Fluid Morphic Computing

A connection is two spheres touching — no wires, no routing, one leak

1 · The gig

A gig at night. Beside the mixing board sits the block, one row of spheres per mic. The crowd mic's spheres swell, brush and fall back; the band's rows, driven again before they can shrink, press together all the way across, and with the engineer's hands off, the faders ride on what the block lets through. When the singer stops between lines her row leaks apart and her fader dips; when she sings again both come back.

2 · The crossing

Later that night, the same block runs a traffic light. A car stops at the red. From inside the signal head, one car crossing on its green reaches in and leaks away, while the waiting car's row keeps being driven until it joins across and the light turns green. Then the head turns to its cross-street face: two cars queue, their row builds and joins, and the light turns for them — the same block runs the light both ways.

What this is

A computer whose wiring is not drawn but grown. Fluid Morphic Computing places blank spheres at fixed points in a volume; an input swells the spheres it drives, and two spheres are connected exactly while they touch — while their radii span the distance between their centres. The depth of the overlap is the strength of the path. There is no weight table, no address and no router: one leak shrinks every radius back toward blank, so whatever is not driven again comes apart, and forgetting and pruning are the same act. What lasts is what keeps being driven. The two films show the same block doing two jobs: riding a gig's faders, and running a street corner's traffic light.

The technical shape

Synapse: r₁ + r₂ > d — two spheres touching is the connection; nothing else allocates one
State: radii only · centres fixed · n·|C| scalars generate the weights instead of n² stored ones
Strength: overlap depth δ = r₁ + r₂ − d sets conductance
Plasticity: one leak — dr/dt = α·s − λ(r − r₀); forgetting and pruning are the same act
Wiring: none inside the block — drive enters one face, current is read at the other
Stage: concept specification · working technical note · reference simulator

The deep end · full technical outline

Outline v2 · expanding

Connection by contact

Each channel at a node is a sphere of radius r ≥ r₀ at a fixed centre; at rest no sphere reaches its neighbour. Channels a at i and b at j form a contact if and only if r_i,a + r_j,b > d_ij. The contact is not allocated, routed or named: it appears when the inequality becomes true and vanishes when it becomes false. That inequality is the entire synapse condition.

Strength is depth

While two spheres overlap, the crater depth δ = r_i + r_j − d sets a conductance through a monotone map — g₀·δ, or g₀·(1 − e^(−κδ)). Deeper interpenetration lowers resistance across that interface. The map is a modelling choice; the claim of the architecture is only that strength is a function of overlap, never a stored scalar that can be written to directly.

One leak, every job

There is no Hebbian update, no STDP table, no prune-and-regrow scheduler. Drive grows a radius at rate α; when drive stops, the radius falls back toward r₀ at rate λ, and any contact it was holding comes apart. Forgetting and structural pruning are one process. Reinforcement is simply drive that returns before the leak has emptied the radius.

Persistence beats a shock

A drive that repeats faster than the leak accumulates depth; a single spike can open a deep contact for a moment and then lose it. In a population, coordinated repeating drive builds a connected path while scattered drive makes brief, shallow touches. That is the films' whole plot: the band holds its faders up and the crowd does not; a waiting car turns the light and one passing car does not.

What it does not do

The update rule has no pairwise term: each radius integrates only its own drive, so two channels meet when they are individually loud and already near each other — co-location, not correlation. Without a correlational term or an embedding whose distances already encode the input's statistics, credit assignment is undefined. Whether placement alone can recover correlation is the central open problem.

Neighbours and the open part

The closest prior art is activity-dependent neurite outgrowth (van Ooyen and van Pelt, 1994–95), where neurons connect while their fields overlap — with the sign reversed and as a model of development, not a computer. The predicate itself is the Boolean model of continuum percolation; FMC is a driven, dissipative version of it. No physical sphere computer has been built, and the materials problem is unsolved.

Related in the lab

Morphological computing

Built at this bench

The machine, the instrument, the film that explains it — all designed and built in-house. The same bench works for hire.

Services →