M03 / FOUNDATIONS

Features. From edges
to eyes, in five steps.

Every CNN climbs the same ladder: pixels → edges → textures → parts → objects. The features get smaller, deeper, and more abstract. The first layer sees brightness. The last sees "cat".

EXPERIMENT 03.01 · FEATURE.HIERARCHY LEVELS · 5 ABSTRACTION · low → high

INPUT IMAGE

Drag the slider →
walk up the abstraction ladder.

LAYER L2

FEATURE HIERARCHY conceptual visualization

WHAT THIS LAYER "SEES" L2 · textures

ACTIVATION ON YOUR IMAGE

28×28×16 · 12.5k activations

// 02

The full ladder · what each level captures

LAYER	SHAPE	DETECTS	RECEPTIVE FIELD	EXAMPLE FEATURES
L0 · pixels	224×224×3	raw RGB intensities	1×1	— (input)
L1 · edges	112×112×64	oriented edges, brightness gradients	~7×7	horizontal, vertical, diagonal lines
L2 · textures	56×56×128	corners, T-junctions, repeated patterns	~20×20	stripes, dots, fur, scales
L3 · parts	28×28×256	eye-shapes, wheel-shapes, leaf-shapes	~60×60	eyes, noses, ears, wheels, beaks
L4 · objects	14×14×512	whole-object templates	~150×150	cat-face, dog-face, car, building

Key intuition → Receptive field grows with depth. Layer 1 only sees a 7×7 patch, so it can only learn 7×7 things — edges. Layer 4 sees nearly the whole image, so it can learn whole objects. Spatial resolution drops as channel depth rises, packing more "meaning per pixel".