Module 4 — Hiker’s Cheat Sheet (Rosetta Stone)

Module 4 — Hiker’s Cheat Sheet (Rosetta Stone)#

Use this table to translate between the Hiker in the Fog story and ML/DL technical terms.

Technical Term	Hiker Analogy Component	Practical Meaning (So what?)
Model	The Hiker	A function that maps inputs to outputs.
Weights (w)	The hiker’s internal dials	Adjustable numbers that control how strongly inputs influence outputs. Learning changes these.
Bias (b)	An offset	Extra adjustable number that shifts predictions, increasing flexibility.
Loss Function	Height of the mountain	Numeric “how wrong?” score that training tries to reduce.
Gradient	Slope under the boot	Direction and steepness telling how to change weights to reduce loss.
Learning Rate (α)	Step size	How big each update is; too big overshoots, too small is slow.
Optimizer	The walking strategy	The algorithm that applies updates (e.g., gradient descent, Adam).
Epoch	One full pass of the trail	One full pass through training data.
Convergence	Reaching a flat valley floor	Loss stops improving significantly.
Overfitting	Memorizing one path perfectly	Great training results, poor new-data performance.
Validation Set	A scout checking another route	Detects overfitting during training.
Local Minimum	A pothole	“Good-enough” valley that isn’t the best overall.
Activation Function	“Should I signal?” decision	Non-linearity enabling complex patterns (e.g., ReLU, Sigmoid).
Backpropagation	Radio messages uphill	Sends gradient information backward so every hiker can adjust dials.

Training vs Inference: training moves and adjusts dials; inference stands still and reports location.
Nothing is learned except weights (and bias).
Loss is the only feedback signal; gradients provide direction.