Meta-Gradients in Reinforcement Learning

 Some person or entity named Wiki P Edia has the following thoughts on this whole 'meta' concept.  

Any subject can be said to have a metatheory, a theoretical consideration of its properties, such as its foundations, methods, form and utility, on a higher level of abstraction. In a rule-based system, a metarule is a rule governing the application of other rules, and "metaprogramming" is writing programs that manipulate programs. 

So extrapolate as you wish when thinking about meta-gradients or meta-learning in deep learning neural nets and in reinforcement learning systems (because they are somewhat different).

I like the concept that machines that learn should learn how to train themselves more efficiently.


Popular posts from this blog

Simulating the Universe with Machine Learning

CycleGAN: a GAN architecture for learning unpaired image to image transformations

Pix2Pix: a GAN architecture for image to image transformation