Google Summer of Code 2021

Hi everyone,

I would like to greet everyone and introduce myself as this year’s GSoC student :slight_smile:
My name is Ante and I am currently a graduate student in computational bioelectromagnetics at the University of Split.
Even though my formal training is in electronic engineering at Uni of Split, and I did my masters at the department of electronic systems at Aalborg University, I decided to take a slightly different route after discovering that I was extremely interested in bioelectromagnetism and bioinformatics in general.

The theme of my project at GSoC is the integration of the model fitting toolbox with the sbi package to enable simulation-based inference on brian2's models.
Summary of the project is available here, and my progress will be available in the Projects section of brian2modelfitting.

I am looking forward for this summer of code and for being able to participate in such a wonderful open source community :slight_smile:


Hi @antelk,

Welcome to the forum!

I’m really looking forward to your results!

Last winter, I attempted to move from my standard tools for parameter fitting (various genetic algorithms) to SPI. It looked convenient to have a generative model after fitting instead of just a set of parameters that fit a neuron dynamics. The sbi worked pretty well for a small number of parameters (4-5), but it stuck in the real-world 24-parameter model. I documented this in the bug report here. I wonder whether it is a problem of the specific model or a general disadvantage of the inference-based fitting.

So, good luck with your GSoC. I think this community will appreciate if your report here results from any sbi benchmarks and tests.

1 Like

Hi @rth,

Thank you for the warm welcome, I really appreciate it :slight_smile:

In the issue you opened, you mentioned that the total size of the summary statistics vector is 242.
To the best of my knowledge, sequential neural posterior estimation (SNPE) algorithm implemented in sbi should have no problems with high-dimensional outputs, unlike classical likelihood-free inference approaches such as ABC, as described in (Gonçalves2020).
The authors claim that SNPE can even be applied to raw data without explicitly deriving any summary features.
Regarding the number of parameters, in the same paper, the full posterior distribution for the somatogastric ganglion over 31 parameter in total is given as an example of a real-world application of sbi.
The code associated with the mentioned example is available here, you can check it out and maybe figure out what’s going wrong in your case.
If I find out anything, I will let you now.

Also, I will try to create short blog posts of my progress each week and I will post a link in this thread as soon as the first post is up online.


Thank you, @antelk ! I’ll take a look. As jan-matthis@github suggested, I reduced the number of parameters to 4, and sbi performed very well. I could be any of my 24 parameters. However, as soon as I released more than 5-7 inference stuck. I think (entirely not sure) that combination Ca2+ dynamics, SK channel, L- and T-types calcium channels are the culprits. Specifically, T-channel with GHK equation on the board.

I’m looking forward to your blog post.

1 Like