cs.CV
VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training
Zhan Tong and others
8 visuals, 21 sections
Read the explainerSee the idea
behind any paper
Paste an arXiv link. ArcVisual reads the paper’s source, finds the parts that are genuinely hard, and turns them into animations you can play, pause and scrub. Every claim points back to the paper’s own words.
Try one:
- paste a link from
- arXiv
- kinds of animation
- 3
- of claims cite the paper
- 100%
Input
Visuals
Grounded
Any arXiv field
with LaTeX source
csmathphysstatbio
Every token brings a query and a key.
0:00 / 0:30
Each query is scored against every key; softmax turns the scores into weights that mix the values.
Drawn live in your browser, from Attention Is All You Need · §3.2.1
Papers already explained
Open to everyone
Every explainer is public and keeps its link. If someone has explained a paper before, it opens instantly for you.
From link to explainer in a few minutes
The article becomes readable as soon as the paper is analysed. Visuals fill in beside the text as each one is checked.
- 1
Paste a link
Any arXiv paper with LaTeX source. No account, no upload.
- 2
We read the source
Sections, equations and figures come from the LaTeX itself, so nothing is lost to PDF extraction.
- 3
We pick what is hard
Ideas are ranked by difficulty and how central they are. Most sections stay as prose, on purpose.
- 4
You read and replay
The hard parts become animations beside the text. Scrub back to the exact step you missed.
Four ways an idea can move
These run live in your browser, from the same kind of parameters the pipeline writes for a paper. Press play, drag the bar, or slow them down.
Mechanisms
When an idea is a process, it is drawn as one: the parts appear, data flows between them, and the cells that matter light up.
Every token brings a query and a key.
0:00 / 0:30
Each query is scored against every key; softmax turns the scores into weights that mix the values.
Attention Is All You Need · §3.2.1
Derivations
When a paper walks through an equation, each step is written out and transformed into the next.
Scaled dot-product attention
step 1 of 3
Start with how well a query matches a key: their dot product.
0:00 / 0:14
Dividing by √dₖ keeps dot products from growing with dimension, so the softmax never saturates.
Attention Is All You Need · §3.2.1
Results
When a figure carries the argument, the chart is built up so you see what changed and by how much.
Two networks, 18 and 34 layers deep, on ImageNet.
0:00 / 0:15
Making a plain network deeper made it worse; making a residual network deeper made it better.
Deep Residual Learning · Table 2
Systems
When a model is a pipeline of parts, the blocks appear in order and data is traced through them.
Each token starts as an embedding.
0:00 / 0:21
An encoder layer is two sublayers, each wrapped in a residual connection and normalisation.
Attention Is All You Need · §3.1
Made for people who read papers
Researchers
Get the core of a paper outside your field before committing an afternoon to the PDF.
Students
Watch a derivation unfold one step at a time instead of decoding it from a dense page.
Professors
Share a readable companion that cites the exact passages it explains, so nothing is paraphrased away.
Educators
Pause on the step a textbook skips, scrub back while you talk it through, and enlarge it for the room.
Good to know
This is a tool for people who read critically, so here are its limits up front.
- arXiv papers with LaTeX source
- PDF-only submissions are turned away with a reason, never half-explained.
- English, under 60 pages
- Longer or non-English papers are declined in the first few seconds.
- A companion, not a replacement
- Every section links back to the original, and figures stay with their authors.
- Free while it lasts
- Explanations run on a free model budget. If today's is spent, try again in a few hours.
Start with a paper you already know
It is the fastest way to judge whether the explanation is any good, and whether anything was gotten wrong.