Three reservoirs, same capacity, completely different anatomy
A physical reservoir is sold on one number: its total information processing capacity. Chen, Nishioka and colleagues show that this number hides where the computation actually happens, and that two reservoirs with near-identical totals can be anatomically unrecognizable. Their local capacity maps, tested on an echo state network, an ion-gating reservoir and a spin-wave reservoir, turn out to be useful: states selected by the map beat states selected by readout weight magnitude or at random, at equal size.
Source: Local information processing capacity: anatomy of computation in neuromorphic systems, arXiv:2611.05959, 2026-11-04. Primary source. Read the full HTML text of the preprint, all figures and methods.
What the work claims
This is a methods paper with a demonstrated result. The authors introduce local information processing capacity, abbreviated L-IPC, which assigns each individual dynamical state of a reservoir a profile: how much of its behavior is linear versus nonlinear, how much memory it carries, and how much of it is computationally irrelevant residue.1 The global IPC score of a reservoir, the figure everyone quotes, is just the sum over states. The claim is that the distribution underneath that sum is the physically meaningful object, and that it differs dramatically between systems whose global scores are comparable.
The second claim is practical. Because an L-IPC profile identifies which states carry task-relevant nonlinearity and memory, it can guide pruning: keep the states that matter for a given task, discard the rest. On the benchmark they report, this task-guided selection outperforms both magnitude-of-readout-weight selection and random selection at the same number of kept states, in systems whose computation is spatially or spectrally structured.1
How it works
The construction is a local decomposition of an existing global idea. Information processing capacity measures how well a reservoir's states reconstruct a basis of temporal, nonlinear functions of its input. L-IPC treats each dynamical state as a vector in an infinite-dimensional Hilbert space and projects that vector onto subspaces corresponding to nonlinearity orders (first, second, up to fourth order in this paper) and memory steps (zero to ten steps back). What remains after the projections is an explicit residual, called computational irrelevance: the part of that state's behavior that contributes nothing to any computation in the tested basis.1
Three physical systems are anatomized, each initially providing 120 states. The first is a conventional echo state network, a randomly connected recurrent neural network, where states are the unit activations. The second is an ion-gating reservoir: a graphene channel covered by ion gel, where ionic transport at the gel-graphene interface builds and relaxes an electric double layer, and that slow interfacial physics supplies both the nonlinearity and the memory; states are extracted by sampling electrode currents at several points within each input pulse. The third is a spin-wave reservoir: a yttrium iron garnet crystal patterned with coplanar waveguides under a bias field, where nonlinear spin-wave interaction and dispersive propagation do the computing; states are extracted by spectrally filtering the detected response and envelope-detecting each filtered band.1
The headline comparison scales each system up. The ion-gating reservoir is operated at eight input pulse intervals, from 80 microseconds down to 4 microseconds, expanding its state pool to 960. The spin-wave reservoir is read out through seven detector waveguides, giving 840 states. The echo state network is enlarged to 900 states. All three are then tested on NARMA tasks, the standard nonlinear autoregressive moving-average benchmarks that demand long memory and nonlinearity from a reservoir. For NARMA-10 with 160 states selected out of the ion-gating pool of 960, L-IPC-guided selection reaches a normalized mean square error of 0.112, against 0.145 for weight-based selection and 0.177 plus or minus 0.040 for random selection; the error using all 960 states is 0.088.1 The protocol uses 20,000 time steps per run, 1,000 discarded as washout, 12,000 for training and 7,000 for testing, with 20 independent repetitions for the random-selection statistics.
Where a skeptic should push
The load-bearing assumption is that a state's profile, computed against a specific polynomial and memory basis, is a stable property of the substrate rather than an artifact of how the states were defined. But the states themselves are a measurement choice: in the ion-gating reservoir they are time samples within a pulse, in the spin-wave reservoir they are frequency bands. The "anatomy" is co-produced by the device and the readout geometry, and the paper does not fully separate the two. A differently instrumented version of the same device could draw a different map. That does not make the map useless, but it bounds the claim: L-IPC anatomizes this device as wired, not the device in itself.
Second, the pruning benchmark is a family of synthetic tasks. NARMA is the field's standard yardstick, but a state that is irrelevant to NARMA-6, NARMA-8 and NARMA-10 may be exactly the state you need for a different task, and the paper's own cross-task analysis shows the selected set shifts between tasks. Pruning guided by one task is a bet that your deployment resembles your benchmark. Third, the demonstrated gains are largest in the structured systems; in the homogeneous echo state network, where computation is spread evenly, map-guided selection has little to grab onto, which is the honest null result and the authors report it as such.1 Finally, everything here is done with a linear readout and offline training; nothing yet demonstrates that the map survives online learning or hardware drift.
Mapping where computation lives in living tissue
For organoid intelligence, the black box is literal: you cannot open a neural organoid and label which cells do the computing, and the only windows are the electrodes or optical reporters you happened to build in. This paper hands the field a protocol for that situation. An organoid-electrode system is a physical reservoir whose "states" are exactly as arbitrary as the ones here: spike counts per electrode, filtered field components, bursts binned in time. L-IPC says the meaningful question is not "how much can this tissue compute" but "which of my measured states carry nonlinear, memory-bearing computation and which are residue". A capacity map of a given organoid preparation would tell an operator which electrodes sit on computationally live territory and which are reading out mostly noise, at commissioning time, without training a single task.
The opportunity is concrete. Long-lived organoid interfaces degrade: electrodes fail, tissue shifts, cells migrate. Today's implicit assumption is that all channels are exchangeable until proven otherwise. If computation in living tissue is as structured as it is in the ion-gating and spin-wave devices here, then a channel map could support deliberate redundancy planning, and task-aware selection across a superset of states, for example multiplexing stimulation rates the way the eight pulse intervals multiplex the ion-gating reservoir, could recover performance after channel loss without retraining everything. The pruning result also cuts the other way, and that is the threat: a map that labels regions of living tissue "computationally irrelevant" invites you to starve or deprioritize them, but irrelevance was defined relative to a task basis the experimenter chose. Tissue labelled residue under one protocol may be doing exactly the slow, diffuse processing that another task needs, and unlike an FPGA, you cannot re-synthesize the pruned part. Treating a capacity map of living tissue as a license to ignore the low-capacity majority would be the same mistake this paper warns about for global scores, repeated at finer resolution.
The bottom line
Established: on three physically distinct reservoirs, a state-resolved decomposition of information processing capacity reveals structure that the global score conceals, and task-guided state selection from that map measurably outperforms weight-based and random selection at fixed size, with the NARMA-10 numbers above as the headline evidence. Hypothesis, not yet established: that the same maps stay stable across tasks, readout geometries and time, and that they transfer to substrates as messy as living neural tissue. What would confirm it: a reproducible L-IPC map computed on a living neural interface, showing that map-guided electrode selection preserves task performance across weeks of drift. What would break it: evidence that the local profiles are dominated by instrumentation geometry rather than substrate dynamics, in which case the map measures the probe, not the tissue.
Frequently asked questions
What is local information processing capacity?
It is a per-state decomposition of a reservoir's computational ability. Each dynamical state is projected onto subspaces for nonlinearity order and memory length, and the leftover residual is labelled computational irrelevance. Summed over all states it reduces to the familiar global capacity score.
Which physical systems were tested?
An echo state network, an ion-gating reservoir built from graphene under ion gel whose electric-double-layer dynamics provide memory, and a spin-wave reservoir made from yttrium iron garnet with coplanar waveguide readouts.
How big is the pruning advantage?
For NARMA-10 with 160 of 960 states kept in the ion-gating reservoir, map-guided selection reached an NMSE of 0.112, versus 0.145 for selection by readout weight magnitude and 0.177 plus or minus 0.040 for random selection, with all 960 states giving 0.088.
Why does this matter for organoid intelligence?
Organoid interfaces can only measure the states their electrodes define. A capacity map computed at commissioning could show which channels sit on live computation and which are residue, informing redundancy and task selection before any training begins.
What is the main caveat?
The map is co-produced by the device and the measurement geometry, and irrelevance is defined relative to the tested task family. A region scored as irrelevant for one benchmark may be essential for another, which matters when the substrate is living tissue you cannot rebuild.
References
- J. Chen, D. Nishioka, Y. Song, R. Iguchi, S. Hikasa, H. Kitano, A. Hirose, T. Tsuchiya. Local information processing capacity: anatomy of computation in neuromorphic systems. arXiv:2611.05959. 2026. https://arxiv.org/abs/2611.05959. Accessed 2026-11-05.