a very brief summary of client selections in federated learnings
[TOC]
IEEE TWC 2019
Motivation
existing works focus on the problem in individual learning ⇒ overlooking long term performance
Client Selection
Goal & Criterion:
-
energy efficient
- transmission power (Shannon's formula)
- local training energy consumption
-
later training result is better (exp. verification, ascend > uniform || descend) in terms of loss, accuracy, robust
- remark: mostly empirical, may need more experiments on temporal pattern
Solve:
incorporated client selection into overall performance optimization using Lyapunov method
use decision variable to decide whether a client is selected
IEEE IOTJ 2021
Motivation:
current complete/quasi random client selection cannot handle the heterogeneity of the client devices ⇒ some of the client may fail to complete task ⇒ many discarded learning rounds ⇒ affect the model accuracy
Client Selection (FedMCCS)
Goal:
- consider the heterogeneity of the clients
- their limited computation and communication resources
- achieve maximum number of those that can complete FL rounds with out system cash
- maintain high-performance Fl mode in the mean time
Criterion:
- resource utilization is under budget per device type
- CPU frequency
- memory cost
- energy cost
- time needed to complete a round is under threshold
- download time
- predicted update time (RUPred-LR)
- upload time
- maximize event rate (?)
IEEE IOTJ 2021
Motivation:
optimize accuracy in stateful FL with a budgeted number of candidate clients
improve test accuracy
Client Selection:
Goal: select the
Criterion: test accuracy
Formulation & Solve: similar to secretary problem [4] [5]
- evaluate the test accuracy of the first
$\alpha^*$ clients and reject them all - select the next
$R$ clients with test accuracy better than the best test accuracy of the first$\alpha^*$ clients - if none is found then select the last clients
How to select the value of
IEEE TPDS 2021
Motivation:
- the number of clients could be sufficiently large & bandwidth is limited ⇒ should select clients
- clients with low priority are simply being deprived of chances to participate at the same time ⇒ data imbalanced
Client Selection (C2MAB + RBCS-F)
Goal: select clients to minimize long-term model exchange time while guaranteeing fairness
Criterion:
-
long-term fairness
-
availability of clients
-
selection fraction: sometimes available clients is not enough
Solve:
Lyapunov method
and one more thing to answer: how to calculate the model exchange time?
here using Contextual Combinatorial Multi Arm Bandit
The methods and formulations in this paper is highly similar to TWC2020
IEEE TC 2021
Motivation
- address problems in federated learning in extreme conditions
- mitigate the impacts of stragglers, crashes and model staleness
IEEE ICASSP 2020
Federated Machine Learning: Survey, Multi-Level Classification, Desirable Criteria and Future Directions in Communication and Network Systems
IEEE Communications Survey & Tutorials 2021