Analysis

Overview

Info

Estimated time for section: 30 min

The following section details downloading the results from Google Cloud, and then running several analysis scripts.

Downloading Data From GCP

Once LFADS w/ PBT has finished running, we can now download the data back to our local computer for analysis.

Warning

At this point, you can stop all your machines (server and clients). Information on stopping machines can be found here

To download data from the bucket, navigate back to https://console.cloud.google.com/storage.

Click on the newly created zip file, and then click 'Download.' The output of the run should now be downloaded to your local computer.

Post Processing

Now that you have downloaded the data back to your local computer, open up the tutorial package.

Open up pbt_plot.m in Matlab. First, set the tutorial_package to your current working folder.

Then, inside the pbt_plot.m script, we need to set the data_folder variable to the location of your pbt_run folder, which is located inside the run folder you downloaded. For instance, on my computer its C:\\Users\tutorial\output\runs\pbt_run.

Set the output_folder variable to the folder where you want the plots generated.

Run the pbt_plot.m script. This script will show the evolution of HPs over successive generations. These are some of the plots we got from the tutorial run.

Application-specific analyses

Spiking dataset: compare to true rates

Note

This section is only for users using a synthetic spiking dataset (as in the tutorial_package) where ground truth firing rates are available.

If you used the synthetic dataset derived from Lorenz, then we can compare to the true rates.

First, open the spikes/compare_rates.m script in the tutorial package. Fill in the first lfads_output_dir with the address of the lfads_output which is located inside your run folder. For instance, on my computer its C:\\Users\tutorial\output\runs\lfads_output.

Then, you can run the script, which will generate R^2 value, which represents the error in the inferred rates compared to the true rates, as well as plot the inferred rates of several example neurons against their true underlying rates.

The following plot was generated from this tutorial's run.

EMG dataset:

Note

This section is only for the specific EMG dataset in the tutorial_package

If you used the locomotion EMG dataset provided in the tutorial package, then we can visualize the output of the AutoLFADS model and also perform a joint angular acceleration decoding analysis similar to what is presented in Wimalasena et al., 2021.

First, open the emg/emg_post_analysis.py script in the tutorial package. - Change the data_dir with the address of the lfads_input which is located in your run folder - Change the lfads_output_dir with the address of the lfads_output (also in run) For instance, on my computer its C:\\Users\tutorial\output\runs\lfads_output.

Then you can run the script. AutoLFADS inferred means from the training and validation set used for training will be plotted for 4 example EMG channels. Additionally, an optimal linear estimator will be fit to map the AutoLFADS output onto joint angular accelerations of the hip, knee, and ankle. Cross-validated R^2 calues are included in the title of each subplot.

The following plots will be generated from this script in the tutorial.

Below, is a plot of the single-trial inferred muscle activations for four muscles from a trained AutoLFADS model on the tutorial EMG dataset. Output has been split into two columns based on whether the EMG data was included in the training or validation set used during AutoLFADS training.

Below, is a plot of the decoded single-trial joint angular acceleration predictions from a trained AutoLFADS model on the tutorial EMG dataset. True behavior is shown in black, predictions overlaid in blue. Cross-validated R^2 value is included in the title.

Calcium dataset: map to true latent states

Note

This section is only for users using a synthetic calcium dataset (as in the tutorial_package) where ground truth latent states are available.

If you used the synthetic calcium dataset derived from Lorenz, then we can measure how well RADICaL predicts the true latent states.

First, open the calcium/map_to_latents.m script in the tutorial package. Fill in the first lfads_output_dir with the address of the lfads_output which is located inside your run folder. For instance, on my computer its C:\\Users\tutorial\output\runs\lfads_output.

Then, you can run the script. RADICaL inferred rates are mapped to the ground truth lorenz states by training a ridge regression. The script will then generate R^2 (coefficient of determination) value, which represents how well the predicted latent states match to the true latent states for the testing trials, as well as plot the predicted latent states of several example conditions against their ground truth latent states.

The following plot was generated from this tutorial's run.