There are two paths to doing research.
The bottom-up approach starts by learning the basics, through such things as reading and coursework, maybe online. A masters might lie on this path, but note that not all masters degrees have a research focus. Courses do give you the basics and reading can get you closer to research. But some papers are difficult without having the basics. Authors can make a lot of assumptions about the knowledge of their readers.
The top-down approach starts with a problem that you find interesting and is likely to be interesting to others and then do "just enough" reading to understand the problem and design and carry out a solution. Along with reading you can discuss your problem with others who might have the background you lack, provided you can connect with them and get them to talk to you. Looking for possible extensions to existing papers might be a way to find a problem.
The first is sort of breadth-first and the second is like depth-first searching for help.
But, everyone on the planet at the moment, along with all their siblings, seems to want to be studying and researching in ML at the moment. It is among the hottest of the hot topics currently. It will be hard, using either approach to be successful since the ideas are "out there" accessible to all those others. So, there is almost certainly a ton of parallel research going on on a very large number of approaches. Getting to the front of this pack is going to be challenging unless you already have gone some way on one of those two approaches.
You find collaborators at conferences pretty regularly, but you have to read the papers and meet the authors. You also need something to add if you want to get a collaboration going. The bottom up approach (via coursework) gives you access to professors who might give you a boost.
My advice is to give one or the other approaches a try and see what you can do. But, keep flexible. Don't commit too much time or intellectual effort to a too-narrow approach. Keep your options open. Either of the approaches gives you skills, not only in ML but also for what might be the next-big-thing.