100 Years of Embeddings w: Akshay Agrawal скачать в хорошем качестве

100 Years of Embeddings w: Akshay Agrawal

This talk is about Minimum Distortion Embedding (MDE), an embedding framework that generalizes over 100 years of embedding methods, including PCA, Laplacian Eigenmap, and UMAP. I'll also introduce a novel Quasi-Newton algorithm for non-convex constrained optimization problems that can be used to compute these embeddings efficiently and at scale. But this talk is also about the tools we use to do research: not only how new research motivates new tools (like PyMDE, a GPU-accelerated library for dimensionality reduction that scales to millions of items), but also how the tools we use shape the way we think, and ultimately the quality of the research we produce. I'll talk about why my experience as an engineer on Google Brain, and as a researcher in Stephen Boyd's lab at Stanford, led me to create marimo, a new kind of interactive programming environment for Python that solves widely-known problems with Jupyter notebooks, such as reproducibility and reusability, while also giving bringing data to life in new ways that can accelerate research. ​​​​​​​​​​Check out the full paper here: https://www.alphaxiv.org/abs/2103.02559 Example notebooks and documentation: https://pymde.org/examples/ marimo: https://github.com/marimo-team/marimo molab: https://molab.marimo.io/notebooks Careers: https://marimo.io/careers Discord: http://marimo.io/discord Stay up to date with future events on alphaXiv.org!