wantedly / machine-learning-round-table

Gather around the table, and have a discussion to catch up the latest trend of machine learning 🤖
https://www.wantedly.com/projects/391912
297 stars 2 forks source link

[2023/08/30] Machine Learning 輪講 #210

Open nogawanogawa opened 10 months ago

nogawanogawa commented 10 months ago

Why

Machine Learning 輪講は最新の技術や論文を追うことで、エンジニアが「技術で解決できること」のレベルをあげていくことを目的にした会です。

prev.https://github.com/wantedly/machine-learning-round-table/issues/208

What

話したいことがある人はここにコメントしましょう! 面白いものを見つけた時点でとりあえず話すという宣言だけでもしましょう!

zerebom commented 10 months ago

MLflow2PROV: Enhancing Provenance in Machine Learning Workflows

https://github.com/mariusschlegel/mlflow2prov

概要

問題提起

解決策

image

Output例

図が細かいので一部だけ切り取っている。

image

感想

nogawanogawa commented 10 months ago

A Survey on Efficient Training of Transformers

論文URL

https://arxiv.org/abs/2302.01107

著者

Bohan Zhuang, Jing Liu, Zizheng Pan, Haoyu He, Yuetian Weng, Chunhua Shen

会議

IJCAI 2023 survey track

背景

最近のTransformerは大きな計算リソースが要求される。

計算量の観点では、モデルの大規模化によって消費電力等が増大している。 またモデルサイズの大規模化に伴い、要求するメモリサイズも指数関数的に増大する。

計算量とメモリリソースを効率的に使用することにより、Transformerの学習を低コスト・高精度で行う事ができるはずである。

目的

Transformerの学習における計算量・メモリリソースの効率化

アプローチ

調査概要

image

学習効率は大きく3つのカテゴリに分解でき、それぞれについて調査している。

memo

https://github.com/nogawanogawa/paper_memo/issues/96

Hayashi-Yudai commented 10 months ago

SHARK: A Lightweight Model Compression Approach for Large-scale Recommender Systems

推薦モデルに対してうまくfeature selectionと量子化を行うことで軽量化と処理の高速化を実現した