Skip connection in a residual block. The input passes through two layers to give the residual , while an identity shortcut carries around them so the block outputs . Learning a residual on top of the identity makes very deep networks easier to optimize. Inspired by the ResNet paper.
#import "@preview/cetz:0.5.3": canvas, draw
#import draw: content, hobby, line
#set page(width: auto, height: auto, margin: 8pt, fill: none)
#set text(size: 12pt)
#canvas({
let node-sep = 2.5 // Horizontal separation between nodes
let arrow-style = (mark: (end: "stealth", fill: black, scale: 0.5))
// Layers get a colored box, activations a transparent one with a caption below.
for (column, body, name, fill) in (
(0, [layer 1], "l1", rgb("#ffd699")),
(1, $a(arrow(x))$, "act1", none),
(2, [layer 2], "l2", rgb("#7dc3c3")),
(3.75, $a(arrow(x))$, "act2", none),
) {
let padding = if fill == none { (0, 3pt) } else { (3pt, 6pt) }
content(
(column * node-sep, 0),
body,
fill: fill,
name: name,
frame: "rect",
padding: padding,
stroke: none,
)
if fill == none { content((rel: (0, -0.3), to: name + ".south"), "activation") }
}
content((3 * node-sep, 0), text(size: 1.8em)[$plus.o$], name: "add")
content((rel: (0, -0.3), to: "add.south"), "add")
for (start, end) in (
("l1", "act1"),
("act1", "l2"),
("l2", "add.west"),
("add.east", "act2"),
) { line(start, end, ..arrow-style) }
line((rel: (-2, 0), to: "l1"), "l1", name: "input", ..arrow-style)
content((rel: (0, -0.2), to: "input.10%"), $arrow(x)$)
hobby(
(rel: (-1.5, 0), to: "l1"),
(rel: (0, 2.2), to: "act1"),
"add.north",
close: false,
tension: 0.8,
..arrow-style,
name: "skip",
)
content((rel: (0, 0.3), to: "skip.mid"), $arrow(x)$)
content(
(rel: (0, -0.3), to: "skip.mid"),
align(center)[skip connection\ (identity)],
anchor: "north",
)
// Residual branch F(x) spans both layers
content(
(rel: (0, -1.2), to: "act1"),
math.underbrace(box(width: 17em), text(size: 1.4em)[$cal(F)(arrow(x))$]),
)
content(
(rel: (0.8, 0.8), to: "add"),
$cal(F)(arrow(x)) + arrow(x)$,
name: "fx-label",
frame: "rect",
stroke: none,
padding: 1pt,
)
line("fx-label.south", "add.north-east", stroke: .2pt)
})