Hi Cebra team, does this model accept multiple discrete variables?

[XuanMabard]

our dataset has behaviors code and position on the screen(upper-left, lower_left) as discrete variables. we are wondering if the model is capable of consume multiple discrete variables as input along with neural data?

[stes]

Hi @XuanMabard , thanks for asking. Yes, you can feed in discrete variables to the model. I would still recommend to cast everything to a float before feeding it into the model.

[XuanMabard]

Hi Cebra team,
If we are using the solution one you mentioned above, we have two questions: 1) whether this make any sense that if, in our case, the discrete labels are categorical variables with no inherent ordering. 2)would this reduce the importance of the categorical data, because it's lower dimensionality compare to the neural data.

now if we implement option two, this might not be the best for our goals, because our goal is to identify latent variables that represent each of the categorical variables independent to the others? also, Is the requirement of having only one discrete variable a fundamental limitation of the algorithm?

We were thinking one possibility, is to train separate model for each discrete variables, do you think this make sense?

Thank you so much for your follow-up answers!

[MMathisLab]

You can train independent models to compare the goodness of fit, but you can also train models with >1 discrete variable (correct me if I'm wrong @stes )

[XuanMabard]

Hi cebra team,
I'm really interested in how many discrete variables the model can take, i would love to hear more from you.
thank you so much!

[vince-jq-sun]

I have the same question. It seems that recode discrete 2D behavioral variables as 1D changes the meaning of the 2D variables. Is there any way to indeed feed the training with a 2D discrete variable?

My current solution is to dummy code each value in the 2D discrete variables as continuous discrete variables of higher dimension?.For example, if my discrete variables have two columns; each column has two values, can I re-code them as a 4D continuous vairable (each columns represent the presence of one value of one variable using 1, and absence of that value as 0).
Is such approach appropriate? Is there a direct way to input a 2D discrete variable without flattening it?

[MMathisLab]

as stated, you can test the code here: #135