Overview of Brian2

This page describes some of the concepts in Brian2 and how the different parts are meant to work together. Of course, nothing here is set in stone at the current stage...

Magic and clocks

For a more detailed description, see New magic and clock behaviour.

The clock system has been simplified and is now more explicit: The user either defines a Clock (and passes it for example to the NeuronGroup) or the default clock is used. Changing the dt of a Clock after the creation of a NeuronGroup or between runs is no longer a problem.

The “magic” system is also more stringent and easier to explain now: Brian keeps track of all instances that are created (they also have a unique name, all this is taken care of in BrianObject) – independent of the execution frame. All of these objects are used when using the magic run method. There are also less issues with circular references as for example synapses or monitors only store weak references to their targets and therefore do not keep them alive on their own.

Equations

For user-centered documentation see Equations, a short developer document is available as Equations and namespaces.

Equations objects are the core of model descriptions in NeuronGroup (and eventually Synapses). Internally, they are implemented as a list of SingleEquation objects (which should never be used by users directly), exposing additional information via properties (e.g. the names of all state variables, a dictionary of all differential equations, etc.). SingleEquation objects are basically a wrapper around a single line of equations, providing access to the expression, the unit, the type of equation, etc.

Equations do not deal with namespaces, they do not carry more information than the equation strings (this has the consequence for example concatenating Equations object is possible and does not cause any problems). They are also immutable, so using the same object in different groups does not lead to any problems, either. They offer the possibility to specify values or exchange variable names (this also increases backward-compatiblity), but this does nothing more than string replacements.

The Equations class performs only very generic checks of the equations (e.g. whether illegal names such as t are used for state variables, whether the unit definition is a valid unit, etc.) – it does not check for consistent units (because this requires knowledge of the external namespace, or the pre- and postsynaptic groups in the case of synapses) or correct flags (because they differ between NeuronGroup and Synapses, for example).

Variables and namespaces

Objects referring to variables and functions, in particular NeuronGroup and Synapses provide the context for resolving names in code strings. This is done via the Group.resolve() method that returns a Variable or Function object for a name. All internal names (state variables defined in the group and also all variables referenced by this group, e.g. the pre- and postsynaptic state variables in a Synapses object) are stored in the variables attribute which can be used as a dictionary but is actually a Variables object. Note that Variable objects only exist once for every variable, e.g. the Synapses class contains references to the Variable objects of the pre- and postsynaptic classes, not new objects

Groups can also specify a group-specific, explicit namespace that is stored in their namespace attribute, used for resolving external names. For more details on the resolution of names, see Equations and namespaces.

State Updaters

For a more detailed description, see State update.

State updaters convert equations into abstract code. Any function (or callable, in general) that is able to convert an Equations object into a string of abstract code can therefore be used as a state updater. Many state updaters can be described very easily by creating an ExplicitStateUpdater object with a textual description such as x_new = x + dt * f(x, t) (which should be understood as \(x_{t+dt} = x_t + dt \cdot f(x_t, t)\)).

The StateUpdateMethod class provides a mechanism for registering new state updaters, registered state updaters will be considered when no state updater is explicitly specified. State updaters expose their capabilities via a can_integrate method, that specifies whether they are able to integrate the given equations or not (for example, if a state updater does not support stochastic equations but the equations are stochastic). The order of the registration also provides an implicit priority: If not state updater is specified explicitly, the first from the list that claims to be able to integrate the equations will be chosen.

Code objects and code generation

The actual computations during a simulation – the state update, the threshold testing, the reset code – are performed by CodeObject objects. A group such as NeuronGroup creates the code objects for a certain target language (at the moment: Python or C++), providing the abstract code, the specifiers and the namespace to the code generation module. The CodeObject objects are executed every timestep and either directly update state variables (state update, reset) or return a result of their computation (threshold).