Signal Lab — a hand-tunable neural network

Operation

Each operation keeps its own weights.

Training

lr 1.4

Weight compression

bits

full precision — no compression

Edit the network

drag a wire up/down → weight
drag a node up/down → bias
click an input → flip its bit

Network tanh · tanh · sigmoid

2-bit A, 2-bit B in → 4 bits out

Performance

	A + B
loss (this input)	—
accuracy (all 16)	—
precision	full

What you're seeing. A real multilayer network (4 → 4 → 4 → 4) with tanh hidden units and a sigmoid output; a bit reads as 1 once its output passes 0.5. The answer chips show the correct bits, so you can see the gap.

Addition is the one you can realistically reach by hand — drag the output-bit wires and biases until the answer chips turn green.
Multiplication and division are nonlinear. The network can learn them to 100% (only 16 cases to fit), but you'll almost never find the weights by hand — that's what Train is for. One press gets most of the way; press again to converge.
Train addition to 100%, then switch operation: the wires that solved + are useless for ×. Each task needs its own shape.
Compression rounds every weight to a coarse grid. Slide it down and watch the wires snap to fewer thicknesses while accuracy falls. Training still runs in full precision; the slider squeezes the result afterward.