The document discusses machine learning (ML) model stores and some unique software engineering challenges for building ML applications. It identifies several key information elements that ML model stores typically provide, including the ML algorithm, framework, training set, license, size, usage statistics, and user feedback. The document also examines differences between how ML engineers and end users package and integrate ML models, and considerations for representing model quality, pricing schemes, and capacity planning for ML applications hosted in these stores.
1 of 42
Downloaded 13 times
More Related Content
An Exploratory Study on Machine Learning Model Stores
12. ML Model Stores
January 2018
June 2018
November 2018
November 2018
April 2019 (beta)
What are the unique SE
practices/challenges for
building ML applications?
29. To package an ML application for release
- Source code
-Trained ML model(s)
-Training set info
Integrating such a model …
- API calls, re-train on new data?
30. To package an ML application for release
- Source code
-Trained ML model(s)
-Training set info
Integrating such a model …
- API calls, re-train on new data?
36. How unique are the ML models
provided by each model store?
Very few cross-store ML applications
- Vendor lock-in
- Need for cross-platform frameworks
41. Minor Releases Different Product Pages
A new version for an ML application is:
- same ML model re-trained on another dataset? or
- same dataset with different ML implementations?