Insights Provably expressive graph neural networks

Twitter describes how to design local and computationally efficient provably powerful graph neural networks.

Insights Sharing learnings about our image cropping algorithm

Twitter shares a technical analysis of its assessment for potential bias in its image cropping algorithm as part of its efforts to be more transparent around how it uses machine learning to improve pe

Open source Dropping cache didn’t drop cache

Twitter engineers found and fixed a Linux kernel bug in memory shrinker which caused OOM for us.

Insights ICLR Invited Talk on Geometric Deep Learning

This blog post is based on Michael Bronstein’s keynote talk at ICLR 2021 and the paper M. M. Bronstein et al., Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges (2021)

Insights Fusing Elasticsearch with neural networks to identify data

Twitter annotated its data to a defined taxonomy by leveraging machine learning in its data platform by fusing Elasticsearch with neural networks.

Infrastructure Adopting RocksDB within Manhattan

Learn about Twitter’s adoption of RocksDB within Manhattan, its internal distributed key-value store.

Insights Simple scalable graph neural networks

In this post, we describe a graph neural network architecture (SIGN) that is of simple implementation and that works on very large graphs.

Insights Harnessing second order optimizers from deep learning frameworks

Introducing the dict-minimize package

Infrastructure Sharding, simplification, and Twitter’s ads serving platform

Learn how Shardlib, a new sharding library at Twitter simplifies the management of sharded ads services and enables dynamic resharding of these services without redeploying their clients.

Insights Building a data stream to assist with COVID-19 research

Twitter built a data product to help with coronavirus research by opening our APIs to academic researchers, to access the public conversation around COVID-19.

See more