Home

Welcome to the Sir Isaac wiki!

Sir Isaac : Efficient dynamical inference

Sir Isaac is a software package for efficiently inferring dynamical models from noisy and incomplete time series data. By striving for predictive rather than mechanistic accuracy, the software is able to start from minimal assumptions to produce a dynamical model that approximates training data, the complexity of which is adaptively adjusted as new data become available. The resulting models are interpretable and are able to predict responses to new dynamical perturbations.

The mathematical foundation of Sir Isaac consists of ordered hierarchies of dynamical systems that can grow to approximate arbitrary dynamics, including arbitrary nonlinearities and latent state variables. By limiting the model search space and relying on a variation of Bayesian Model Selection, Sir Isaac inherently minimizes overfitting and efficiently infers dynamical models even in challenging data situations.

In initial tests on noisy simulated data, Sir Isaac successfully discovered the second-order differential structure of Newton's equations of motion using fewer than 200 simulated observations. Recently, the software successfully inferred dynamical rules from experimental data describing behavioral dynamics of escape in C. elegans. The resulting dynamical model produced accurate out-of-sample predictions and hypotheses for neural circuits underlying the behavior.

Project Roadmap

Last updated August 2019

version 1.0 (6 months)

• test suite
• user manual
• standardized code documentation
• switch to python 3

version 1.5 (1 year)

• efficiency and optimization updates
• gpu support

version 2.0 (2 years)

• simplified user iterface through web portal

version 2+ (3 years)

• integrate model simplification schemes, e.g. MBAM (Manifold Boundary Approximation Method)
• integrate dimensionality reduction schemes to more efficiently deal with datasets of large dimension