal neurons, showing correlated synaptic structures.

Furthermore, synapses in the "Large" state were significantly more likely to contain a spine apparatus, a specialized organelle, providing a biological marker for the statistically identified discrete state.

Discussion and Implications

These results suggest that synaptic plasticity may operate on two levels. The binary state likely reflects Hebbian learning driven by activity signals that are uniform across neuronal arbors, providing a stable foundation for memory. In contrast, analog variation may be dominated by local molecular noise or spontaneous dynamical fluctuations.

While the study is limited by the physical boundaries of the imaged volume—meaning reported synapse counts are lower bounds—it provides a new framework for understanding cortical architecture. Future work will investigate whether these binary-analog rules generalize across different brain regions and species, potentially informing more robust learning algorithms in artificial neural networks.

Summary updated

Abstract

Abstract Learning from experience depends at least in part on changes in neuronal connections. We present the largest map of connectivity to date between cortical neurons of a defined type (L2/3 pyramidal cells), which was enabled by automated analysis of serial section electron microscopy images with improved handling of image defects. We used the map to identify constraints on the learning algorithms employed by the cortex. Previous cortical studies modeled a continuum of synapse sizes (Arellano et al. 2007) by a log-normal distribution (Loewenstein, Kuras, and Rumpel 2011; de Vivo et al. 2017; Santuy et al. 2018). A continuum is consistent with most neural network models of learning, in which synaptic strength is a continuously graded analog variable. Here we show that synapse size, when restricted to synapses between L2/3 pyramidal cells, is well-modeled by the sum of a binary variable and an analog variable drawn from a log-normal distribution. Two synapses sharing the same presynaptic and postsynaptic cells are known to be correlated in size (Sorra and Harris 1993; Koester and Johnston 2005; Bartol et al. 2015; Kasthuri et al. 2015; Dvorkin and Ziv 2016; Bloss et al. 2018; Motta et al. 2019). We show that the binary variables of the two synapses are highly correlated, while the analog variables are not. Binary variation could be the outcome of a Hebbian or other synaptic plasticity rule depending on activity signals that are relatively uniform across neuronal arbors, while analog variation may be dominated by other influences. We discuss the implications for the stability-plasticity dilemma.
BibTeX
@misc{dorkenwald2019binary,
  author = {Sven Dorkenwald and Nicholas L. Turner and Thomas Macrina and Kisuk Lee and Ran Lu and Jingpeng Wu and Ágnes L. Bodor and Adam Bleckert and Derrick Brittain and Nico Kemnitz and William Silversmith and Dodam Ih and Jonathan Zung and Aleksandar Zlateski and Ignacio Tartavull and Szi-chieh Yu and Sergiy Popovych and William Wong and Manuel Castro and Chris S. Jordan and Alyssa M. Wilson and Emmanouil Froudarakis and JoAnn Buchanan and Marc Takeno and Russel Torres and Gayathri Mahalingam and Forrest Collman and Casey M Schneider-Mizell and Daniel J. Bumbarger and Li Yang and Lynne A. Becker and Shelby Suckow and Jacob Reimer and Andreas S. Tolias and Nuno Maçarico da Costa and R. Clay Reid and H. Sebastian Seung},
  title = {{Binary and analog variation of synapses between cortical pyramidal neurons}},
  year = {2019},
  doi = {10.1101/2019.12.29.890319},
  howpublished = {bioRxiv},
}
Figure