Let’s consider first a simple rate-coded neuron of the leaky-integrator type, which simply integrates the weighted sum of its excitatory inputs:

\[ \begin{align}\begin{aligned}\begin{split}\tau \frac{d \text{mp}(t)}{dt} &= ( B - \text{mp}(t) ) + \sum_{i}^{\text{exc}} \text{r}_{i} * w_{i} \\\end{split}\\\text{r}(t) & = ( \text{mp}(t) )^+\end{aligned}\end{align} \]

where \(mp(t)\) represents the membrane potential of the neuron, \(\tau\) the time constant of the neuron, \(B\) its baseline firing rate, \(\text{r}(t)\) its instantaneous firing rate, \(i\) an index over all excitatory synapses of this neuron, \(w_i\) the efficiency of the synapse with the pre-synaptic neuron of firing rate \(\text{r}_{i}\).

It can be implemented in the ANNarchy framework with:

```
LeakyIntegratorNeuron = Neuron(
parameters="""
tau = 10.0
baseline = -0.2
""",
equations = """
tau * dmp/dt + mp = baseline + sum(exc)
r = pos(mp)
"""
)
```

The only required variable is `r`

, which represents the instantaneous firing rate and will be used to propagate activity in the network. All other parameters and variables are freely decided by the user. More detailed examples can be found in the section Examples.

Custom functions can also be defined when creating the Neuron type and used inside the `equations`

field:

```
LeakyIntegratorNeuron = Neuron(
parameters="""
tau = 10.0
baseline = -0.2
""",
equations = """
tau * dmp/dt + mp = baseline + sum(exc)
r = sigmoid(mp)
""",
functions = """
sigmoid(x) = 1.0 / (1.0 + exp(-x))
"""
)
```

Make sure that the name of the function does not overlap with existing mathematical functions (cos, exp), existing variables (tau, r) or built-in functions (pos, t, dt).

The ODE can depend on other parameters of the neuron (e.g. `r`

depends on `mp`

), but not on unknown names. ANNarchy already defines the following variables and parameters for a neuron:

- variable
*t*: time in milliseconds elapsed since the creation of the network. - parameter
*dt*: the discretization step, default is 1 ms.

The `sum(target)`

term gives a direct access to the weighted sum of all inputs to the neuron having the corresponding target (see Projections to see how the target is defined). These inputs are organized in a data structure called `Dendrite`

.

It is possible to modify how weighted sums are computed when creating a Rate-coded synapses.

**Note:** The connection type, e.g. *exc* or *inh*, needs to match with the names used as a `target`

parameter when creating a `Projection`

. If such a projection does not exist when the network is compiled, the weighted sum will be set to 0.0 for all neurons.

Using only `sum()`

in the equations sums over all defined targets. For example, if two projections with targets `"exc"`

and `"inh"`

reach a neuron, `sum()`

is equivalent to `sum(exc) + sum(inh)`

. Inhibitory weights must then be defined as negative.

One has the possibility to use global operations on the population inside the neuron definition, such as the maximal activity in the population. One only needs to use one of the following operations:

`min(v)`

for the minimum: \(\min_i v_i\),`max(v)`

for the maximum: \(\max_i v_i\),`mean(v)`

for the mean: \(\frac{1}{N} \sum_i v_i\),`norm1(v)`

for the L1-norm: \(\frac{1}{N} \sum_i |v_i|\),`norm2(v)`

for the L2-norm: \(\frac{1}{N} \sum_i v_i^2\)

Example where neurons react to their inputs only where they exceed the mean over the population:

```
WTANeuron = Neuron(
parameters="""
tau = 10.0
""",
equations = """
input = sum(exc)
tau * dr/dt + r = pos(input - mean(input))
"""
)
```

Note

The global operations are computed using values at the previous time step (like weighted sums), not in the step currently evaluated. There is therefore implicitely a delay of `dt`

, but it cannot be changed.