Reservoir computing in three parts
I've drawn it the way the 2019 review you'll find below describes it.
Input
Whatever data you've got, turned into signals the reservoir can take in.
Reservoir
A tangle of loops that remembers the last few inputs for a while. You don't touch it.
Readout
Usually just a linear regression on top, and it's the only bit you'd train.
The idea in three parts
If you haven't met it before, I'd picture it in three pieces (there's a sketch above). Your input goes into the reservoir first, and that's just a big network full of feedback loops. Each new input stirs it up in a nonlinear way while it still carries some trace of whatever came in a moment ago (the trace fades after a while). Then there's a readout that looks at what the reservoir is doing and learns to turn that into an answer.
What I like about it is that you never train the reservoir. Gouhei Tanaka and his co-authors say as much in their 2019 review in Neural Networks, where the reservoir stays fixed and you only fit the readout with something like linear regression. Training is quick and cheap because of that, and the review gives it as the main reason you'd pick this over other recurrent networks.
Where it came from
It grew out of two ideas that turned up at about the same time. In 2002 Wolfgang Maass, Thomas Natschläger and Henry Markram described the liquid state machine in Neural Computation. They were trying to explain how circuits of spiking neurons in the brain could handle a stream of changing input in real time, and their answer was that a big enough, varied enough circuit works as a fading memory that a readout can learn from.
The other branch is the echo state network from Herbert Jaeger. The paper most people cite is Jaeger and Harald Haas in Science in 2004. On a standard test of predicting a chaotic time series they reported accuracy better by a factor of 2,400 than earlier methods, and on cleaning up a wireless signal they cut the error rate by two orders of magnitude. The Tanaka review treats both as the roots of what's now called reservoir computing.
Why physical reservoirs work
Since you never adjust the reservoir, it doesn't have to be software. Anything with rich enough dynamics might do, as long as you can feed it inputs and read its state. The 2019 review is mostly about these physical reservoirs (built from different materials and devices), and it sorts them by what the reservoir is made of.
I think this is what makes it interesting for wetware. With living neurons you can't set the weights or write a program for them (there's no compiler, as I keep saying on the CL1 page). Reservoir computing doesn't ask you to. You only need the cells to respond in a rich and repeatable way, and then you train the readout on an ordinary computer.
A living reservoir: Brainoware
The best-known example is Brainoware, from Indiana University in 2023. The team used a human brain organoid on an electrode array as the reservoir, and after two days of training it picked which of 8 speakers was saying a vowel 78% of the time. An ordinary AI system still did better on the same task, so I'd read it as a proof of concept.
The weak point, as far as I can see, is the "repeatable" part. FinalSpark's paper on its Neuroplatform says organoid activity drifts from day to day and the current needed to make one fire goes up as it ages. A reservoir that changes under you means the readout you trained last week may not fit any more. If you want the wider picture of computing with cells, I've covered it in what is biocomputing.
Questions
What is reservoir computing in simple terms?
I'd say it's machine learning where you leave the big network alone. You run your data through a fixed network that remembers recent inputs for a while, and you only fit a simple readout on top (often a linear regression), so it learns quickly.
Who invented reservoir computing?
There isn't one inventor as far as I know. It grew out of the liquid state machine from Maass, Natschläger and Markram in 2002 and Herbert Jaeger's echo state networks (his 2004 paper in Science with Harald Haas is the one you'll see cited).
Can living neurons be used for reservoir computing?
They have been, in small tests. Brainoware used a human brain organoid as the reservoir in 2023 and got 8 speakers right 78% of the time, but I'd point out that a normal AI system still beat it.
Sources
- Maass, Natschläger and Markram, Real-time computing without stable states: a new framework for neural computation based on perturbations, Neural Computation, 2002
- Jaeger and Haas, the echo state network paper on chaotic time series and wireless channels, Science, 2004
- Tanaka et al., Recent advances in physical reservoir computing: a review, Neural Networks, 2019
- Cai et al., Brain organoid reservoir computing for artificial intelligence, Nature Electronics, 2023
- Jordan et al., Open and remotely accessible Neuroplatform for research in wetware computing, Frontiers in Artificial Intelligence, 2024