Shared neural cells

Signal Bloom

cell12→8→4 MLP×field76×76

World

Grid side Boundary

Playback speed 24 t/s

Deterministic seed

Cell network

Architecture12 perception → 8 hidden → 4 state One shared residual micro-network is evaluated independently inside every cell.

Residual step size 0.28 Neighbor coupling 1.52 Cell heterogeneity 0.20 Intrinsic noise 0.002

Floating synapses

Every number below is a signed, non-integer influence from one neighbor direction. Plastic mode gives every cell its own evolving copy.

NW N NE W

CELL

E SW S SE

Local plasticity 0.000 Return-to-base decay 0.030

Seed, layers & brush

Initializer Seed density 18%

Brush

Channel Operation

Brush value 1.00 Radius 2 Soft edge 0%

View & data

Palette Display channel

Zoom grid Probe selected synapses

Experiment files

NEURALIS field manual

Two architectures, one interactive medium

Neural cells

Each grid location owns a continuous four-channel state. A shared 12→8→4 micro-network reads the cell, weighted neighbor aggregates, local texture, and cell-specific traits. Every directed neighbor edge has a real-valued signed weight; plastic presets update those weights locally while the world runs.

Learned weighted rule

Draw or choose an input/target pair, then train. The optimizer unrolls the same local 3×3 neural rule for several cellular steps and backpropagates through the whole trajectory. The learned object is therefore an iterative local physics, not a network that directly paints the final target.

Pointer controls

Input	Action
Left drag	Paint the configured channel or layer
Right drag	Erase regardless of brush operation
Middle drag / Alt-drag	Pan the field
Wheel	Cursor-centered zoom
Hover	Move the live microscope probe

Keyboard

Spaceplay / pauseNone tickRresetGtrain / stop1–4change viewCcycle channelPcontrols panelOinspectorHomefit viewFfullscreen?this help

Suggested experiment

Open Plastic Wiring, switch the view to Synapses, and select a bright cell. Increase local plasticity slowly while watching its eight directed weights diverge. Then switch to Learned rule → Translate East, train for a few hundred epochs, and inspect the learned signed 3×3 kernels.