mackelab/stg_energy

Repo for STG energy paper. Michael (@michaeldeistler) and Pedro (@ppjgoncalves).

Energy-efficient network activity from disparate circuit parameters

Purpose of this repostitory

This repository contains all code and data to create the figures in the paper Energy-efficient network activity from disparate circuit parameters by Deistler, Macke*, Goncalves* (2022). If you only want to use the machine learning tools developed in this work, see the sbi repository and the corresponding tutorial. If you only need the pyloric simulator, see this repo.

Installation

First, create a conda environment:

conda env create --file environment.yml
conda activate stg-energy

Then clone and install this package:

git clone [email protected]:mackelab/STG_energy.git
cd STG_energy
pip install -e .

Please note that all neural networks were trained on sbi v0.14.0. In v0.15.0, the training routine of sbi changed (z-score only using train data). Thus, training on a newer version give slightly different results.

Structure of this repository

Roughly, the workflow for this work can be divided into three sections: (1) Running the pyloric simulator for many parameter sets, (2) Training the neural density estimator to approximate the posterior and (3) Generating plots.

(1) is implemented in stg_energy/generate_data/simulate... and was run on a compute cluster with SLURM. (2) is implemented in stg_energy/generate_data/train... (3) is implemented in paper/

Commands to generate data and train the network

cd stg_energy/generate_data/simulate_11deg
python simulate_11deg.py
python 01_merge_simulated_data.py

python train_classifier.py

cd stg_energy/generate_data/simulate_11deg_R2
python simulate_11deg.py
python 01_merge_simulated_data.py

python train_classifier_R2.py

cd stg_energy/generate_data/simulate_11deg_R3
python simulate_11deg.py
python 01_merge_simulated_data.py

python train_flow_R3.py

Git LFS

To store the data files, we use Git LFS.

Citation

@article{deistler2022energy,
  title={Energy efficient network activity from disparate circuit parameters},
  author={Deistler, Michael and Macke, Jakob H and Gon{\c{c}}alves, Pedro J},
  journal={bioRxiv},
  pages={2021--07},
  year={2022},
  publisher={Cold Spring Harbor Laboratory}
}

Contact

If you have any questions regarding the code, please contact [email protected]

stg_energy's Issues

Name for effective prior

We need a name for the distribution implied by the classifier.

The classifier happens without knowledge of the observation x_o. It only looks at simulations. Hence, a good name might be:
simulation informed prior

We think that the name is great, but it's a bit wordy.

Other options:
effective prior / simulation prior / usable prior

Figure that shows how much energy is consumed by each conductance

This could be added in figure 2 or figure 3. It will be interesting because it is different from the Eigenvector obtained by the active subspace.

Understand effects for 4h

We will need panels that explain the conductances that are highlighted

Should we move the regression performance of the regression into the main text?

Merge figure 0 and figure 1

Storytelling: how do you build models???
Well, you search for parameters, but often they give bad results (prior samples)
Then, you find parameters that give realistic activity (simulation-informed prior)
Then, you look at data -> posterior

What do we mean by 27 degree robustness?

We have to show that it indeed is robust

Is it a problem that we do not correct for the rejection step?

in fig2, show more samples

To demonstrate that we have a HUGE database of models that fit the data, we could show many more samples than just 4

Figure 4 still has min and max labels

Reducing pairs gets you almost all the way? If yes would be cool.

Fig5_cc: Does optimizing pairs give you as much energy boost as optimizing all 8 parameters of the neuron?

visualize space after classifier

what are min and max

not label them as min and max but with values

Should we get rid of the eigenvalue spectra?

We could replace them with sketches of the regression nets in fig5

do we need the name active subspace?

People do not like it when one uses a new term for something simple. Hence, we should say something like:

We search for directions that correlate with energy. We find that it is one. We call this the active direction. Then cite active subspaces.

supersize 4f

Storytelling for temperature

We will fit to experimental data at 27 degree.

Then, we'll make the following observations:

1. For any parameter set, it is possible to find Q10 values that allow it to be robust and thus energy consumption is not really influenced.
2. we checked how energy efficiency at 11 is related to efficiency at 27 degree. We find that some models are equally efficient at 11 degree but very different at 27 degree. This adds yet another layer of complexity to identifying optimal solutions

How to best connect the active subspace figures in the text

Quantitativley show that postreior is different from the simulation informed prior

Inkscape edits after finishing figures

• contour lines for active subspace figure
• non-orthogonal plot

Illustrate the posterior within the active subspace

The active subspace only gives the dimension in which energy is expected to change. But it does not necessarily mean that the parameters can change along this direction. We should plot a contour of a posterior in the sketch of the active subspace.

Change colormap in fig4e

should be autumn for energy

How should we best visualize the "simulated but not close enough scenario (fig5b)"

This refers to figure5b: it looks a bit ugly right now with the white background and then the black part around it.

Should we just remove the black part (simulated but not close enough)?

We really need to redo figure4e

they are not orthogonal.

At least explain clearly

Oblique projections / pictorial

Decomposing things onto axes that are not orthogonal

Gradient in fig4e is the one shown in fig3

Murphy miller 2009. Trefethen might be an overkill (non-normal dynamics)

We can make this a feature and visualize this nicely

Units of energy in fig3 and fig4 are not yet muJ/s

/40000 as in fig2

Do we keep linear regression in supplementary or move it to the main manuscript?

Which trace do we use for the main manuscript?

In figure 3 (histograms) should we show the max and min trace, or the one were the spikes are more similar?

should 4f be a log-log plot?

Labels for fig2a

To Jakob, it seemed as if in fig2a (histogram figure) the second row would also be spikes. We need ylabels to show that it is energy

Wrong untis in figure 3a

In the correlation plot, we say that the unit is in /cm**2. But it's not because we have already multiplied by the area.

start_val is not always the same

The initialization for the gating variables has been 0.5 in my MSc thesis but is 0.0 for all energy experiments I have run so far.

Thus, the functions simulate and simulate_energyscape currently use 0.0, whereas simulate_general uses 0.5.

In fig4, what parameter set should panel h be?

Currently, we show a set with high energy per spike and high number of spikes. Maybe we can show one just in between f and g?

Try how good prediction performance is with linear regression

why do we have the density estimator???

we could not create 55,000 well-fitting models otherwise...

re-order summary stats

when I plot the summary stats (e.g. fig1 supp), we have to change the order of the names to match the new order of the summary stats.

Also, make sure that AppFig2 is good and summstats are ordered in the same way for panel a and b

What about the most efficient models at 11 degree?

Are they also robust?

Add figure for temperature robustness

Message:

enforcing temperature robustness does not change energy requirements.

Does the inverse hold?

Rotate labels and ticklabels of y-axis in fig4 by 180 degrees

Use PM, AB/PD, AB consistently

see e.g. fig5 where we use PM and AB

voltate in fig4 is 200 to 300

Explicitly make the point that we can predict energy

We should add a panel in fig3 that shows the regression net.

More thoughts:

• panel a into a 2d image plot?

Do we need a sample from the prior with NaN summstats?

label in 4e is confusing

we say 11 and 27 degree when, what we mean is "Posterior fit to 11 degree" and "Samples that are temperature robust". We should also add titles to the panels.

Should panel 4f be bigger?

We could get rid of the eigenvalue spectrum, and move 4f (scatterplot) to the bottom right.

Discussion -> tim oleary rule for being robust

correlations between parameters are captured -> suggesting that compensation between parameters

Do we want a figure that demonstrates the classifier results?

We could show something like:

1. Samples from prior
2. Samples from simulation-informed prior
3. Samples from posterior
4. Accepted samples

Also show the figure that compares accuracy in main text?

Additional figures

Two more figures would be possible:

1. Independence of energy consumption
2. Sample from the conditional and check if the pairs can get you "all" the way from least to most energy efficient.

A row that shows you can not change single parameters

For the conditional correlations figure, we should explicitly show that, by changing just one parameter, one can gain much less.

Somehow indicate the angle in fig4

Maybe add a small legend figure with just axes and angle

Should we add a figure that expicitly demonstrates the independence of neurons

We could also just refer to fig4b,d. There, it becomes clear that E/spike and #spikes only depend on parameters of the respective neuron. All others have almost no influence.

d and e are poorly visible in fig5_cc, panel c

two options:

1. make the dots and letters a bit bigger
2. add an arrow to the letters