MLSys: The New Frontier of Machine Learning Systems

[dexhunter]

https://arxiv.org/pdf/1904.03257.pdf

[dexhunter]

@gaocegege 感觉是个私人库，我都添加不了Label 😿

[dexhunter]

This paper is an overview of the conference on Machine Learning and Systems (MLSys). The conference covers a range of topics including hardware systems for ML, software systems for ML and ML optimized for metrics beyond predictive accuracy. In my personal opinion, this conference attracts both academia and industry since companies such as Google, Microsoft are deploying MLaaS at a rapid speed. Contributors to projects such as kubeflow can learn more papers from this conference.

The paper mentioned some issues to be solved:

• software
  • write ML programs easier
  • measure performance of ML model
  • depolyment/monitoring/interprestation/debugging/adaptation/tuning/maintainance of ML applications (also data/labels)
• hardware
  • tradeoff between precision, stability, fidelity and more
  • distributed system to support ML training/serving
  • specialized, heterogeneous hardware for ML
• metrics beyond predictive accuracy
  • constraints such as power, latency and memory limits
  • full-stack privacy & security
  • accessibility of ML (more users)

A list of past papers could be found here

[gaocegege]

需要权限。。