ilabcode/hierarchicalgaussianfiltering.jl

The Julia implementation of the generalised hierarchical Gaussian filter

hierarchicalgaussianfiltering.jl's Issues

Drag 'n Drop HGF specification

set_state function

Make presentable table with variables, notation and interpretations

Make node-specific update equations

Optimization

remove prediciton step from categorical state node

Make no HGF version compatible with Julia 1.7

fix get_prediction for binary and categorical nodes

make sure that they update predictions form their parents first, recursively

Symbol for predicted volatility

Choose something else than nu

Testing suite

We should use a shared testing suite with the Python toolbox for the update equations and whole model fitting.

• Single update equations
• Utility functions

Canonical tests

• Noisy binary HGF
• Categorical HGF
• State transition HGF

make get_surprise general

optional: provide a list of node names

Nodes in update order not recognized as HGF nodes

Example of update order: update_order = ["u", "x1", "x2"], returns error.
The nodes are recognized as a string and not an HGF node

Timed inputs

change coupling names

value -> mean
volatility -> variance

Enhanced HGF

Make alternative set of update equations

Update equations for multiple parents

Visualizations of HGF structures

Of generative models

of simple belief updating structures

of full message passing structures

Nonlinear transformations between nodes

Preprint: Julia packages software paper

Shared parameters

Plot function categorical HGF

Drift parameter

The Rho drift parameter

Pretty printing of nodes

Autoregression HGF

Autoregression HGF, where drift cannot be applied

More RejectParameter checks for negative precisions

Change names of variables:

evolution rate -> volatility
evolution rate for input nodes -> input noise
auxiliary prediction precision -> volatility-weighted precision

Decide on update equations for multiple parents and multiple children

input precision greek notation

input precision is an omega in continuous input nodes, and is a pi in binary nodes.
And it is currently called en evolution rate in the input node as well.
How should we name these things?

previous parent predictions: not agnostic, but based on parent priors

categorical HGF: use get_prediction for initialization

Minimize stored information

Don't store anything in nodes that is not necessary for computations at next timestep.
Create functions for calculating these values if necessary

Don't save history as default

Code Style

Make sure that all code style is to best Julia practice standards

find_optimal_parameters

Function for finding the HGF parameters that minimizes surprise
can be done by gradient descent, sampling, variational bayes etc

Automatize simple update order

And provide a warning if the update order is non-trivial

Appreciation and Suggestion for HierarchicalGaussianFiltering.jl

I recently read the paper "The Generalized Hierarchical Gaussian Filter" and found it an exciting extension of the Hierarchical Gaussian Filtering (HGF) framework. I appreciate your efforts and the paper's potential applications in cognitive neuroscience and computational concepts of mental disorders.

I want to bring to your attention some prior works such as "Online Message Passing-based Inference in the Hierarchical Gaussian Filter" and "The Switching Hierarchical Gaussian Filter," and also "Bayesian joint state and parameter tracking in autoregressive models" by Senoz et al seem to be relevant to this research. It might be helpful to consider these works in the context of HierarchicalGaussianFiltering.jl and include them in the documentation or examples to provide additional user insights.

Most components for implementing the generalized HGF are already available in RxInfer.jl, which treats everything as message-passing algorithms. In the near future, we will probably assign someone from the BIASlab to implement the generalized HGF within RxInfer.jl toolbox. In fact, RxInfer.jl has already some vanilla implementations of online inference MP in HGF .

Thank you once again for your valuable contributions.

Fix README

• PkgEval badge
• Deps badge
• Version badge
• General setup

Add custom node

Within-trial time course

Update equations for multiple children

Vectorize Turing

Turing might need to use vectorization for speed.

Check / remove coupling strengths from input nodes

Plot HGF structure

TagBot trigger issue

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

If you'd like for me to do this for you, comment TagBot fix on this issue.
I'll open a PR within a few hours, please be patient!

Add premade models

Premade HGFs

• Kalman filter HGF (Continuous 1-level)
• HGF with filtering of category means

Action models

• Confidence HGF
• Linear regression

Children learning which parents re reliable

transforms the uncertainty for binary state nodes when plotting

Input transform

Make optional mapping for inputs before they are given to the HGF.
For example state transition sets to the encoding that works for the state transition HGF

proper get_prediction function for online inference

initial/ first predictions from binary children should be drawn from the parents whose predictions comes from parents ect.

Current agnostic solution (mainly concerned categorical HGF) is to initialize the prediction in the binary state nodes. Therefore, parents initial predictions need to be equal.

Documentation

Some examples of nice documentation in related packages:

• https://probml.github.io/dynamax/
• https://gpjax.readthedocs.io/en/latest/

• The GitHub action needs to be fixed.

README:

• A logo
• A paragraph describing the toolbox
• Figures of the premande HGFs
• A hello world example
• PkgEval badge
• Deps badge
• Version badge

User guide:

• Theory module with figures
• API examples
• Add the equations to the docstrings

Notebooks:

• Multilevel fitting

• Categorical HGF

• Kalman filter HGF

• Add to Documenter REPO

Control nodes

Nodes for dynamically changing states and parameters of the HGF based on context or actions