The Ins and Outs of Working with Embeddings and Embedding Models

Ready to zoom all the way in on a timely technical topic? We hope so, because this week’s Variable is all about the fascinating world of embeddings.

Embeddings and embedding models are essential building blocks in the powerful AI tools we’ve seen emerge in recent years, which makes it all the more important for data science and machine learning practitioners to gain fluency in this area. Even if you’ve explored embeddings in the past, it’s never a bad idea to expand your knowledge and learn about emerging approaches and use cases.

Our highlights this week range from the relatively high-level to the very granular, and from theoretical to extremely hands-on. Regardless of how much experience you have with embeddings, we’re certain you’ll find something here to pique your curiosity.

• How to Find the Best Multilingual Embedding Model for Your RAG
  As Iulia Brezeanu emphatically states, “Besides having quality data, choosing a good embedding model is the most important and underrated step for optimizing your RAG application.” Follow along her accessible guide to learn how to make the best choice for your project.
• OpenAI vs Open-Source Multilingual Embedding Models
  For another perspective on current options in the field of multilingual embedding models, we strongly recommend Yann-Aël Le Borgne’s post, which provides a detailed comparison of the performance of OpenAI’s latest generation of embedding models with that of their open-source counterparts.
• How to Create Powerful Embeddings from Your Data to Feed into Your AI
  Taking a step back from the question of model selection, Eivind Kjosbakken’s deep dive outlines the different approaches available for converting your data “from formats like images, texts, and audio, into powerful embeddings that can be used for your machine learning tasks.”

• Statistical Method scDEED Detects Dubious t-SNE and UMAP Embeddings and Optimizes Hyperparameters
  Walking us through their latest paper, Jingyi Jessica Li and Christy Lee provide a framework for identifying data distortions in projection from a high-dimensional to two-dimensional space and for optimizing hyperparameter settings in a 2D dimension-reduction method.
• Editing Text in Images with AI
  Scene text editing—the process of tweaking textual elements within images—is a surprisingly complicated task. Julia Turc shares some of the recent progress researchers have made in this area and expands on the role of embeddings within STE model architectures.
• A Real World, Novel Approach to Enhance Diversity in Recommender Systems
  For another concrete demonstration of the power of embeddings, we recommend Christabelle Pabalan’s new article. It unpacks the difficulty of boosting diversity in recommender systems’ outputs, and shows how choosing the right embedding model proved to be a key step towards achieving very promising results.

For readers who’d like to explore other topics this week, we’re thrilled to recommend some of our recent standouts:

• For their TDS debut, Skylar Jean Callis shared a comprehensive technical walkthrough of vision transformers (ViT), complete with a PyTorch implementation.
• Stay up-to-date with the latest ML research by following along Maarten Grootendorst’s deep dive on Mamba, a new state space model architecture that aims to become an alternative to transformers.
• In a thought-provoking post, Louis Chan reflects on some of the common sources of tension between data scientists and engineers (and what teams can do to mitigate them).
• As model sizes ballooned in recent years, the importance of model compression grew in lockstep. Nate Cibik just launched an excellent series on streamlining approaches, dedicating part one to pruning.
• Curious to learn about the power of simulations? Hennie de Harder’s new explainer focuses on Monte Carlo methods and how they can help solve complex problems.
• History buffs, rejoice: Sachin Date just published the latest installment in his series on the origins of key mathematical concepts, this time focusing on Pierre-Simon Laplace and the central limit theorem.

Thank you for supporting the work of our authors! If you’re feeling inspired to join their ranks, why not write your first post? We’d love to read it.

Until the next Variable,

TDS Team

The Ins and Outs of Working with Embeddings and Embedding Models was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.