Zero-Shot Learning Device Using Correspondence Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Zero-shot learning methods require constructing new predictors for unseen tasks or classes, which is costly in terms of computational resources, time, and labor, especially when dealing with numerous prediction targets.
Innovation Solution
A learning device and method that utilize correspondence inference to calculate outputs of existing predictors for seen tasks or classes and infer correspondences with attribute information for unseen tasks or classes, allowing for predictions without constructing new predictors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a new predictor is constructed for all seen tasks or seen classes using input/output data and attribute information, then prediction capability for unseen tasks or classes is achieved, but computational resources, computation time, and labor costs increase significantly
Solution Approach 1:
The patent segments the prediction process into two independent phases: (1) training predictors only for seen tasks/classes using available input/output data, and (2) inferring predictions for unseen tasks/classes by matching attribute information with seen tasks/classes. This segmentation eliminates the need to construct new predictors for unseen tasks, reducing computational resources and time while maintaining prediction capability.
Solution Approach 2:
The patent performs preliminary action by pre-training predictors for all seen tasks and classes before encountering unseen tasks. The attribute information for seen tasks is pre-processed and stored. When unseen tasks appear, the system directly matches their attribute information with pre-processed seen tasks, avoiding the need for new predictor construction and enabling rapid adaptation.
2Adaptability or versatility
If predictors are constructed for numerous prediction targets, then comprehensive prediction coverage is achieved, but the cost increases according to the number of prediction targets
Solution Approach 1:
The patent makes the seen task predictors universal by designing them to handle multiple task types through attribute-based matching. A single predictor trained on seen tasks can be applied to unseen tasks by matching attribute information, eliminating the need for separate predictors for each task. This multi-functionality achieves comprehensive prediction coverage while maintaining cost efficiency.
Solution Approach 2:
The patent uses attribute information as a copy or representation of task characteristics. Instead of constructing new predictors for unseen tasks, the system copies the attribute information of unseen tasks and matches it with attribute information of seen tasks, thereby reusing existing predictors. This copying approach enables comprehensive coverage without proportionally increasing costs.
Data Source
AI summary
A learning device 100 includes correspondence inference unit which calculates outputs of predictors, which have learned for seen tasks or seen classes, for test input data, and infers correspondences between the calculated outputs and attribute information corresponding to an unseen task or an unseen class, and prediction unit which calculates a prediction output for the attribute information corresponding to the unseen task or the unseen class, using the inferred correspondences.