Divergence of EM updates due to possible memory aliasing/overwrite

[ebattenberg]

Using the cilk backend, I've experienced some unexplained behavior when using the train method. Sometimes when using a large dataset N = 7.5x10^6, D = 13, the EM iterations result in nonsense sparse mean vectors with large covariances. But when making copy of the dataset (data = data.copy()) before training, the result converges to meaningful mean vectors as expected.

Steps to reproduce:

1. initialize parameters using sklearn's kmeans++
2. initialize gmm specializer with diag covariance, any number of components
3. set components to initialized parameters.
4. run gmm.train(data)
5. get crazy mean vectors.

If I add "data = data.copy()" or anything else that causes the memory to be copied to a new location before step 4, the strange behavior doesn't occur.
Also, when loading smaller datasets this does not occur.