In plain language
A multiplies each input by a , adds a , and applies an . A can strengthen an input or reverse its effect.
02 / The building block
One small decision.
Inputs carry information. decide how much it matters. Adjust them to see a compute its output.
The neuron workbench
Adjust weights. Inspect the signal.Input x₁
Input x₂
Select a node to understand its role. No training happens when you move these sliders.
Differentiable neuron demonstrationForward pass · weights edited manually
The technical idea
The original uses a hard threshold and a mistake-driven learning rule. A is a differentiable alternative; it is not identical to the historical .
From data to proteins
In a simple protein classifier, two inputs could encode normalized hydrophobicity and charge. The numbers in this exhibit are generic , not a validated biological prediction.
Look deeper: mathematics, methods & limitations
z = w₁x₁ + w₂x₂ + b; a = f(z)
A positive increases z when its input increases. A negative decreases z. The shifts the response even when both inputs are zero.
Threshold outputs 1 for z ≥ 0 and 0 otherwise. maps z to (0,1). returns max(0,z), so its output is not inherently a probability.
uses the chain rule to compute derivatives through differentiable operations. A hard step is not useful for ordinary -based ; the original uses a different update rule.