inm-6 / nnmt Goto Github PK

Firing rate integration might not converge to the correct fixed point. Maybe try with different starting points. Check for hardcoded values.

Right now we show a plot of power spectra as an example in the README.
The application area of this tool would be more intuitively understandable with a more detailed example figure, e.g., something similar to Fig. 1 of Bos et al. (2016).

• replace frequency omega by complex frequency lambda in function arguments?
• generalization for spatial networks
• theory for single neurons in network (in contrast to populations), e.g., distribution of firing rates?

The h5py_wrapper is no longer updated regularly and causes a lot of deprecation warnings. We should think about replacing it.

Currently it is quite difficult to figure out which tools the toolbox contains and how they are depending on each other. Maybe one could add some kind of table of visualization to improve this.

• main documentation in README.md
  • structure (visualization: DAG?), usage and further development
• provide references or equation for implementations, detailed docstrings
• maybe use Sphinx

• simple plotting options for analytical results (e.g., power spectra, transfer function, sensitivity measure)
• just correct axes labels

Currently acknowledgments and authors are still duplicated on the top level of the repository and in the documentation. The old readme should also be removed as soon as all content is in the new documentation.

Hi, this looks like an amazing tool and I am interested in using it for my own work. But I am primarily concerned with mean field statistics of rate model with ReLu like nonlinearity. From a quick look, it seems to me that there is no direct support for it, but I want to check about it here. Thanks!

As described in the title. This leads to ugly code in nnmt.lif.exp._sensitivity_measure.

Show error and where it comes from to the user. Enable the user to set/change the accuracy where possible. Compare with firing rate integration in #7.

The implementation of the error function integration in the firing rates used to be numerically unstable. This was improved by replacing _Phi by _Phi_neg and _Phi_pos. But the function _derivative_of_firing_rates_wrt_mean_input is still using the old version. This probably could be solved in the same way as for the firing rates.

Hi,

Would it be possible to include a lognormal delay distribution? Not sure how analytically tractable that would be.

Best,
Aitor

Add a test incorporating the 'old' method for calculating the firing rates used in Hahne, Dahmen and Schuecker 2017.

• environment.yaml with needed Python packages
• microcircuit example in Snakefile

Currently, many functions imply LIF neurons and parameters as used in the microcircuit model.
We should think about how to disentangle main functionality and specific (derived) parameters.

Running test_reproduce_Bos2016.py takes very long, which probably is due to calculating the transfer function for many different frequencies. This could be solved by reducing the number of analyzed frequencies to a subset of the original ones.

Some unit tests are not encapsulated well enough. We should go through them once again and make sure that each tests stands for itself.

In the coverage report, the targeted code lines 244/245 in input_output.py are marked as missing. Depending on which error the h5py_wrapper throws it could rather be OSError instead of IOError.

Anyways, I think, it would be nice to print a user warning at this point, but this is an issue for another pull request.

Originally posted by @essink in INM-6/lif_meanfield_tools#48 (comment)

• check that US English is used consistently (in particular in README)
• run autopep8 on Python code