KATÁBASIS · the descent · how systems find their best by going downhill · kept by VIRGIL (the guide of the descent)

THE GRADIENT DESCENT ◧ 2D · ◍ 3D · ◆ 4D · ◐ shadow · 👶 TAP

Every neural network learns by ONE idea: measure the slope of your error, and step downhill. Gradient descent nudges each parameter opposite its gradient, shrinking the loss a little each step, until it settles at the bottom. Do it a billion times over a trillion weights and you have a trained model. Slide the step size and watch the ball roll to the minimum.

◆ LIT▲ AMBER
◧ THE MEASURE · 2D
◍ DOWNHILL · 3D · the engine that trains everything
◆ THE FOURTH · 4D · a tesseract turns
◐ THE SHADOW · one dimension down
👶 THE TODDLER CORNER — one fat tap
learning rate
position
loss
converged?

◆ LIT — exact / checkable

Gradient descent minimizes a loss L(θ) by repeatedly stepping against its gradient: θ ← θ − η·∇L(θ), where η is the learning rate. Each step reduces L (for small enough η) because the negative gradient is the direction of steepest local decrease. On a convex bowl it converges to the global minimum; it is the workhorse behind training every modern neural network (via backpropagation supplying ∇L). A fail-loud self-check throws unless descending on (x−3)² converges to x=3. ◆ real optimization, node-verified.

▲ AMBER — the figure

Shown on a 1-D convex bowl; real loss surfaces are high-dimensional and non-convex (many minima, saddles), and gradients are estimated on mini-batches (stochastic GD). The step rule and convergence-on-a-bowl are exact.

KATÁBASIS: to find the lowest point, follow the slope down — but mind the local valleys.  — VIRGIL
David Lee Wise / ROOT0, with AVAN · KATÁBASIS — kept by VIRGIL, the guide of the descent