This toolbox allows the user to find the best fit of the unknown free parameters for recorded traces and spike trains. It also supports simulation-based inference, where instead of point-estimated parameter values, a full posterior distribution over the parameters is computed.
By default, the toolbox supports a range of global derivative-free
optimization methods, that include popular methods for model fitting:
differential evolution, particle swarm optimization and covariance matrix
adaptation (provided by the
nevergrad, a gradient-free optimization
platform) as well as Bayesian optimization for black box functions (provided
scikit-optimize, a sequential model-based optimization library). On the
other hand, simulation-based inference is the process of finding parameters of
a simulator from observations by taking a Bayesian approach, in our case, via
sequential neural posterior estimation, likelihood estimation or ratio
estimation (provided by the
sbi), where neural densitiy estimator, a deep
neural network allowing probabilistic association between the data and
underlying parameter space, is trained. After the network is trained, the
approximated posterior distribution is available.
Just like Brian 2 simulator itself, the
brian2modelfitting toolbox is
designed to be easy to use and to save time through automatic parallelization
of the simulations using code generation.