Synaptic unreliability, a foundational concept, found in deep learning, and in computational neuroscience, has been undermined by a math proof that shows us that MVR (multi-vesicular release) is wide spread in neocortical synapses.

This new research may impact companies like Numenta, Google, Facebook, Deepmind, Tesla, OpenAI, and the way neural networks are designed in the future.
Feb 9, 2021 · 7 min read

Article by Micah Blumberg

The concept of synaptic unreliability that is a foundation for computational neuroscience, spike timing models, and deep neural networks is fundamentally undermined by the mathematical proof of widespread MVR (multi-vesicular release) in neocortical synapses.

Background: “Impact of synaptic unreliability on the information transmitted by spiking neurons” “This suggests that synapses represent the primary bottleneck limiting the faithful transmission of information through cortical circuitry.” “The postsynaptic neuron can be viewed as an input-output element that converts the input spike trains from many presynaptic neurons into a single-output spike train. This input-output transformation is the basic computation performed by neurons. It is the foundation upon which cortical processing is based.” “

How could MVR affect the capacity neurocomputation? Well even before this researchers had the idea that a single synapse could individually set its own neurotransmitter release, dynamically, through a local feedback regulation.

“The probability of neurotransmitter release: variability and feedback control at single synapses” “(…)neuronal compartments might perform regional integration operations, acting as semi-independent computation units18,130,131. In this scenario, in which a neuron can be thought of as a multiple-unit network, it makes sense that signal/noise adjustments are performed separately for each unit rather than for the cell as a whole. Also, having synapses with different prs in different dendritic branches means that information from a single axon can be dynamically filtered in a different way at each dendritic compartment.” source

“experimental data from electrophysiological, molecular and imaging studies have demonstrated that synaptic terminals can individually set their neurotransmitter release probability dynamically through local feedback regulation. This local tuning of transmission has important implications for current models of single-neuron computation.” source

So what is this math proof:

“analysis revealed that the number of vesicle release sites exceeded the number of anatomical synapses formed by a connection by a factor of at least 2.6, which challenges the current understanding of synaptic release in neocortex and suggests that neocortical synapses operate with multivesicular release, like hippocampal synapses”

“neocortical synapses (…) may modulate their strength more flexibly than previously thought, with the corollary that the canonical neocortical microcircuitry possesses significantly higher computational power than estimated by current models.”

The above paper was also cited in this article below

“The researchers were able to use a novel mathematical analysis to prove that each synapse in fact has several sites that can release packets of neurotransmitter simultaneously. “This means that synapses are much more complex and can regulate their signal strength more dynamically than previously thought. The computational power and storage capacity of the entire neocortex therefore seems to be much greater than was previously believed,” says Kevan Martin.”


The idea of multivesicular release (MVR) being a widespread phenonemon among many synapses was previously covered in a 2015 paper on multivesicular release: “Nevertheless, functional data from many studies strongly suggests that MVR is a widespread phenomenon among synapses — more prevalent than originally assumed.”
“The ubiquitous nature of multivesicular release”

“Presynaptic action potentials trigger the fusion of vesicles to release neurotransmitter onto postsynaptic neurons. Each release site was originally thought to liberate at most one vesicle per action potential in a probabilistic fashion, rendering synaptic transmission unreliable. However, the simultaneous release of several vesicles, or multivesicular release (MVR), represents a simple mechanism to overcome the intrinsic unreliability of synaptic transmission.”

The key question here is what causes the MVR to change? The answer is the Pr.

Lets define Pr: Pr is the probability that a vesicle is released:

  1. the number of release sites (N) see Glossary,
  2. the probability that a vesicle is released (Pr),
  3. the amplitude of the postsynaptic response elicited by the content of each synaptic vesicle (q) [2].” source

“Desynchronization of MVR on a supra-millisecond time scale is also evident at inhibitory synapses between MLIs [13] where 0 to 3 vesicles can be released per action potential at single release sites”

So what modifies the Pr? Its the APsyn, and what modifies the APsyn? Potassium.

“The potassium channel subunit Kvβ1 serves as a major control point for synaptic facilitation”

“We believe our data provide evidence that the APsyn waveform is a critical modulator of synaptic facilitation in excitatory nerve terminals and that further study of presynaptic K+ channels is warranted across neuronal cell types.”

“Our central finding is that an important mechanism of synaptic facilitation in excitatory hippocampal neurons is APsyn broadening. We find that the surprisingly rapid frequency-dependent broadening of APsyn is enabled by a unique molecular combination of Kv1.1/1.2 channels with the Kvβ1 subunit. Indeed, this small broadening of the APsyn mediated by Kvβ1 has a tremendous impact on synaptic transmission as the loss of the Kvβ1 subunit blocks synaptic facilitation even during paired-pulse stimulation without altering initial vesicle fusion (Figs. 46).”

“APSYN waveform serves as an important regulator of synaptic function.”

“adaptive plasticity: manipulations that increase presynaptic Ca2+ channel abundance and release probability result in a commensurate lowering of the APSYN peak and narrowing of the waveform, while manipulations that decrease presynaptic Ca2+ channel abundance do the opposite. “

In conclusion the sodium/potassium gradient in the pre-synapse changes the amplitude of the action potential which changes the strength of the post-synaptic activity including how many vesicles are activated. This process that I am outlining allows far more information to travel over each node (or neuron) in a neural network, from the dentrite/soma over the AP to the post synapse. Widespread MVR in neocortical synapses overcomes the information bottle neck imagined by the theory of synaptic unreliability.

So maybe we really need to think about modeling the information criteria of the brain starting at the neurotransmitter level and the ion level? If this hypothesis is right the ideas of simulating a human brain is even further off than the current leading consensus.

We may need to model the ionotropic brain. That means modeling how calcium and potassium ions, stored in the dendrite soma, are changing amplitude & frequency of the action potential, which changes the synaptic output, ie how many vesicles are activated, and what types of neurotransmitter are released downstream, which further alters the ionotropic balance of dendrites on the receiving end. If one neuron is connected to 200 downstream neurons, when that one neuron fires, with varying strength, varying amounts of vesicle activation, varying quantities and types of neurotransmitters, how are those 200 downstream neurons affected in a different way?

What do you think about the idea of MVP Multi vescular release changing the paradigm of synaptic unreliability, and what it means when the bottle neck of a single neuron is no longer a bottle neck?

Neurohaxor WebXR discord

Silicon Valley Global News

Silicon Valley Global News: Stories, Research, Advanced…

Silicon Valley Global News

Silicon Valley Global News: Stories, Research, Advanced Concepts RE: Virtual Reality, AR, WebXR, AI Semantic Segmentation on 3D volumetric data, Medical Imaging, Neuroscience, Brain Machine Interfaces, Blockchain, Cryptocurrency, Drones, Light Field Video, Homomorphic Encryption.

Written by

Silicon Valley Global News: VR, AR, WebXR, 3D Semantic Segmentation AI, Medical Imaging, Neuroscience, Brain Machine Interfaces, Light Field Video, Drones

Silicon Valley Global News

Silicon Valley Global News: Stories, Research, Advanced Concepts RE: Virtual Reality, AR, WebXR, AI Semantic Segmentation on 3D volumetric data, Medical Imaging, Neuroscience, Brain Machine Interfaces, Blockchain, Cryptocurrency, Drones, Light Field Video, Homomorphic Encryption.