Why don't we treat the mean and variances in EM algorithm as latent variables I know how the Expectation Maximization works. What I fail to understand is why only the mixture components are treated as latent variables and why not the mean and variances values of the K gaussians?
I read a note saying latent variables are something that we don't observe.
Here in EM, we don't observe the mean and variances either, right? 
 A: That's because they do not belong to the same class of objects. In the classical description of the EM algorithm, the latent (also called hidden) variables are random variables, while the means and variances of each Gaussian are parameters (values that are used to describe a probability density).
The goal of the EM algo is to infer the value of parameters (i.e. the mean and variance of each Gaussian in your example), which is done in the M-step. Hidden variables are not a final product of the algorithm, but are merely used as a tool to infer the value of the parameters. Computing their distribution in the E-step is a required step before the M-step.
However, I agree that, in some settings, it might not be obvious to separate what is a "hidden variable" and what is a "parameter" that needs to be estimated. Typically, in this paper, they infer not only the parameters of a mixture of distributions, but also compute an estimate of which component has generated each data point.
Okada, Makoto, Kenji Yamanishi, and Naoki Masuda. "Long-tailed distributions of inter-event times as mixtures of exponential distributions." arXiv preprint arXiv:1905.00699 (2019).
Besides, keep in mind that the EM algo is not only applicable to mixtures of Gaussians: it can be used to infer the parameters of many different statistical models (mixtures of Gaussians are only one classical example). So, depending on your problem, "using the mean and variance of the K components" may not make sense. For instance, in the paper I mentioned above, they use the EM algo to infer the parameters of mixtures of exponential distributions.
If you are looking for a very detailed and clear derivation of the EM algo, I strongly recommend Andrew Ng's notes on the subject.
