Fix: one symbol vocabulary across the optimizer cards - #12
Merged
Conversation
The card equations drifted letter by letter as they were written; a
reader comparing adjacent cards could reasonably conclude a symbol had
changed meaning. Unified under one convention (recorded in
DEVIATIONS.md):
- The squared-gradient memory is one bold s everywhere: AdaGrad and
RMSProp promote plain s to \mathbf{s} (matching AdaDelta), Adam's
plain ŝ becomes the bold ŝ its own repair cards were already using
(formula + all five prose mentions).
- Prodigy no longer resurrects v — the letter that meant velocity two
acts earlier — for Adam's denominator: it is the family's s. Its
paper-internal ℓ1 tally (confusingly also called s in the paper)
becomes w, and the prose now names r and w as the two running
tallies of the ⟨g, x₀−x⟩ alignment. The d-update drops subscripted
recursion for the ← style every other card speaks.
- Nesterov's lone ∇ℒ(θ) argument decoration is gone — gradients are
at the current point on every card, so it only invited a hunt for a
difference that isn't there.
- One minus: the tight (1{-}β) in Nadam and Lion becomes the spaced
(1-β) of every other card.
- Sophia's clip ρ gets the house defusal against RMSProp's decay ρ
('the alphabet is small and the field is greedy' — third verse).
The reveal-the-code files keep each paper's own names; the ledger now
states that split explicitly. Verified in the live DOM: all fifteen
rendered formulas checked via their KaTeX annotations; 194 tests,
svelte-check clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Deploying gradient-lab with
|
| Latest commit: |
7c2fe77
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://5fa33072.gradient-lab.pages.dev |
| Branch Preview URL: | https://fix-card-notation.gradient-lab.pages.dev |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The inconsistencies
Reviewing all fifteen card equations side by side turned up five places where notation silently shifted meaning between cards:
s, AdaDelta wrote bold𝐬for the same quantity, Adam wrote plainŝ— while its own repair cards (Nadam, AdamW, RAdam) wrote boldŝright underneath it.v— the letter that meant velocity in the momentum act — for Adam's squared-gradient denominator, and usedsfor a different, paper-internal tally two symbols away from the family'ss.d_{t+1} = max(d_t, …)) where every other card uses the←update style.(θ)and left the other bare in the same line — inviting a hunt for a difference that doesn't exist.(1{-}β), everyone else the spaced(1-β).The convention (now recorded in DEVIATIONS.md)
←for updates,=for one-shot definitions; no time subscripts unless the quantity genuinely indexes time (β^t, ρₜ).Prodigy's prose also now names
randwas the two running tallies of the ⟨g, x₀−x⟩ alignment, so no symbol in its formula is unexplained.Verified
All fifteen rendered formulas checked in the live DOM via their KaTeX annotation sources; Prodigy card eyeballed in dark mode; 194 tests green,
svelte-checkclean. Content-only — no optimizer behavior changes.🤖 Generated with Claude Code