# Loss Landscape

> The A.I Loss Landscape project explores the morphology and dynamics of the fingerprints left by
> deep learning optimization training processes, in the intersection between research and art. It
> produces high quality visualizations (video and image) of how loss landscapes change as neural
> networks train. Led by Javier Ideami (researcher, engineer, creative director).

## Pages
- [Home](https://losslandscape.com/): Overview, latest news, featured articles/apps/talks, selected visualizations and the project mission.
- [Moving Lands — Videos](https://losslandscape.com/videos): Video visualizations (DROP, CROWN, ICARUS, LOTTERY, LATENT, SWAG, SENTINEL, EDGE HORIZON and more), all produced with real deep learning training data.
- [Still Lands — Gallery](https://losslandscape.com/gallery): High resolution image visualizations, opened full-screen in a lightbox.
- [FAQ — The Method](https://losslandscape.com/faq): How and why loss landscapes are visualized — dimensionality reduction, random directions, the Hessian and its eigenvalues, dynamics vs. shape, and more.
- [Knowledge](https://losslandscape.com/knowledge): Curated list of academic papers on loss landscapes, loss surfaces, mode connectivity and generalization.
- [A.I Fine Art](https://losslandscape.com/art): Fine art prints and NFT collections of the visualizations.
- [About — The Mission](https://losslandscape.com/about): The project's mission and gratitude.
- [LL Explorer](https://losslandscape.com/explorer): A free interactive tool to explore loss landscapes (separate application).

## Key facts
- Loss landscapes are representations of a network's loss values across its weight space; the loss function is multivariable and multidimensional.
- Visualizations reduce dimensions using pairs of random (near-orthogonal) directions plus the loss value as a third axis, typically in the -1..1 range, often log-scaled.
- The project studies landscapes in motion, riding along the minimizer to examine its changing nearby surroundings during training.
- Key pieces: ICARUS (mode connectivity, NeurIPS 2018, arXiv:1802.10026), DROP (dropout), CROWN (ReLU/Mish/Swish), LOTTERY (pruning, arXiv:1803.03635), SWAG (arXiv:1902.02476).
- Collaborators include Pavel Izmailov, Timur Garipov, the Landskape research group, and others.

## Links
- Contact: ideami@ideami.com
- Ideami's Central site: https://ideami.com
- YouTube: https://www.youtube.com/@ideami
- Medium: https://medium.com/@ideami