Reservoir Computing is the notion of using complex, nonlinear, mostly deterministic but borderline chaotic systems to obtain solutions to concrete well-defined problems. Roughly the idea consists of using an approach similar to Machine Learning: use a dataset, feed the inputs to the reservoir computer, pick some outputs from it to feed to a regular ANN. The weights of the readout ANN are trained to match the dataset targets.

At first learning about this concept, it struck me as roughly equivalent to a shaman making predictions watching sticks fall on the floor, and I want to formalize what I mean exactly by this: consider the shaman as the reservoir computer, using his auditory subsystem in order to hear regular town's folk problems, loosely coordinating the shaking of the sticks with the inputs before throwing them, and using visual interpretative subsystems in order to make useful predictions. As the shaman student trains for years, he might train the equivalent of the readout ANN using actual outcomes from the world.

My question is: Is my analogy faulty in some fundamental way? or are shamans poorly specified reservoir computers?

Note that my outsider perspective of the whole subject of reservoir computing stricks me as wishful thinking, and I'm surprised that the subject doesn't seem to be considered more controversial or suspect

  • $\begingroup$ Note to moderators: I couldn't find a good tag for this question, I tried "exotic computing", "nonstandard computing" but none of these seem to exist $\endgroup$
    – lurscher
    Commented May 8 at 15:09

1 Answer 1


Your analogy seems valid to me.

Reservoir computing strikes me as a high likelihood of being more or less baloney, or at least adding little of real use in the long run.

Of course there could be instantiations that are valid (but limited in value), and instantiations that are total nonsense. It is such a broad and vague concept that one cannot say that it is categorically useless, but if it is useful, I expect will be in a narrow and limited and unexciting way, not in the broad breakthrough that the high-level description might mislead unwary readers into expecting.

Don't assume that you'll be able to tell what is considered controversial or suspect. I suspect many experts might consider reservoir computing suspect, at best. There are many concepts in computer science that are trendy and get some attention but (in my opinion) are basically baloney, yet are rarely publicly labelled as such anywhere that a member of the public can readily find. Why? Perhaps because most researchers are polite. Or because most researchers prefer to focus their attention where they can make constructive progress, rather than on "calling out" pointless nonsense. Or perhaps because it only becomes baloney if one makes claims about it, and many authors either avoid making claims (by using big-sounding words in a vague, handwavy way) or use it in very limited ways that are valid but far less than what the broader marketing might lead one to expect (an instance of the motte-and-bailey technique, arguably).

Also keep in mind that science can be subject to "marketing", where some stuff gets a lot of attention because it has been marketed in a way that seems exciting (at least to non-experts) and grabs the attention of an audience. In computer science, attention from glossy prestige journals like Nature or Phys. Review isn't necessarily well-correlated to what's likely to have true lasting value, as they seem to be swayed by such marketing.

In research, there's a lot of crap out there. You might know Sturgeon's law: 90% of everything is crap -- and that applies to science, too. You have to know how to find the good stuff. Experts in the area typically know how to judge work, but as a non-expert, it's harder. Some heuristics are to look at work that is published at top-tier conferences in the discipline (e.g., for ML, places like NeurIPS, ICML, ICLR, etc.), published by researchers at top-tier research institutions, and that is getting cited by others in top-tier conferences or from top-tier institutions. Those are highly imperfect and unreliable proxies for the quality of work, but I'm hard-pressed to suggest a better way that a non-expert can evaluate the quality of deeply technical work.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.