VAE accuracy/confidence?

[jelrod14]

Hello,
I was just wondering if there was a way to pull a performance metric from the contour invariant VAE model after clustering that says how confident the model is in putting a particular USV into a certain cluster. For example, if my clustering yields 8 clusters, is there a range as to how "cluster 1" a call is and towards the end of that range things get more variable? If it's possible to know and pull this metric from deepsqueak that would be super helpful!
Thanks!

[DrCoffey]

Hey @jelrod14, that information is definitely available but not easily accessible. Line 72-77 of UnsupervisedClustering_Callback.m is:

[clustAssign,D] = knnsearch(C,data,'Distance','euclidean');

        %% Sort the calls by how close they are to the cluster center
        [~,idx] = sort(D);
        clustAssign = clustAssign(idx);
        ClusteringData = ClusteringData(idx,:);

D is the distance to the centroid of the assigned cluster. You could try to add D to the ClusteringData table and then it would get saved along with the clustering.

I won't have time to add this and test it etc for a little while, but if you change the code first you can always submit a pull request to get it added to the main code.

-Kevin