LEARNING: how it learns

LEARNING is the model room: what the model weighs, which way liking rises on the map, how its guesses have scored, your taste as JSON, and the math behind all of it.

LEARNING only shows; nothing here changes the model, except a style’s name. It is the level above TASTE, out at the top, for when you want to see inside.

1. Make a few picks in EVOLVE. LEARNING fills in once the model first fits: at the sixth pick, or right after the warm start.
2. Go to LEARNING: its stop at the top of the levels, or ⌥5.
3. Choose a style, read what it weighs, and name it.

The line under the title says what the model was fitted from (From 18 picks.), or how far its first fit is (3 more picks and it fits your taste.).

On a phone LEARNING is one screen that stands still: three tabs under the title show one of its panels at a time, where liking rises (the map, first), what it weighs (a list that scrolls) and its forecasts, with COPY AS JSON and THE MATH under them.

What it weighs

The model hears every sound as 44 numbers: 18 about how it sounds and 26 about how its patch is built. For each style it holds one weight per number, and WHAT IT WEIGHS shows them, largest first.

The styles. A chip for each style that claims at least 2% of the pool, with its share. Click a chip to show its weights; a dot at the chip’s left marks the one shown. Click its name and type to name it, and every place the style appears uses your name; a refit keeps it on the style it named. ▶ plays the sound that style rates highest today, with a dotted amber ring while it is rendered first.

The rows. Each row is one quality: its name (body, grit, drive & fold), the technical name beside it, a bar, and the weight.

• A bar to the right of the center line: it likes more of this. To the left: less.
• The thin line across the bar is how far the weight could be off, one standard deviation each way.
• Solid: that line clears the center. It is sure which way the style leans, however short the bar.
• Hollow, with a ? in the margin: the line crosses the center, so the lean could go either way. Early on nearly every row is hollow.

The bars move with every pick. A pick reweights the model’s draws, and after each one LEARNING asks for the weights under the reweighted draws and moves every bar to them; a refit, a generation or a file you open moves them too. A pick's new weights that land while your pointer is on the bars move only their lengths; the rows take their new order, sliding there, when you move off them or rest a second. A refit that lands then waits whole, the same way.

Point at a weight (or Tab to it) and the small map beside it shades every sound by how much of that quality it has: brighter and larger for more, dimmer for less. The legend names it (dots: grit).

REPLAY (or R) steps the bars through the moments this session kept, one after another, and each bar moves to where that moment left it. The label beside it says which moment: after 21 picks for a pick, refit · after 24 picks for a refit, and likewise a generation, an opened taste file, a star or a cut. A step is credited to what it was. Only a step from one pick to the very next draws where each bar was as a dashed outline and lights the weight the pick moved most; a step across a refit, a generation or a file says that it moved, and a step across moments that weren’t kept (a quick run of picks keeps only the last one’s weights) says what moved since the moment before. It ends on now, and starts from the first moment the session kept, as TASTE’s track does.

On a narrow window the six that weigh most show, and ALL 44 WEIGHTS opens the rest. Before the first fit there are none: none yet: it weighs nothing until it first fits.

Where liking rises

The pool on the map, each sound glowing by how much the model guesses you’d like it, with the sound you’re playing ringed and named. While you play it, its face’s outline lights up on the ring, as loud as it sounds. The arrow is the way liking rises across the map: a summary fit, made by the model, of each sound’s liking on the map’s two axes, and the legend gives that fit’s r², how much of the glow one direction explains (the arrow: liking rises · explains 16%). A low share means your taste doesn’t lie along one line across this picture, which is common: the map is flat, and your styles can pull different ways.

The arrow follows every pick. When a pick moves the ratings, it turns to its new heading, and the old one stays a moment as a dashed line. Before the first fit there is nothing to point along: no direction yet.

Its forecasts

Before every pick between two sounds, the model writes down the chance it gives each side. Then your answer arrives, and the forecast is scored against it.

• The count: how many it guessed right, out of every forecast (14 / 22).
• Under it: what it expected against what it got (expected 64% · was 58%). Expected is the average confidence it had in its own guesses; close to what it got means its confidence is honest.
• The strip: each forecast as the chance it gave the sound you picked, from 0% to 100%. Bright where it guessed right, dim where it guessed the other. The newest drops in as it lands.
• The skill line: calibrating · 4/20 until it has 20 guesses, then how much sharper than a coin flip it has been (13% sharper than chance), on the fair-test picks once there are 20 of them.
• By kind, once there is more than one: dealt pairs, edits you heard, edits you asserted, offers you took or passed, each with its own skill.

Pairs in EVOLVE and PATCH’s strip, KEEP AS NEW’s comparison and PERFORM’s offers are all forecast. Nothing is forecast before the first fit, so the warm start’s picks have none, and after it the panel says none yet: it guesses before each pick from here until your next pick.

Copy as JSON

COPY AS JSON puts your taste’s data on the clipboard, as the engine posted it: what it learned from, the math’s numbers, every style’s weights, every sound’s rating, every forecast, and the calibration. Where the browser won’t let a page write to the clipboard, the JSON appears in a box below to select and copy. Nothing is downloaded; Download your taste, in ⌘K, saves the whole record as a file.

The math

THE MATH opens the model in one line and five sentences, with every number read from the engine:

P(A over B) = σ(u(A) − u(B)), u = max over styles of w · φ

φ is a sound’s 18 audio and 26 structural features, standardized, and w holds one weight for each. A pick moves w along φ(picked) − φ(passed), so every sound’s rating moves at once, not only the two you heard. It holds 500 draws of w. Each pick reweights them, and every 6 picks it fits them again. It is allowed one more style for every 20 things it learns from, up to 5. It rates a sound by the style that likes it most.

“Things it learns from” are what TAUGHT counts: picks, stars and cuts. The reference goes further.

In ⌘K

At LEARNING, ⌘K’s list starts with THIS LEVEL: Replay what it weighs (R), Copy as JSON, and The math (it says open while the math is open).

What to try next

• Name your largest style for what it leans on, then find the name in PATCH’s catalog (point at a module).
• Make six picks, and watch a hollow row turn solid at the refit.
• Copy the JSON and look at a forecast: the chance it gave, and what you picked.