Multicompartmental modelling with PyNN: design goals

The three principal aims of the PyNN project are:

1. to make it easy to run models on different simulators (and on neuromorphic hardware), in order to facilitate cross-checking, reproducibility and re-use;

2. to provide a simulator-independent platform on which to build an ecosystem of tools for computational neuroscience (visualization libraries, workflow engines, etc.) rather than having tools that only work with one simulator;

3. to support modelling at a high-level of abstraction (populations of neurons, layers, columns and the connections between them) while still allowing access to the details of individual neurons and synapses when required.

The scope of PyNN was originally limited to networks of point neurons (integrate-and-fire and related models). The primary reason for this was that at the time only the NEURON simulator had both support for multicompartment models and a Python interface.

This situation has now changed, with the release of Brian 2, the addition of Python support to MOOSE, development of the Arbor simulation library, and support for multicompartment models in the future versions of the BrainScaleS and SpiNNaker neuromorphic chips.

We are therefore adapting the PyNN API to support both point neuron models and morphologically-and-biophysically-detailed neuron models (and mixed networks of both model types). The principal design goals are:

1. maintain the same main level of abstraction: populations of neurons and the sets of connections between populations (projections);

2. backwards compatibility (point neuron models created with PyNN 1.0 (not yet released) or later should work with no, or minimal, changes);

3. integrate with other open-source simulation tools and standards (e.g. NeuroML) wherever possible, rather than reinventing the wheel;

4. support neuromorphic hardware systems.