ML Pipeline Deployment via Composable Model DAGs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Building IT applications involving machine learning and deep learning models is a time-consuming and resource-intensive process due to the need to perform various compute-intensive steps from scratch, including data collection, cleaning, and model building, with most applications being built without leveraging existing composable models for efficient deployment.
Innovation Solution
A method and system for efficient deployment of machine learning and deep learning models' pipelines that utilize existing composable models by creating a directed acyclic graph (DAG) composition of source models and transformation functions, predicting resource consumption, and identifying optimal target systems based on constraints such as throughput, response time, and cost, allowing for modular and accelerated application building.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If IT applications are built from scratch using traditional machine learning and deep learning processes, then the applications can be customized and optimized for specific needs, but the development time and resource consumption increase significantly
Solution Approach 1:
The patent segments the application development process into modular components including pre-built model templates, transformation functions, and configurable parameters. This allows developers to assemble applications from standardized building blocks rather than building everything from scratch, significantly reducing development time while maintaining customization capability through configuration of these modular elements
Solution Approach 2:
The patent implements preliminary action by pre-training machine learning and deep learning model templates beforehand and storing them in a model repository. These pre-trained models can be directly deployed or lightly customized for specific applications, eliminating the need to train models from scratch and dramatically reducing both development time and computational resource requirements
2Reliability
If IT applications are built from scratch with complete data collection, cleaning, and model training processes, then the models can be optimized for specific application requirements, but the computational resources and time required increase substantially
Solution Approach 1:
The patent performs model training and optimization in advance, storing pre-trained model templates in a repository. When deploying applications, these pre-optimized models are retrieved and configured rather than re-trained, significantly reducing computational resource consumption and energy usage while maintaining model reliability and optimization for specific requirements
Solution Approach 2:
The patent enables parameter changes by allowing users to configure and adjust parameters of pre-trained model templates to suit specific application requirements. This approach maintains model reliability and optimization benefits while avoiding the need to re-train models, thereby reducing computational resources and energy consumption
3Productivity
If existing composable models are utilized for application deployment, then the deployment process is accelerated and resource requirements are reduced, but the ability to handle unique application requirements may be limited
Solution Approach 1:
The patent implements universality by creating model templates that are designed to be multi-functional and adaptable to various application scenarios. These universal templates can serve multiple purposes and can be configured through parameter adjustments to handle different unique requirements, thus maintaining both deployment speed and adaptability
Solution Approach 2:
The patent uses parameter changes to enable existing composable models to adapt to unique application requirements. By allowing configuration and adjustment of model parameters, the system maintains high deployment speed while achieving the flexibility needed to handle diverse and unique application scenarios without sacrificing adaptability
Data Source
AI summary
This disclosure relates generally to configuring/building of applications. Typically, a deep learning (DL) application having multiple models composed and interspersed with corresponding transformation functions has no mechanism of efficient deployment on underlying system resources. The disclosed system accelerates the development of application to compose multiple models where each model could be a primitive model or a composite model itself. In an embodiment, the disclosed system optimally deploys a composable model application and transformation functions on underlying resources using performance prediction models, thereby accelerating the development and deployment of the application.