search

rossfadely / mofa

Mixture of Factor Analyzers
MIT License
5 stars 2 forks source link

Track down bug which causes logdet to go negative for certain initializations #9

Closed rossfadely closed 12 years ago

rossfadely commented 12 years ago

The culprit seems to be the psi initializations - too large of variance causes the initial cov to have negative logdet sometimes.

Setting the initialization of the psis to 0.5 * datavar seems to be ok for 100 seeds..

davidwhogg commented 12 years ago

oh I understand why that is! Remember me saying to Dilip "why is this positive definite?" It relies on the psis being not-insane -- same with the lambdas. In general the safe move is to initialize everything to a small variance -- maybe even set by the smallest eigenvalue of the empirical data variance tensor....

rossfadely commented 12 years ago

see initialization thread