WhoBPyT Architecture Plan

[JohnGriffiths]

WhoBPyT Architecture Planning

This thread will be used for discussions on two related things:

a) the whobpyt code base architecture, and

b) the general development model

There is a general consensus that progress so far has been hampered by lack of clarity on these two points.

The location of the current consensus on the whobpyt architecture is this file

Please read this document. The proximal result of discussions in this thread should be updates to this file.

Please submit PRs with updates to the document contents as we go.

[Andrew-Clappison]

Just as a start - Decisions to make on included features and other choices (which we need to decide is in/out of scope for base code V1 of WhoBPyT). This will influence architecture in terms of classes we need, and the expected input/output of these classes.

Applicable to all:

• Implement Option of batching
• GPU Support
• Optional Laplacian
• Boundary for fitted parameters
• Mean and variance of Priors
• Local Node Parameter Fitting
• What Objective Function Components
• "Sampling" in training function
• Boundary Conditions on State Variables
• fNIRS BOLD feature
• How to combine multiple objective function components
• Standard data structures/dimensions for input/output data
• Allowing use of other optimizers other than ADAM
• How we will represent simulated and empirical data internal to the class.
• What output we want by default

Jansen-Rit

• Delays
• How connections will be implemented

RWW

• Staying with RWW Parameters or with g_IE, g_EE, etc...
• Time Changes
• Method to fit simulation in memory by having forward run to get initial conditions then running in GPU
• Choise of solution to RWW firing rate equations unstable
• Whether we allow any additional tanh() to be optionally added
• Adding gammaI term and J_new term to RWW

Linear

• What options to give