Asynchronous Multi-Task Learning Framework for Feature Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing synchronous multi-task learning methods perform poorly due to data loss caused by asynchronous execution of tasks and input data over time, as they assume synchronized tasks and input data.
Innovation Solution
An asynchronous multi-task learning framework that integrates common feature vectors across time steps using a pre-feature integration unit and input feature integration unit, allowing for the extraction of output feature vectors and execution of tasks even with asynchronous data, minimizing information loss through serial data and feature integration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If synchronous multi-task learning methods are used, then task execution is simplified and coordinated, but data loss occurs due to asynchronous task execution and input data in real-world scenarios
Solution Approach 1:
The patent transitions from static synchronous task execution to dynamic asynchronous task execution. Each task can be executed independently at different time steps based on data availability, with the system adapting to varying input data arrival times. This is achieved through separate task execution modules that process tasks asynchronously while maintaining coordinated updates to shared parameters through gradient accumulation and periodic synchronization.
Solution Approach 2:
The patent implements preliminary action by accumulating gradients from asynchronous task executions before performing parameter updates. Instead of requiring all tasks to complete before any update, the system accumulates gradients from available tasks and performs preliminary parameter updates, reducing data loss and improving task execution reliability in asynchronous environments.
2Productivity
If synchronous multi-task learning methods are used, then coordinated task execution is achieved, but performance deteriorates due to data loss from asynchronous real-world data
Solution Approach 1:
The patent segments the multi-task learning system into independent task execution modules, each capable of processing its assigned task asynchronously. This segmentation allows tasks to be executed independently based on data availability rather than waiting for all tasks to be synchronized, improving both productivity and reliability by handling asynchronous real-world data effectively.
Solution Approach 2:
The patent implements feedback mechanisms where gradient information from asynchronous task executions is accumulated and used to update shared parameters. The system continuously receives feedback from completed tasks, accumulates gradient information, and performs parameter updates, ensuring that prediction accuracy improves over time even as tasks execute at different rates.
3Reliability
If asynchronous task execution is implemented, then data loss is minimized and task execution reliability improves, but system complexity increases
Solution Approach 1:
The patent implements universality by creating a multi-functional framework that handles both synchronous and asynchronous task executions through the same core architecture. The task execution modules and parameter update mechanisms serve multiple functions: processing individual asynchronous tasks, accumulating gradients from multiple sources, and performing coordinated parameter updates, thereby managing complexity while maintaining reliability.
Solution Approach 2:
The patent introduces intermediary mechanisms such as gradient accumulators and parameter servers that mediate between asynchronous task executions and the central model updates. These intermediaries buffer and coordinate information flow, allowing tasks to execute independently while maintaining system coherence, thus reducing the direct complexity of managing asynchronous operations without sacrificing reliability.
Data Source
AI summary
The operation method of a multi-task learning model according to an embodiment of the present disclosure may include: obtaining a common feature vector of a previous step; receiving input data corresponding to a current task executed in a current step from among a plurality of tasks; extracting a common feature vector of the current step based on the common feature vector of the previous step and the input data corresponding to the current task; extracting an output feature vector corresponding to the current task based on the common feature vector of the current step; and outputting output data corresponding to the current task based on the output feature vector corresponding to the current task.