sciml
education

vZero explainer

Charts, diagrams and a slide deck explaining Neural ODEs for stiff chemistry, with toy experiments

34 images · 1 min read

Explore in Studio Jump to gallery

vZero explainer

datasheet // vzero-explainer

images: 34
videos: 0
formats: png, pdf
text: 89 words · 1 min read
studio: /solver?tab=ode&preset=4
topics: charts ×15

diagrams ×19

Datasheet

Deck: 35 slides (vZero Neural ODE), source slides.json; the shipped PDF has 34 slides (slide 31 removed)
Guide: 39 pages in the source; the shipped PDF has 37 pages
Sections (slides.json): The Map, The Core, Numerics, Training, Gradients, Representation, Koopman and Reduction, Physics and Measurements, Validation, Deployment, The Next Experiment, Recap
Figures: 15 charts (vector PDF rendered at 2400 px) and 19 slide diagrams
Also shipped: an interactive power-mean explorer (self-contained HTML)
Content type: teaching toy experiments; no H2/air accuracy, chemical speed-up or deployed INT8 results are claimed

Gallery

Slide 7 chart: A complete small learning loop — Slide 7 chart: A complete small learning loop Explain Playground

Slide 8 chart: Test a rollout beyond the training window — Slide 8 chart: Test a rollout beyond the training window Explain Playground

Slide 9 chart: Stiffness means competing clocks — Slide 9 chart: Stiffness means competing clocks Explain Playground

Slide 10 chart: The solver is part of the model you run — Slide 10 chart: The solver is part of the model you run Explain Playground

Slide 11 chart: Normalize the microphones — Slide 11 chart: Normalize the microphones Explain Playground

Slide 12 chart: Loud loss is not loud gradient — Slide 12 chart: Loud loss is not loud gradient Explain Playground

Slide 13 chart: The power mean is an emphasis knob — Slide 13 chart: The power mean is an emphasis knob Explain Playground

Slide 14 chart: The landscape is over network weights — Slide 14 chart: The landscape is over network weights Explain Playground

Slide 17 chart: Change the coordinates, keep track of the meaning — Slide 17 chart: Change the coordinates, keep track of the meaning Explain Playground

Slide 19 chart: Log transforms need a boundary policy — Slide 19 chart: Log transforms need a boundary policy Explain Playground

Slide 21 chart: One extra observable closes this toy — Slide 21 chart: One extra observable closes this toy Explain Playground

Slide 22 chart: DMD fits a linear step from snapshots — Slide 22 chart: DMD fits a linear step from snapshots Explain Playground

Slide 23 chart: POD compresses snapshots — Slide 23 chart: POD compresses snapshots Explain Playground

Slide 25 chart: A robot arm makes state and rate tangible — Slide 25 chart: A robot arm makes state and rate tangible Explain Playground

Slide 32 chart: Quantization is controlled rounding — Slide 32 chart: Quantization is controlled rounding Explain Playground

Slide 1 diagram: vZero Neural ODE — Slide 1 diagram: vZero Neural ODE Explain Playground

Slide 2 diagram: Five jobs, five different questions — Slide 2 diagram: Five jobs, five different questions Explain Playground

Slide 3 diagram: A state is a snapshot — Slide 3 diagram: A state is a snapshot Explain Playground

Slide 4 diagram: A rate is not the next state — Slide 4 diagram: A rate is not the next state Explain Playground

Slide 5 diagram: What does the neural network learn? — Slide 5 diagram: What does the neural network learn? Explain Playground

Slide 6 diagram: Where this appears in vZero — Slide 6 diagram: Where this appears in vZero Explain Playground

Slide 15 diagram: Forward prediction and backward credit — Slide 15 diagram: Forward prediction and backward credit Explain Playground

Slide 16 diagram: Check the gradient on a problem you can solve — Slide 16 diagram: Check the gradient on a problem you can solve Explain Playground

Slide 18 diagram: A new map does not erase every fast clock — Slide 18 diagram: A new map does not erase every fast clock Explain Playground

Slide 20 diagram: Koopman changes what we observe — Slide 20 diagram: Koopman changes what we observe Explain Playground

Slide 24 diagram: A latent Neural ODE adds a learned evolution — Slide 24 diagram: A latent Neural ODE adds a learned evolution Explain Playground

Slide 26 diagram: Build a valid mechanical mass matrix — Slide 26 diagram: Build a valid mechanical mass matrix Explain Playground

Slide 27 diagram: Use constraints that match the domain — Slide 27 diagram: Use constraints that match the domain Explain Playground

Slide 28 diagram: A graph tells the model who interacts — Slide 28 diagram: A graph tells the model who interacts Explain Playground

Slide 29 diagram: A camera sees an observation of the state — Slide 29 diagram: A camera sees an observation of the state Explain Playground

Slide 30 diagram: A progress coordinate is not a clock — Slide 30 diagram: A progress coordinate is not a clock Explain Playground

Slide 33 diagram: ONNX carries a computation graph — Slide 33 diagram: ONNX carries a computation graph Explain Playground

Slide 34 diagram: Change one layer, keep the comparison honest — Slide 34 diagram: Change one layer, keep the comparison honest Explain Playground

Slide 35 diagram: Learn the rate. Respect the solver. Check the physics. — Slide 35 diagram: Learn the rate. Respect the solver. Check the physics. Explain Playground

Documents