Neural Network Task Combination for Inference Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multitask learning techniques face challenges in determining the optimal combination of tasks for deep neural networks to improve inference accuracy, as they often require manual configuration and lack efficient methods for identifying and correcting erroneous detections across multiple recognition tasks.
Innovation Solution
An information processing apparatus and method that determines a first task with the lowest inference accuracy and combines it with a second task based on erroneous detection patterns, using a combination determination unit to rewrite the network structure and improve inference accuracy by identifying common regions between different neural networks' outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple tasks are allocated to separate neural networks, then each task can be optimized independently, but the overall inference accuracy decreases due to erroneous detections
Solution Approach 1:
The patent combines multiple tasks into a single unified neural network that performs all tasks simultaneously. This merging approach allows the network to learn shared features across tasks and reduce erroneous detections by contextualizing results from different tasks, thereby improving overall inference accuracy while maintaining manageable complexity through shared computational structures.
Solution Approach 2:
The patent creates a universal neural network that can perform multiple different tasks (object detection, semantic segmentation, keypoint detection, etc.) within a single model. This multi-functional network shares common layers and parameters across tasks, improving inference accuracy by leveraging shared representations while avoiding the complexity of maintaining separate specialized networks for each task.
2Productivity
If manual configuration is used for task combinations, then task allocation can be controlled, but the process becomes time-consuming and inefficient
Solution Approach 1:
The patent implements an automatic task combination determination mechanism that evaluates multiple tasks and autonomously decides which tasks should be combined in the unified neural network. This self-service approach analyzes task characteristics, compatibility, and potential for shared features, automatically generating optimal task combinations without requiring manual configuration, thereby significantly improving productivity while maintaining ease of operation.
Solution Approach 2:
The patent incorporates feedback mechanisms where the system evaluates inference results from different tasks and uses this information to automatically adjust and optimize task combinations. The feedback from detection results, erroneous detections, and performance metrics guides the automatic determination process, enabling continuous improvement of task allocation while maintaining operational simplicity.
Data Source
AI summary
There is provided with an information processing apparatus. A first determining unit determines a first task from among a plurality of tasks based on a result of inference in the plurality of tasks, in which each of a plurality of trained models executing different tasks performs inference for detecting a different detection target on evaluation data. A second determining unit determines a second task to be combined with the first task based on a result of inference in which an object that is not a detection target of the first task was erroneously detected for evaluation data corresponding to the first task by a trained model executing the first task among the plurality of trained models.