Reference

Synapses and modulators

How cells connect in Spike: static, depressing, plastic and dopamine-gated synapses, gap junctions, rate and mean-field connections, wiring modes, and the NE, ACh and DA neuromodulators.


A synapse is a connection from one cell to another. In the lab it is drawn as an axon: a line from one group of neurons to another. This page covers what an axon can be, what each setting means, and how neuromodulators change a circuit from the outside.

If a word is new, the first mention on each page explains it in one line. For the cells themselves see the parts bin. For the controls on screen see how to read the lab.

The basics of an axonLink to this section

You make a connection in one of two ways. In Connect mode, click a source group, then a target group. Or drag from the rim of a neuron disc onto another disc. Click the axon to open its inspector. The editor keeps one axon per ordered pair of groups, so a second drag between the same two groups selects the existing one.

SettingWhat it doesRange in the lab
WeightHow strong each connection is. The unit depends on the target cell (pA, nS or mV, see below)0 up to about 20 typical synapses of the target model
E / IFlips the sign of the weight. Positive weights excite, negative weights inhibit
DelayHow long a spike takes to arrive0.1 to 250 ms
SpreadGives every synapse of this axon its own random weight around the value above0 to 100 %
DeleteRemoves the axon

Excitatory vs inhibitory. An excitatory synapse pushes the target toward threshold. An inhibitory one pushes it away. In the circuit file the sign of the weight is the whole difference. A neuron group is coloured sienna (excitatory) or Prussian blue (inhibitory) by the sign of the weights it sends out. The E / I button on a neuron flips the sign of every axon that leaves it.

Weight. The weight slider is square-law: it has fine steps near zero and coarse steps at the top. You can also click the number next to the slider and type a value. Typed values are clamped to the slider range. The range is set by the target cell: twenty times its default input weight, and at least twice the largest weight the level's own reference circuit needs. If a level wants more strength than a slider offers, the verdict tells you to use more cells or another route instead.

Weight units by target cell. A synapse onto a current-based cell is in picoamps (pA). Onto a conductance-based cell it is in nanosiemens (nS). Onto a cell with delta synapses it is a voltage step in millivolts (mV). The Izhikevich cells also take a voltage step in mV. The parts bin lists the unit of every cell.

Hard limits. The circuit loader refuses a weight above 500 nS onto a conductance-based cell, or above 10000 pA onto a current-based cell. The reference circuits stay under 60 nS and 1500 pA. A cell can still blow up numerically inside these limits. The engine then stops the run and says "numerical instability"; reload the circuit to continue.

Delay. The default delay of a new axon is 2 ms. The delay slider is square-law too. One to three milliseconds fits a single synapse. A hundred or two hundred milliseconds stands for a whole chain of cells you do not draw, which some burst generators need.

Spread. With spread above 0, each synapse of the axon gets its own weight, drawn from a normal distribution whose standard deviation is that percentage of the weight. The draw is fixed by the circuit seed, so the same circuit always gets the same weights. Test x5 draws it again for every seed. A circuit that only works with identical weights fails here. This is the lab's way of asking whether your design is robust.

Synapse types at a glanceLink to this section

TypeNEST modelWhat it doesWhere you meet it
Staticstatic_synapseThe same strength for every spikeEverywhere. The default
Depressingtsodyks2_synapseGets weaker with use, recovers slowlyThe Aplysia chapter
Plastic (STDP)stdp_synapseLearns: strengthens when the sender fires just before the receiverWriting a memory with plasticity
Dopamine-gatedstdp_dopamine_synapseWeakens when its input was active and a dopamine source firesThe fly mushroom body chapter
Gap junctiongap_junctionElectrical coupling that synchronises two cellsSandbox and your own designs
Rate connectionrate_connection_instantaneous, rate_connection_delayedThe gain of one rate unit onto anotherThe Rate models chapter
Mean-field connectiondiffusion_connectionJ and K of the spiking network a Siegert unit stands forThe network and its mean field

In the lab the synapse type stepper appears in the axon inspector only where the chapter uses it: in the Aplysia levels (static or depressing), in the fly mushroom body levels (static or dopamine-gated), and in Writing a memory with plasticity (static or plastic). The sandbox and your own files can use any NEST connection model through the circuit file, see a circuit is data. Also available through the file are tsodyks_synapse (the older kinetic form; prefer the tsodyks2 model) and quantal_stp_synapse (stochastic release with short-term plasticity).

Static synapsesLink to this section

A static synapse keeps its weight. Every spike gives the target the same push. Its timing is set by the delay and by the target cell's synaptic time constant (see tau_syn_ex and tau_syn_in in the parts bin). Alpha and exponential shapes both rise or jump quickly, then fade over a few milliseconds.

Depressing and facilitating synapses (tsodyks2_synapse)Link to this section

Short-term plasticity is a change in a synapse's strength over milliseconds to seconds, caused by its own recent use. A synapse that is used hard runs short of transmitter, so it releases less each time. That is depression. It refills slowly. Some synapses do the opposite: use leaves calcium behind, and the next spike releases more. That is facilitation.

The model is Tsodyks-Markram. In the lab, a spike gives a current of weight * u * x, where x is the share of transmitter still available (starts at 1) and u is the share released by each spike.

SettingUnitRange in the labMeaning
Unone, 0 to 10.01 to 1The share of the available transmitter released by each spike. Larger means deeper depression
Recovery (tau_rec)ms50 to 5000How long the synapse takes to refill

Switching a synapse to depressing in the lab sets x to 1, tau_rec to 1000 ms and tau_fac to 0 (depression only). A non-zero tau_fac adds facilitation in files you write yourself.

Note: NEST starts the first spike of a connection from u = 0.5, not from U. Later spikes use U when tau_fac is 0. The lab sets u equal to U for you, so the first spike is as strong as it should be. If you write your own circuit file, set "u" equal to "U".

Depression is how the Aplysia chapter shows habituation: the sensory neurons keep firing at the same rate, but the connection to the motor neuron gets weaker with each touch. See campaign-bestiary.

Plastic synapses (stdp_synapse)Link to this section

Spike-timing-dependent plasticity (STDP) lets a connection learn from the order of spikes. If the sender fires just before the receiver, the connection strengthens. If it fires just after, the connection weakens. This is a form of Hebbian learning ("cells that fire together wire together"), with the arrow of time added.

In Writing a memory with plasticity you can turn a connection among the pattern's cells plastic. The settings:

SettingMeaningLab rangeStarting value
Learning rate (lambda)How much one spike pair changes the weight0.0001 to 0.030.0015
Weakening (alpha)How much stronger a late pairing weakens than an early one strengthens0.1 to 30.5
Ceiling (Wmax)The weight a connection can never exceed10 to 3000 pAthe larger of ten times the weight and 100 pA
tau_plusTime window of the pairingfile only20 ms

Only excitatory connections can be plastic. NEST's stdp_synapse is excitatory only, and the circuit loader rejects an inhibitory STDP synapse. So the inhibition between two patterns in the memory levels is built in, not learned.

Dopamine-gated synapses (stdp_dopamine_synapse)Link to this section

A three-factor rule changes a synapse only when three things line up: the sender was active, the receiver was active (or just was driven), and a dopamine signal arrives afterwards. It is how a reward or a punishment can teach a circuit.

In the fly mushroom body chapter the synapse is used as a depression. When the odour input was active and the dopamine source fires, the synapse loses strength. That is how pairing an odour with punishment teaches the fly to avoid it. The settings:

SettingMeaningLab range
Dopamine sourceA group of cells whose spikes release dopamine onto this synapseany group of the circuit
LearningHow much one pairing takes away. The lab sets the signs so that the synapse depresses0 to 3000
Trace (tau_c)How long a synapse stays eligible after its input fired5 to 500 ms (30 ms at first)
Max weightThe largest strength it can have1 to 10000

Behind the scenes, the engine creates one NEST volume_transmitter for each dopamine source and treats every spike of that group as a dopamine pulse. It also sets a private copy of the synapse model for each projection, because NEST cannot set these properties per connection. The engine's own DA level drives every transmitter too, at 100 Hz per unit of DA.

Note: The real mushroom-body rule needs no spike in the output cell. NEST's rule looks at the output cell's own spikes, so the odours must drive those cells during the pairing. This is a game choice. See what is real.

Gap junctionsLink to this section

A gap junction is an electrical synapse: a pore that joins two cells so their voltages pull together. It is instant (no delay) and works both ways. Gap junctions are a quick way to synchronise a cluster of pacemaker cells.

Only two cell models accept them: hh_psc_alpha_gap and hh_cond_beta_gap_traub. The rules, enforced when a circuit loads:

  • Both sides must be one of those two models.
  • The weight is a conductance in nS and must be above 0.
  • The delay is not allowed.
  • It is always symmetric. With one_to_one wiring the loader makes it symmetric. With all_to_all it is allowed only within one population. An edges list gets its mirror edge added.
  • A gap junction cannot be modulated by the neuromodulator table.

Note: The heart levels do not use gap junctions. They use excitatory synapses between the pacemaker cells as a stand-in for electrical coupling, and say so. The crab chapter also approximates the electrical coupling between AB and PD cells with fast reciprocal excitation.

Rate connectionsLink to this section

Rate units have no spikes. They are joined by rate connections: the weight is the gain of one unit onto another, and a negative weight inhibits. Rates are in arbitrary units. There are two flavours:

  • rate_connection_instantaneous: no delay. This is what the lab creates when you connect two rate units.
  • rate_connection_delayed: a delay is required. Available in circuit files.

The default weight of a new rate connection is 1 (excitatory) or -1 (inhibitory). A unit that excites itself needs allow_autapses, which the lab sets for a self-connection.

Any other synapse type between rate units, or a rate connection between spiking cells, is rejected when the circuit loads. The axon inspector of a rate connection shows the weight but not the delay or the spread.

Mean-field connections (diffusion_connection)Link to this section

The mean-field idea: in a large network each cell hears so many weak, independent inputs that you can describe the whole population by its average rate. A Siegert unit applies a formula to the mean and the variance of the input and gives the population rate.

A diffusion_connection joins one Siegert unit to another. Its weight is J: the size of one input's voltage step in mV, in the spiking network the unit stands for (negative for inhibition). A second number, K, is how many such inputs each cell gets. The lab shows K as a slider from 1 to 200 with the unit "per cell". The default weight of a new one is 0.1 mV (excitatory) or -0.1 mV (inhibitory).

The engine turns J and K into NEST's drift and diffusion factors: the mean input grows with J * K * rate and its variance with J^2 * K * rate. The result is in spikes per second. Delay and spread do not apply.

Wiring modes: which cells does an axon connect?Link to this section

When both groups have more than one neuron, the axon inspector shows a wiring stepper.

ModeWhat it connectsWhere it is offered
Every pair (all)Every cell of the sender to every cell of the targetAlways
Neuron k to k (segment)Neuron k of the sender to neuron k of the target, one synapse per segmentAlways
Random few inputs (random)Each target cell listens to just a fixed number of sender cells, picked at random. A fresh draw for every circuit seedFly mushroom body, Rate models and Attractor memory chapters

Segment wiring also has a shift stepper (-2 to +2). Neuron k of the sender connects to neuron k plus the shift of the target. It builds chains and travelling waves with a few clicks. In the fly compass chapter the ends of the chain are joined into a ring. Random wiring has an inputs per cell slider: 1 to 12, and up to 100 in the Rate models chapter.

Note: In the circuit file, "Every pair" is the rule all_to_all, "Neuron k to k" with no shift and equal group sizes is one_to_one, a shifted chain or a ring is an explicit list of edges, and "Random few inputs" is fixed_indegree. Any simulator reads these the same way.

NeuromodulationLink to this section

A neuromodulator does not make the next cell fire. It changes how cells and synapses respond. Acetylcholine slows the heart's pacemaker. Noradrenaline speeds it up. Serotonin strengthens a synapse. Dopamine gates learning. In Spike a modulator is a number that is the same everywhere in the circuit: a concentration, in arbitrary units, where 0 is baseline and 1 is strong.

There are three channels:

ChannelNameTypical role in the game
NENoradrenalineSympathetic nerves: speeds up the heart, excites the breathing burster. In the Aplysia levels the NE channel stands for serotonin (5-HT), and its dial says so
AChAcetylcholineVagus nerve: slows the heart, drives the gut's pacemaker
DADopamineGates learning in the dopamine-gated synapse

Where the level comes fromLink to this section

A modulator level is the sum of two things:

  1. An external level, set by the task or by you in the sandbox with the dials.
  2. Release pools. An output of the kind modulator turns the spikes of its neurons into a concentration. Each spike adds gain to a pool. The pool decays with time constant tau. At a steady rate f the level settles near gain * f * tau / 1000.

In the lab you turn a motor output into a releaser with the Rate readout / Releases modulator stepper. The heart levels, for example, read ACh (vagal) and NE (sympathetic) this way. See inputs and outputs.

SettingUnitLab rangeStarting value
ModulatorNE, ACh or DAthe one the level wants
Gainconcentration per spike0.001 to 0.50.05
Taums50 to 50001000

Example from the inspector: at gain 0.05 per spike, tau 1000 ms and 20 spikes per second, the level settles near 1.0.

What a modulator acts onLink to this section

A modulation row maps a modulator to one parameter. It can act on every neuron of one group, or on every connection of one axon. Two laws:

  • Add (gain): value = value + gain * level.
  • Scale (scale_per_unit): value = value * (1 + scale * level).

An optional min and max clamp the result. A row can target one group or one axon, and uses add or scale, never both. Rows are applied between steps, never in the middle of one.

Cells that follow modulators by themselves. The NESTML cells sa_node_pacemaker, butera_burster and icc_slow_wave have parameters called ne and ach. The engine adds an implicit row for each: NE to ne and ACh to ach, gain 1, minimum 0. They follow the circuit-wide levels with no setup. A circuit overrides this for one cell and one parameter by listing a row of its own. The baroreflex level, for example, maps ACh to ach with a gain of 0.5. Other cells ignore the modulators until you add a row.

In the lab. Select a neuron group and press Advanced. The section Responds to neuromodulators lists the rows for that group. A row has a modulator stepper, a parameter stepper, an add / scale stepper, a per unit slider and an X to remove it. Add row appends one. Cells that expose ne or ach offer those as the first parameters; every other parameter in the Advanced list can be modulated too. The lab warns when a model has no parameter a modulator can act on.

On a synapse. In the Aplysia levels, the axon inspector has a button Strengthen with serotonin (NE channel). It adds a row that acts on the axon: on U for a depressing synapse (add law), or on weight for a static one (scale law). A per unit slider (0 to 8) sets how strongly. Remove modulation deletes the row. This is the engine's way to make a sensitised reflex: shock the tail, release serotonin, and the sensory synapse gets stronger.

Synapse rows have rules. They are checked when the circuit loads:

  • A modulated weight is clamped to the target model's weight ceiling.
  • A gap junction cannot be modulated.
  • The weight of a plastic (STDP or dopamine-gated) synapse cannot be modulated: the learning rule owns it.
  • delay, synapse_model and similar fixed fields cannot be modulated.
  • Two axons that share the same sender, target and synapse model cannot both carry a row, since their connections cannot be told apart.

The modulator dialsLink to this section

In a level whose task declares modulators, gauges on the scopes side show the levels the circuit sees: read-only sliders with a chip for each channel (NE amber, ACh cyan, DA violet). In the Aplysia levels the NE chip reads 5-HT, serotonin. In the sandbox the dials are real sliders from 0 to 1, in steps of 0.01. You set a level by hand and watch what the cells do. This is the quickest way to learn what ACh or NE does to a pacemaker. See the sandbox.

Levels that declare modulators include Neuromodulation, the heart tasks, the baroreflex, fight or flight, the autonomic boss and the gut level (ACh). See campaign-body.

A circuit file exampleLink to this section

This fragment shows a depressing synapse, a modulation row that sensitises it, and a modulator output:

"projections": [
  { "pre": "S", "post": "M", "rule": "all_to_all",
    "synapse": { "synapse_model": "tsodyks2_synapse", "weight": 400.0, "delay": 2.0,
                 "U": 0.2, "u": 0.2, "x": 1.0, "tau_rec": 1000.0, "tau_fac": 0.0 } }
],
"outputs": [
  { "name": "serotonin", "kind": "modulator", "modulator": "NE", "neurons": [{ "pop": "F" }],
    "gain": 0.05, "tau_ms": 1000 }
],
"modulation": [
  { "modulator": "NE", "projection": 0, "param": "U", "gain": 3.0, "min": 0.0, "max": 1.0 }
]

The numbers here only show the format. They are not a solution to any level.

Something wrong or missing? Edit this page on GitHub (source) or open an issue.