Task-Aware Multi-Task Model Routing for Redundant Output Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models waste computer resources by generating redundant prediction results when not all tasks need to be executed on the same data, leading to inefficiency.
Innovation Solution
A method where first information is added to the input of a machine learning model to determine specific tasks to be executed, allowing adaptive task execution and reducing redundant predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a machine learning model executes N tasks simultaneously on input data, then the model can provide comprehensive multi-task processing capability, but redundant prediction results are generated when not all tasks are needed, leading to waste of computer resources
Solution Approach 1:
The patent applies dynamics by making the model structure adaptive and changeable based on input requirements. A task determination module dynamically identifies which tasks need to be executed, and the model selectively activates only the necessary task processing paths, transforming the static multi-task model into a dynamic system that adjusts its computational scope according to actual needs.
Solution Approach 2:
The patent implements local quality by allowing different parts of the model to have different activation states. Instead of uniformly processing all N tasks, the system selectively activates only the local task processing components that are needed for the current input, enabling precise control over which parts of the model consume resources.
2Adaptability or versatility
If all N tasks are executed on input data through a single model, then comprehensive task coverage is achieved, but processing time increases when only M tasks (M<N) are actually required
Solution Approach 1:
The system dynamically determines the scope of task execution based on input characteristics. The task determination module rapidly assesses which M tasks are needed out of the N available tasks, and the model adaptively adjusts its processing scope to execute only those M tasks, thereby reducing processing time while maintaining necessary task coverage.
Solution Approach 2:
The patent segments the multi-task processing into independent task modules that can be selectively activated. By dividing the N tasks into separate processing paths, the system can choose to execute only the required M tasks in parallel, avoiding the sequential or simultaneous execution of all N tasks and thus reducing overall processing time.
3Adaptability or versatility
If a machine learning model is designed to handle multiple tasks, then the model structure becomes more complex, but this complexity leads to generation of unnecessary prediction results
Solution Approach 1:
The patent applies universality by designing a single machine learning model that can handle multiple tasks through a unified architecture. The model shares common feature extraction and processing components across all N tasks, reducing the need for separate dedicated models for each task. This universal structure is controlled by a task determination mechanism that activates only the necessary task outputs, managing complexity while maintaining multi-task capability.
Solution Approach 2:
The task determination module serves as an intermediary between the input data and the multi-task processing model. It analyzes the input and determines which tasks are needed, then directs the model to produce only the required prediction results. This intermediary layer manages the complexity by filtering task requirements before they reach the complex multi-task model, preventing unnecessary processing.
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
A data processing method, a model training method, and a related device are provided. The method may be applied to a multi-task processing scenario in the field of artificial intelligence. The method includes: obtaining first data and first information, where the first information indicates at least one task executed on the first data; and inputting the first data and the first information into a first machine learning model, and processing the first data by using the first machine learning model, to obtain at least one prediction result that is output by the first machine learning model and that is in a one-to-one correspondence with the at least one task. To be precise, the first machine learning model can learn of, based on the first information, specific tasks that need to be executed on the first data, so that a required task can be adaptively executed on the first data, thereby avoiding generating a redundant prediction result and avoiding a waste of computer resources.