Merged Model Consistent Hashing for Output Stability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Model updates often result in discrepancies and inefficiencies due to differences in output paths between old and new models, leading to inaccurate population assignments and resource reallocations, which can cause overcounting and inefficiencies in resource allocation and memory management.
Innovation Solution
A method that maps structures of a new model to those of an original model, generates a merged model, and applies consistent hashing to assign persistent identifiers, minimizing differences in output and reducing reallocations by reusing existing allocations and maintaining statistical properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a new model is generated from an original model with updated statistical properties, then the model accuracy and population assignment quality is improved, but the output consistency and resource allocation stability deteriorates due to different decision paths
Solution Approach 1:
The patent applies preliminary action by pre-computing the mapping between old and new model structures before actual model deployment. The system identifies corresponding nodes and structures between models in advance, establishes the mapping relationships, and prepares the transformation framework before the new model is fully deployed, thereby reducing the impact on output consistency
Solution Approach 2:
The patent introduces an intermediary mapping layer between the original model and the new model. This mapping structure acts as a mediator that translates outputs from the new model back to the original model's decision paths, ensuring that while the new model provides improved accuracy, the output remains consistent with the original model's stable decision framework
2Manufacturing precision
If model structures are updated to improve population assignment accuracy, then the classification quality is improved, but the resource reallocation frequency increases causing inefficiencies
Solution Approach 1:
The patent uses copying by creating a structural map of the new model that replicates the organization and naming conventions of the original model. This copied structure allows the new model to maintain improved classification quality while presenting its outputs in a format that requires minimal reallocation of existing resources, as the mapped structures directly correspond to original resource locations
3Loss of time
If consistent hashing is applied to maintain persistent identifiers across model updates, then the resource reallocation overhead is reduced, but the system complexity increases due to mapping and classification requirements
Solution Approach 1:
The patent applies universality by designing a mapping system that serves multiple functions simultaneously: it maps model structures, maintains persistent identifiers, enables consistent hashing, and preserves resource allocation relationships. This multi-functional approach reduces overall system complexity compared to implementing separate systems for each function
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods, systems, and computer readable medium for reducing inconsistencies in output between an original model and a new model. The method includes receiving an original model and a new model, mapping structures of the new model to structures of the original model, classifying each structure of the new model as belonging to a group of the original model, an unused group not in the original model, a subset of a group of the original model, or a merged set of a first and a second, different group of the original model, generating a merged model based on the mapping and classifying, and classifying a unique entities, using the merged model, by applying consistent hashing to each of the unique entities.