Distributed Learning Update Filtering for Lower Bandwidth Use
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing distributed learning systems require constant communication between edge devices and a parameter server, even when newly trained parameters do not contribute to improving the global model, leading to increased communication costs and reliance on unreliable connections.
Innovation Solution
Implementing a decision process using Partially Observable Markov Decision Process (POMDP) and reinforcement learning on edge devices to determine if locally trained parameters should be transmitted to the parameter server, reducing unnecessary data transmission by only sending updates that improve the global model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If constant communication is maintained between edge devices and parameter server, then model training convergence is ensured, but communication costs and bandwidth usage increase
Solution Approach 1:
The system performs partial communication by selectively transmitting only those local parameter updates that are likely to improve the global model, rather than constantly transmitting all updates. The worker device evaluates the potential benefit of each update and only communicates when worthwhile, reducing bandwidth usage while maintaining training effectiveness
Solution Approach 2:
The system implements feedback mechanisms where the worker device receives information about the global model state and uses this to determine whether its local updates should be transmitted. The parameter server also provides feedback about which updates were beneficial, enabling the worker to make informed decisions about future communication
2Productivity
If all local parameter updates are transmitted to parameter server, then global model improvement is maximized, but communication frequency and energy consumption increase
Solution Approach 1:
The worker device performs selective transmission by evaluating each local update's potential benefit before transmission. This partial action approach transmits only the most valuable updates, reducing communication energy consumption while maintaining effective global model improvement
Solution Approach 2:
The worker device performs preliminary evaluation of local updates before transmission, assessing whether the updates are likely to improve the global model. This preliminary filtering action prevents unnecessary communications and reduces energy consumption while preserving beneficial updates
3Measurement precision
If frequent parameter transmissions occur, then global model accuracy improves faster, but system reliance on reliable connections increases
Solution Approach 1:
The system uses partial communication where the worker device transmits only selected updates rather than frequent constant transmissions. This reduces the system's reliance on continuously reliable connections while still achieving effective global model accuracy improvement through strategic update selection
4Quantity of substance
If decision process is implemented to filter parameter transmissions, then unnecessary data transmission is reduced, but device complexity increases
Solution Approach 1:
The worker device performs self-service by autonomously evaluating its own local updates and deciding whether transmission is beneficial. This self-service decision process reduces unnecessary data transmission while the complexity is managed through efficient local evaluation algorithms that weigh update quality against communication costs
Data Source
AI summary
A computer implemented method of distributed learning in a system comprising a parameter server configured to maintain a global parameter set of a model to be trained and a plurality of workers. The method comprises transmitting a current global parameter set to a worker. The worker performs a training step based on training data available to the worker, thereby generating a local set of parameters of the model, determines a likelihood of the local set of parameters being suitable for improving the global parameter set and omits transmission of the local set of parameters to the parameter server if it is determined that the local set of parameters is not likely suitable for improving the global parameter set.