A card from a free library — ask anything, no account, works offline. Every card carries its source.
Reservoir computing — a fixed substrate, a trained readout; our use of it
instruments
A reservoir computer never trains its dynamical system. A FIXED, random, high-dimensional substrate is driven by the input and only a LINEAR READOUT y = sum w_i x_i learns, by one closed-form regression (Jaeger 2001, echo state networks; Maass, Natschlaeger and Markram 2002, liquid state machines). The substrate may be hardware: a bucket of water solved XOR and spoken digits from a photograph of its ripples (Fernando and Sojakka 2003); silicon, optics and spintronics since (Tanaka et al. 2019). Of a 100-unit reservoir with one readout the trained weights are 1/101 of the whole - sealed on the reservoir_computing stick (run tick.py reservoir_computing), with the readout's arithmetic: it is a RIEMANN SUM, and zero times infinity is a number only because the width is bound to the count - the left-sum error 1/(2n) - 1/(6n^2), the midpoint error 1/(12n^2), the Darboux gap 1/n (the integral exists iff floor meets ceiling), the UNBOUND sum diverging, Monte Carlo at 1/sqrt(n). OUR USE: the keeping is the reservoir - 933,479 found, verified, cited and sealed cards that are never retrained - and the readout is the only thing that moves: the pheromone trail (one decaying number per card, 0.25 left after two thirty-day half-lives, 49.97 bytes a trail measured on the box), the faces' scope weights and the winning-ticket mask, the chains' priced cost bands, the trajectory of a path (Wilson floor, width 1/sqrt(n)), the physiology load tally (six-hour half-life), the quorum and the elders' counts; in the fuel layer a frozen base with a rank-16 adapter (1/272 trained, sealed). Every one of these fits, and a fit is efficient for one reason: training is a thin linear readout over a fixed substrate, never the substrate itself. The boundary: a model inside the answering loop would be retraining the reservoir, and an unbound weight is a hub. Rests on the instruments, the computer, the mind and the logarithm (fading memory is e^(-t/tau)); Matt's reservoir-and-manifold doctrine says the same in energy terms - the manifold is the readout, and it is linear.
- part of → The instruments — every verifier and validator, and where they join — a low-tech instrument: a bucket of water computes by structure; only the readout is traine
- connects at → Computer science — bits, logic, and what can be computed — a fixed substrate plus a trained linear readout is a complete computer (Jaeger 2001; Maass
- connects at → The logarithm - the instrument the joints share — the echo state property is fading memory e^(-t/tau); a half-life is ln 2 over the rate
- connects at → The mind — matter aware of itself — the liquid state machine models a cortical microcircuit as the reservoir, readouts as the
- connects at → The reservoir and the manifold — value is in the routing layer, not the engine — the reservoir and the manifold: the manifold (routing layer) is the readout, and it is lin
- connects at → Our place — a deterministic instrument you read, not a program you run — our place: the keeping is the reservoir, the faces and trails are the readout - each fit i
- connects at → Resonators — Helmholtz, the cavity, Fabry-Perot: one shape, a Q that fades — the ring-down is fading memory: a cavity with a delay loop or a feedback laser is a reserv
- connects at → Sorting algorithms — the roll of eighty-six and the bound none can beat — the trained readout is a weighted sum and then a sort: top-k by score is a partial sort (a
Is this card incomplete? Tell the library — it will call out for more ↗