In this blog post, Phitchaya Mangpo Phothilimthana and Bryan Perozzi from Google DeepMind and Google Research discuss the advancements in machine learning (ML) for improving the efficiency of ML workloads. They introduce the concept of ML compilers, which convert user-written ML programs into executable instructions for hardware. The authors highlight the importance of optimizing ML compilers to improve the efficiency of ML models. They also present the TpuGraphs dataset, which aims to improve the ML compiler by providing a learned cost model that predicts the runtime of programs. The authors discuss the scale and features of the dataset, as well as the baseline models and Graph Segment Training method. They also mention the recently concluded Kaggle competition based on the TpuGraphs dataset, highlighting interesting techniques employed by the participating teams. The authors congratulate the winners and express their gratitude for their contributions to ML research. Finally, they mention the NeurIPS Expo panel that covered advanced learned cost models and other related topics.
https://blog.research.google/2023/12/advancements-in-machine-learning-for.html