LINCellularNeuroscience/VAME

spread out Latent space

ask5898 opened this issue · 1 comments

ask5898 commented

Hello VAME developers
Im training VAME on a set of 13 videos using individual parameterisation for each video. The latent space I get seems to be really spread out. This would imply that some motifs have no transition probability between them which seems counterintuitive. I have attached the image of my clustered latent space.
My other question is whether motifs for individual videos can be compared if I am using individual parameterisation?
UMAP-motif210617_KPPTN_denv1_2DLC_dlcrnetms5_aedesJun22shuffle1_72500_full
UMAP-motif201211_KPPTN_ctrl5_6DLC_dlcrnetms5_aedesJun22shuffle1_72500_full

pavolbauer commented

Hi Ali,

I think you got this wrong, its good if the space is spread out, that means that you indeed have some clusters / motifs in your distribution that can be regarded as discrete (uniform within). That doesn't imply that there is no transition probability between them, the current state still can jump between those clusters. I dont know your data but when looking at the motif videos do they make sense?

Best,
Pavol