Foreword
Good news on the fight against SARS battlefield are pouring in, the topic of "Data immunity" has been heating up. As the body needs to enhance their immunity against potential virus threats interpersonal contact, corporate and individual user data on how to improve the "immunity" to enhance their data security and defense capabilities in an increasingly wide range of industries and cooperation? Recently, in order to Pozhu rise of the "federal study" into the industry perspective. Federal study, in conformity with data security regulations and policies, helping businesses achieve 13 multilateral collaboration training AI. Federal FATE (Federated AI Technology Enabler) as a federal study the world's first industrial-grade open-source framework to achieve a homomorphic encryption and multi-party computation (MPC) of the secure computing protocols to support federal study architecture with a variety of machine learning algorithms to achieve learning , a federal study areas not open around the "monumental."
Recently, FATE released the first version of the 2020 update --FATE v1.3. In this version, FATE first increase in the federal recommendation algorithm module FederatedRec, the module contains six recommended scene commonly used algorithms, including five kinds of vertical and 1 lateral Federal algorithm Federal algorithm can be used to solve the federal scene under study recommendation problem, as predicted score, sorting and other items. In addition, VMware China R & D Center for Open Innovation native cloud KubeFATE laboratory team also jointly issued this release ushered in the update, the whole reconstruction, and the introduction of FATE-Serving support for the latest version, so users can online reasoning. Finally, a number of modules FederatedML, etc., the new version has also been updated and optimized. The new version will provide developers with a great experience of FATE.
The project is now available on GitHub: https: //github.com/FederatedAI/FATE
FederatedREC: to enhance the algorithm to predict performance and optimize the efficiency of product distribution
In FATE1.3 version, FATE new federal recommendation algorithm module FederatedRec, the federal recommendation module further clarification. This algorithm commonly used algorithm module contains six major recommendation of the scene, are:
-
Vertical Federal algorithm Hetero FM (Factorization Machine)
-
Federal lateral algorithm Homo FM (Factorization Machine)
-
Vertical Federal algorithm Hetero MF (Matrix Factorization)
-
Hetero SVD algorithm longitudinal Federal
-
Vertical Federal algorithm Hetero SVD ++
-
Vertical Federal algorithm Hetero GMF (Generalized Matrix Factorization)
In this algorithm class 6, Hetero FM are Homo FM and FM (factorization machine) algorithm in the longitudinal and transverse federal federal two scenarios, the algorithm can achieve different data-side data federation model joint, different Party data directly explicit features and joint cross scoring. As for the recommendation algorithm scene, Hetero MF, Hetero SVD, Hetero SVD ++ and HeteroGMF also provides a collaborative filtering algorithm package in rich federal model. FATE v1.3 tool based on an algorithm provided, it is possible to realize user-item data between different parties, user-user and the item-item matrix decomposition. For developers, this module is recommended by federal, can significantly improve the efficiency of distribution and product predict the effect of their own algorithms.
KubeFATE: Support FATE-Serving online for federal reasoning, completely reconstructed Kubernetes installation more convenient
With this release, KubeFATE also synchronize updates to 1.3, introduced FATE-Serving support for the latest version, allows users online reasoning, machine learning common functions to further improve.
In addition, for KubeFATE itself, this version also conducted a thorough reconstruction, are:
-
FATE cluster supports full lifecycle management, including a list of clusters FATE inquiry now deployed, see the specific configuration of each FATE cluster of FATE cluster configuration update, delete FATE clusters and other functions;
-
FATE cluster management framework to achieve the task, developers can track the detailed sub-tasks for each task, to facilitate positioning of the infrastructure level;
-
The relationship between the configuration program and FATE cluster separation KubeFATE after subsequent update releases FATE, developers do not need to re-download and install every time KubeFATE to perform the upgrade. Under ideal network conditions, KubeFATE FATE cluster configuration can be automatically downloaded, developers can use directly. Such as the use environment without network support, can be downloaded from Release KubeFATE in tgz package, use the command line KubeFATE uploaded directly to the service, you can deploy to achieve released a new cluster support.
-
Strengthen FATE modular installation and deployment, this version, KubeFATE the Exchange may or several separate modules deployed as a cluster, by Kubernetes Unicom up. This feature will help to complex IT environment of enterprise deployment.
Finally, in the new version, as a service KubeFATE will provide external RESTfulAPI, for developers, the follow-up will be in the form of services KubeFATE resident of cluster management, command-line tool can also be a foreign network even FATE cluster administrator running on a laptop. External interface function with a RESTful API as shown below, can facilitate access to existing enterprise cloud managed system.
FederatedML: training to enhance the efficiency of sparse data, and then optimize memory consumption
In addition to the focus on updating federal recommendation and KubeFATE, the new version for FederatedML also made further enhanced in 1.3, vertical Federal generalized linear models (Hetero-LR, Hetero-LinR, Hetero-PoissonR) began supporting sparse data training, developer when training sparse data can be clearly felt efficiencies to reduce memory consumption. 32M solution in the bin limits, characterized in that the bin also support higher dimensions and more data samples. SecureBoost vertical gradient histogram 32M address the limitations, so FATE could support a higher dimensional feature of training secureboost.
Overall, the 1.3 version of FATE functional and practical value has been enhanced, adding the federal recommendation algorithm sub-module FederatedRec, for the users, the most obvious benefit is that you can use to predict the effect of the federal recommendation to improve their own algorithms and distribution efficiency of the product, the quality of the recommendation service to a higher level. This is rich enough for first-party data, or data services in the initial period of accumulation of less user side, is very helpful. The KubeFATE also opened FATE in a production environment, especially on a cloud native environment using the most optimal management feature updates explore the road. Follow-up, we will further cooperation and VMware, to launch a new project based on multi-party management of FATE.
We welcome contributions are interested in learning together with my colleagues on the Federal codes, or submit Issues Pull Requests. Details can be found FATE official website of the project contributors Guide: https: //fate.fedai.org/contribute/
For inquiries, please leave a message exchange. You can also add us FATE assistant: FATEZS001, further exchanges, thanks!
