Hi, CEBRA team! How can I compare the neural dynamic or latent/decoding differences between two group of mice?

[qweasdf1354]

Hi, CEBRA team! Thanks for great algorisms and analysis methods.

I currently reading the paper and the doc in the web. I wanna to use the great method to my project of two_photon + mice behaviour data. At this time point, I got the [Neuron * Time] (row * column) dataset and continuous behaviour data (from DLC and some user_defined characteristics like the state of mice using discrete variables like 0 or 1) from two group of mice (namely Control and Experiment). To be clarity, the behaviour paradigm is the same,and the only differences is the different group of mice.

I wanna to know how can I compare the neural dynamic or latent/decoding differences between two group of mice?

I have some idea for analysis the data. Could you help me to exam is the following idea work or correct? Thanks a lot!

There are two hypothesis I wanna to test from my data:

One is the control group of mice may decode the state of mice (user_defined characteristics using discrete variables like 0 or 1) correctly and the Experiment may decode less correctly.
I wanna to do the following step to test above hypothesis. 1. Train models per mice. 2. Decoding for every mice. 3. Compare the KNN scores differences between Control and Experiment mice using t-test or non-parametric test.

The other is both of group of mcie decode the state of mice correctly but the dynamic of neural activity is different.
I wanna to do the following step to test the this hypothesis. 1. Train models per mice as the same above. 2. compare the consistency using the method your provide to test the consistency between all mice from all group that may find the high consistency between the same group of mice and low consistency between the different of group of mice. 3. Compare the training loss between the following two model: one is train from Control group, one is train from Experiment group, one is train from both Control and Experiment group.

Thanks for your help! I am looking forward for your reply and appreciate for your help in advance.

[MMathisLab]

Thanks!

Aside from maybe other statistical approaches, all the above sounds fine to me.

You could additionally consider training with our multi-session (all control, all expt), then check decoding, consistency and goodness of fit (loss value).

[qweasdf1354]

Thanks!

[qweasdf1354]

Hi, CEBRA team! Thanks for your reply!
I have a question with multi-session training.
I read the doc and it says

For now, multi-session training can only handle a unique set of continuous labels. All other combinations will raise an error.

Is that possible, I use the continuous_label to represent the discrete label? I mean using the 0 and 1 in continuous_label as the input for the cebra model.

Because the 0 and 1 may be the most important feature of this experiment.

Thanks a lot for help!