On doing (Machine Learning) research without a supervisor

In two words: it sucks.
In more words: it sucks majorly.

Due to funny circumstances of life, I have been out of academia for a while and I'm forced to do research without a supervisor that is immersed in the field. Even worse, nobody around me is doing AI / ML research.

This manifests into a feeling of being lost in an alien jungle: is that ivy poisonous? Is this leaf good to eat?
Is this paper I'm reading now relevant to my interests in the long term, or is it something rehashed from a ten year old idea?
Is this idea I've had interesting enough to work into a fully fledged paper, or will the reviewers laugh at my quaint approach and light their cigars with the printout?
Is this research problem I've identified something I can tackle as it is, or will it take an army of von Neumann-Feynman hybrid clone grad students 100 years to crack?

I simply have no "feel" for the granularity of open problems in the field.

The field is immense, and every day something new appears. I cannot try everything in every direction, of course. But even the areas that I am interested in (feature selection, cluster analysis, deep learning, and such), which are but a subset of the whole, are still immense and feel like a glass wall on which I cannot find any imperfections to grab on to and climb.

I feel like a child who's been given Thoman Mann's "The Magic Mountain" and was told "well, you can learn how to read by looking at the signs on these pages".
Sure I can, but it will take me years. A supervised, guided learning of the alphabet would be much more efficient.

And to everyone who is working under supervision and is in a hurry to tell me that "this is what research is like" and "this is how everyone struggles to find open research problems" -- no, it's NOT.
Research is not done in a vacuum, but in groups and with supervision. That way, you can stand on the shoulders of giants (or a big pile of dwarfs), instead of merely wasting years reinventing some inefficient wheel.

EDIT:  Well, at least a bad ML pun came out of my rant, if nothing else -- I am doing "unsupervised research" :D