IGL Laboratory 2.0.1

THE RESPONSE FIELD

One node, in three dimensions

Starting renderer…

Drag to orbit · Scroll to zoom · Shift + drag to pan · Click to probe

Hover to inspect the surface. Click to set the live input.Sampled mesh

00011011Live

Zero → rise → fall → rise → plateau

Live node

X₁ 0.75X₂ 0.25

IGLP = 16

1.00

Normalized output1.0000II · rising

Show the calculation

Rounding follows JavaScript Math.round: half ties go toward +∞, including negative values.

Four-point truth table

Exact XOR

X₁ X₂	S	Y/P	Region

Select a row to probe that binary input pair.

Compare an experiment

Pin a numerical snapshot, edit the node, then compare its four outputs and target error.

A SMALL NODE. A RICH RESPONSE.

A five-minute exploration

Unlike a monotonic response, this activation can rise, fall, and rise again. Moving four weighted sums across those regions lets one node express all 16 two-input Boolean functions.

Compare input pairs 01 and 10 with 00 and 11. See how the falling segment returns the 11 output to zero.
Move the input pad. Values between the four Boolean corners reveal the entire response field.
Start at P = 4, then increase P. Inspect the cross-section and precision map to see rounding effects.
Try seeded candidate search from a randomized starting node. Compare local and multiscale search.

The activation

IS ≤ 00

II0 < S ≤ PS

IIIP < S ≤ 2P2P − S

IV2P < S ≤ 3PS − 2P

VS > 3PP

Normalize the output by dividing by P. The result Y/P stays between 0 and 1.

What “integer” means here

Quantize each input to Xₚ = round(PX). Round each parameter to an integer. Compute S = round((W₁X₁ₚ + W₂X₂ₚ)/P) + B, then apply the same activation.

JavaScript stores numbers in floating-point form. These bounded integer-valued calculations are exactly representable; this is not a native CPU integer-performance benchmark.

Reading the surface correctly

The surface connects sampled values. In INT mode, polygons approximate discontinuous steps; even P = 1024 does not imply a 1024 × 1024 mesh. Live values, truth rows and search scores use the numerical core. The precision map samples 65 × 65 cell centers to avoid alignment with power-of-two integer grids.

Controls and accessibility

On the focused 3D canvas: arrow keys orbit, + / − zoom and Home resets the camera. On the input pad: arrow keys adjust inputs. Drag the panel dividers, or focus a divider and use its arrow keys. Popups work with click, keyboard focus, and Escape.

Focus view expands the scene inside this tab. Fullscreen requests the browser’s fullscreen mode; Escape exits. Your browser controls whether a new-tab link opens a tab or a window.

Privacy and scope

The application makes no training requests to a server and includes no analytics. Model data is downloaded locally or stored in this browser only when you choose Save to browser. Appearance and layout settings are remembered when storage is available.

One-node demonstrations do not establish network accuracy, generalization, latency, power use, or training-speed advantages.

MLiglon research

U.S. Patents 12,242,946 B1 and 12,488,222 B1, as cited in the supplied laboratory.

Visit MLiglon ↗Detailed user guide ↗

IGL Laboratory

One node, in three dimensions

Quantization difference

Appearance & layout

3D display options

Save, share & export

Copy manually