ML AutoML

[junxnone]

Reference

• 2018 Taking Human out of Learning Applications: A Survey on Automated Machine Learning
• 2019 AutoML: A Survey of the State-of-the-Art -- update - website
• awesome-automl-papers
• AutoML综述更新 【AutoML：Survey of the State-of-the-Art】
• 从Google Visor到Microsoft NNI再到Advisor调参服务接口发展史
• 自动机器学习简述（AutoML）
• AutoML：机器学习的下一波浪潮
• AutoML challenge
• Using Meta-Learning to Initialize Bayesian Optimization of Hyperparameters
• 简述目前AutoML类型与模式

Brief

• AutoML 将特征、模型、优化、评价有关的步骤进行自动化地学习，使得机器学习模型无需人工干预。
• AutoML 目标
  • Better performance
  • No human assistance
  • Lower computation budgets
• HAO - Joint Hyper-parameter and Architecture Optimization - 参数和架构联合优化
• HPO
• NAS

Pipelines

AutoML Steps	Description
DP	Data Preparation = Data Cleaning + AutoAugment + ...
Feature Engineering	DL 中常用于时间或者结构数据
HPO	HyperParameter Optimization
NAS	Neural Architecture Search
Model Compression

Tools

Projects	Author	DP	FE	HPO	NAS
vega	HuaWei
NNI	Microsoft
AutoGluon	Amazon AWS
autokeras	keras-team
auto-sklearn	AutoML -xxx
Advisor
Google Cloud AutoML	Google
katib	Kubeflow