Eren Metin Elçi writes:
I just watched this video the value of theory in applied fields (like statistics), it really resonated with my previous research experiences in statistical physics and on the interplay between randomised perfect sampling algorithms and Markov Chain mixing as well as my current perspective on the status quo of deep learning. . . .
So essentially in this post I give more evidence for [the] statements “simulations are not scalable but theory is scalable” and “theory scales” from different disciplines. . . .
The theory of finite size scaling in statistical physics: I devoted quite a significant amount of my PhD and post-doc research to finite size scaling, where I applied and checked the theory of finite size scaling for critical phenomena. In a nutshell, the theory of finite size scaling allows us to study the behaviour and infer properties of physical systems in thermodynamic limits (close to phase transitions) through simulating (sequences) of finite model systems. This is required, since our current computational methods are far from being, and probably will never be, able to simulate real physical systems. . . .
Here comes a question I have been thinking about for a while . . . is there a (universal) theory that can quantify how deep learning models behave on larger problem instances, based on results from sequences of smaller problem instances. As an example, how do we have to adapt a, say, convolutional neural network architecture and its hyperparameters to sequences of larger (unexplored) problem instances (e.g. increasing the resolution of colour fundus images for the diagnosis of diabetic retinopathy, see “Convolutional Neural Networks for Diabetic Retinopathy” [4]) in order to guarantee a fixed precision over the whole sequence of problem instances without the need of ad-hoc and manual adjustments to the architecture and hyperparameters for each new problem instance. A very early approach of a finite size scaling analysis of neural networks (admittedly for a rather simple “architecture”) can be found here [5]. An analogue to this, which just crossed my mind, is the study of Markov chain mixing times . . .
It’s so wonderful to learn about these examples where my work is inspiring young researchers to look at problems in new ways!