Neural-control family: what deep learning and control enables in the real world
gshi.me
gshi.me
However, I don't quite understand how the authors are encoding "domain invariance" with "a domain adversarially invariant meta-learning algorithm." I'm not sure what that means. If any of the authors are on HN, a more concrete explanation of such "domain invariance encoding" would be greatly appreciated!
Finally, I have to say: The field of deep learning and AI is going to benefit enormously from the involvement of more people with strong backgrounds in physics, specially the theorists who have invested many years or decades of their lives thinking about and figuring out how to model complicated physical systems.
Domain invariance is the ability for a model to deal with multiple problem domains (e.g. wind conditions, type of drone, etc).
Adversarial training means teaching the system to deal with different domains by deliberately giving it the hardest possible examples of different domains - where these difficult examples are in fact typically learned by looking at the gradients from the main algorithm and seeing what would cause it the most problems, and then giving it that problematic domain and forcing it to behave well. The math is similar to GANs where you have multiple neural nets effectively fighting against each other to achieve some desired outcome.
Meta-learning is where you use ML to learn something about how to train your model. This can take many forms. Sometimes this means learning how an optimizer should work. Sometimes it’s a lot like fine-tuning where the meta-learner learns a kind of base model that can easily be adapted to new situations. I’d guess in this situation it’s the latter.
do you have any good references where you think this is done particularly well?
For example, see http://www.stochasticlifestyle.com/the-use-and-practice-of-s...
Doing the behavior that feedback driven control-systems do but even better is a nice and impressive applications. That seems most useful for applications like the application that's being described - swarms of flying drones. Flying generally already yielded to various control system - autopilots work because the skies are mostly empty and so your system working according to your predictions is all that matters. A drone swarm is much more complicated but is still under the system's control.
It's worth saying that the "real world" where a lot of robots fail has different challenges. Whether you're talking self-driving cars, robot dogs accompanying troops or wheeled delivery robots in hospitals, the problem is figuring both what you're looking at and how to respond to it. And this has the problem that nearly anything can show up and require unique responses, causing progress here to never quite be enough. And better physics and better cooperation between controlled elements doesn't seem that useful here and this approach might not help this "real world".
In controlled environments with well-known responses they can probably work (not sure why not go with traditional control approaches though), but really don’t see DNN working outside that.
A DNN cannot guarantee that its predictions respect momentum balance. By proper training you can just only reduce the probability of violating it.
It is plainly a wrong tool for the task.