Pre-trained Model Retrieval for Low-Latency Decision Making

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current recommendation systems, especially empirical model-based systems, face challenges in handling complex systems with many variables, leading to memory footprint and scalability issues, and require substantial resources and time for training, making them unsuitable for quick decision-making in events involving new data.

Innovation Solution

A computer-implemented method that pre-trains multiple models based on specific hypotheses, reducing memory footprint and enabling parallel processing, allowing for immediate model switching during events without retraining, thus addressing memory, scalability, and latency problems.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If empirical models are trained to handle complex systems with many variables, then measurement precision and recall are improved, but memory footprint and device complexity increase substantially

Engineering Contradiction:
Improvemodel accuracyVSAvoidmemory footprint
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the complex modeling task by training multiple specialized models, each handling a specific hypothesis or scenario, rather than using a single large model. This segmentation reduces the memory footprint of each individual model while maintaining overall system accuracy through ensemble processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates multiple copies of simplified models, each trained on different subsets of data or different hypotheses. These model copies can be processed in parallel and their results combined, achieving high accuracy without requiring a single large memory-intensive model.

Inventive Principle:
Principle #26Copying

2Reliability

If empirical models are trained comprehensively, then reliability is improved, but training time and loss of time increase, making them unsuitable for quick decision-making

Engineering Contradiction:
Improvedecision accuracyVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary training of multiple specialized models in advance, each focused on specific hypotheses or scenarios. This preliminary action allows the system to have pre-trained models ready for rapid deployment during events, eliminating the need for time-consuming training during critical decision-making moments.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements a dynamic model selection mechanism where the system can switch between different pre-trained models based on the current event scenario. This dynamic approach allows rapid adaptation to new situations without retraining, maintaining reliability while minimizing training time.

Inventive Principle:
Principle #15Dynamics

3Device complexity

If a single model is used to handle all scenarios, then device complexity is reduced, but adaptability decreases when new data or scenarios emerge

Engineering Contradiction:
Improvemodel managementVSAvoidresponse to new data
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent implements a dynamic model ensemble where multiple specialized models can be selectively activated based on the input scenario. This dynamic configuration allows the system to adapt to new data and scenarios by selecting or combining appropriate models, maintaining high versatility while managing complexity through modular organization.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3188040B1Retrieval of outcomes of precomputed models
Publication Date: 2021.05.05 DASSAULT SYSTEMES SA
  • EP3188040B1 patent drawingFigure 1
  • EP3188040B1 patent drawingFigure 2
  • EP3188040B1 patent drawingFigure 3

AI summary

The invention notably relates to a computer-implemented method for retrieving model outcomes in the course of an event. The method comprises providing variables of a first set of observations of similar events, providing a multiplicity of models, indexing the multiplicity of models with the variables of the first set of observations, querying a model according to one or more variables, and returning, as a result of the query, a model.