How does a neuron decide when to send a signal?

[kolleamm]

In my quest to build better AI, I've been researching a lot on how the human brain works.
I understand the overall structure of the neuron and how it communicates with other neurons.

The only thing I don't understand is how a neuron decides when to send out a signal/response to other neurons?

Is there some sort of algorithm written in the DNA inside the neuron's nucleus?

Having that knowledge, and from what I'm understanding the complete map of a fruit fly neural network already having been mapped, we could have some really interesting insight on how the mind works.

 

[RPinPA]

If the sum of the inputs from the other neurons exceeds some threshold, it fires. There's no algorithm.

As I recall, some inputs are negative, i.e., some connections act as inhibitors to firing. So then the rule would be if the sum of the positive inputs minus the negative inputs exceeds some threshold.

Disclaimer: I'm not a biologist, and I studied biophysics including neurophysiology many years ago. But what you want to investigate is "neurotransmitters", which are the chemicals that carry the signal from the output of one neuron to the input of another.

 

[osilmag]

To add to what RPinPA wrote: When the EPSPs from all the dendrites sum up at the axon hillock above a certain threshold, an action potential will be generated to flow along the axon.

 

[kolleamm]

Thank you all, it's nice learning something new everyday :D

 

[BillTre]

For many neurons, the responses above about inputs from other cells summing together to exceed a threshold level at the axon hillock (about where the axon leaves the cell body) are a good description. However, there is a huge diversity of different kinds of neurons in the nervous system, especially if you consider invertebrate nervous systems. So, its a bit more complicated.

There are neurons that never make spikes (action potentials), yet they can also pass signals to other neurons. Generally they will be smaller neurons that only project axons to nearby cells (they are called local neurons as opposed to projection neurons that project their axons to other areas of the nervous system). Since they don't have far to go, they can get away without using action potential. Instead of just having the on or off type effect on the pre-synaptic axon terminal (like an action potential) where it makes a synapse with another downstream neuron (or other cell), it can produce a more graded output of neurotransmitter release.
In either case (action potential or graded potential) the amount of depolarization that is seen at the site of the synapse (usually the same amount for an action potential or differing amounts for graded potentials) will determine how much calcium is let into the cell by voltage sensitive calcium channels in the axon terminal. The amount of calcium entering the axon terminal will determine the amount of transmitter released.

Alternatively, electrical synapses (basically little holes connecting the insides of the two cells) will drive current from pre-synaptic to post-synaptic cell directly, based upon the voltage differences between the two cells.

Normally I would put a bunch of links in here to wikipedia, but where I at the moment has very poor internet so I you might want to look up several of the terms I am using to get a fuller understanding of their meaning.

 

[atyy]

When there is an action potential ("spike") in a neuron, the neuron releases neurotransmitter which signals to other neurons.

A pretty realistic way of modelling when a spike occurs is via the Hodgin-Huxley equations. There is no fixed threshold above which the neuron spikes in such models.
Eg: Hodgkin-Huxley model https://neuronaldynamics.epfl.ch/online/Ch2.S2.html

A less realistic way, but still accurate enough for many cases, is that the neuron spikes when the membrane potential crosses a fixed threshold.
Eg. Integrate-and-fire neuron https://neuronaldynamics.epfl.ch/online/Ch1.S3.html

In artificial neural networks, the spiking of the neuron is not modeled. What is important is that the output of the neuron is a nonlinear transformation of the summed inputs. The Hodgkin-Huxley models and the integrate-and-fire models have in common the nonlinear transformation of the summed inputs.
Eg. Perceptron https://towardsdatascience.com/what-the-hell-is-perceptron-626217814f53
Eg. Nonlinear activation function in artificial neural networks https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6

 

[Laroxe]

Just to perhaps complicate it a little more,
The brain has two interconnected messenger systems, the nervous system and the endocrine system, the functional unit of the nervous system is the neurone.

A neurone can be thought of as requiring a certain level of activation in order to cause it to fire and potentially effect the action of other neurones, this is its action potential. This action potential is not fixed and is acted upon directly from other neurones and indirectly by the neurochemical environment it occupies. All of these influences potentially change the level of stimulation needed to fire, many of the chemicals that effect the neurone indirectly are referred to as neuromodulators while those that act directly are usually called neurotransmitters, some hormones can function as both.

A neurone in the brain receives signals, potentially from thousands of other neurones, some will be activation signals others will be inhibitory when the activation signals exceed the inhibitory signals and reach the threshold needed the neurone fires. The act of firing itself alters the threshold needed for the next action potential.

I have read recently work that suggests that different groups of dendrites on the receiving neurone can be weighted differently in terms of causing an action potential and that the neuron firing is not a simple all or nothing response. I haven't found the references for these on a quick search but I could try harder if its of interest.
Generally I think its unlikely that the activity of individual neurones will really help in trying to understand how the brain works, basically they don't do much of anything on their own, its making sense of how networks of neurones work that we need to get to grips with.

 

[BillTre]

Generally I think its unlikely that the activity of individual neurones will really help in trying to understand how the brain works, basically they don't do much of anything on their own, its making sense of how networks of neurones work that we need to get to grips with.

There is an interesting transition point in neuroanatomy with respect to size and cell populations.
Smaller organisms have correspondingly smaller numbers of cells in their nervous systems.
Eventually as things get reduced, their nervous systems will have the same or similar parts of the nervous system but number of neurons per part will reduce to one.
This results in individually identifiable neurons in the nervous system, rather that a cell being a member of a population, which may or may not have some other cryptic parameter that could affect cell properties such as position in a field of cells.

Many invertebrates (insects, worms, crustaceans, molluscs) have significant parts of their nervous systems composed of individually identifiable neurons. Other nervous system parts have populations of similar appearing neurons, such as in visual or olfactory systems.
There is a lot of neuro-behavioral work done done on these simple systems because of their research advantages (always being able to return to the exact same cell in an animal when trying to replicate some finding).

Most vertebrates are larger and have in almost all cases no individually identifiable neurons. They have populations of neurons where single cells would have resided in smaller organisms.
The zebrafish embryo/larvae however, which starts out rather small, is an example of a vertebrate with several individually identifiable neurons in their hindbrain.
Some are shown in the third row of this figure (from here:)

top row, picture of embryos of the stage imaged
2nd row Location of nerve roots and nuclei = green
3rd and 4th row named cells are individually IDed, o and s1 are landmarks (ear and first somite), RMO44 and 3A10 are antibodies used in labeling

The larvae (hatched) is less that a mm across and 3-6 mm long (depending on stage). The nervous system of a 20 hour old embryo is less that 200 µm (10^-6m) wide.
As the fish grows, an entire embryo could be contained in the volume of the adult brain.
A lot of stuff is added to the brain, including many cells. A few individual cells can still be found but they are also clouds of more anonymous cells around them, presumably with variants of their properties. Here is a possible example of what this might look like.

One might expect behaviors due to populations of cells to emerge in parallel with the increase of the cell populations.