Declarative Model for ML Labeler Configuration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity of configuring machine learning (ML) models for training and inference requires expert knowledge, and existing platforms and frameworks are not user-friendly, especially for determining the best platforms and algorithms for labeling tasks, leading to a need for a more accessible and flexible configuration method.
Innovation Solution
A declarative model is used to configure a labeling platform, allowing users to specify configurations in a human-readable and machine-readable format, defining processing graphs that include human and machine learning labelers, and enabling easy modification and understanding of configurations through a declarative language, abstracting away the complexity of ML platforms, frameworks, and algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing ML platforms and frameworks are used for configuring labeling tasks, then the system provides comprehensive ML capabilities and algorithms, but the configuration complexity increases and requires expert knowledge
Solution Approach 1:
The patent introduces a declarative model as an intermediary layer between the user and the ML platform. This model provides a standardized, platform-agnostic configuration interface that mediates between user requirements and the underlying complex ML infrastructure, eliminating the need for users to directly interact with multiple platform-specific configuration systems
Solution Approach 2:
The declarative model serves as a universal configuration framework that can describe processing graphs for multiple different ML platforms and algorithms through a single standardized language. This multi-functional approach allows the same declarative model to work across different ML ecosystems without requiring separate configuration systems for each platform
2Measurement precision
If multiple ML platforms and algorithms are tested for labeling tasks, then the system can find the best solution for the task, but the user must develop detailed configurations for each platform and algorithm
Solution Approach 1:
The patent creates a standardized declarative representation (a copy or abstraction) of the processing graph that can be reused across multiple ML platforms. Instead of creating unique detailed configurations for each platform, users work with a single declarative model that can be instantiated or adapted to different underlying ML systems, significantly reducing configuration effort while maintaining the ability to test multiple algorithms
3Ease of operation
If a declarative model is used to configure the labeling platform, then the configuration process is simplified and platform-agnostic, but the system must interpret and implement the declarative model into platform-specific configurations
Solution Approach 1:
The patent extracts the platform-specific implementation details from the user-facing configuration interface. The declarative model contains only the essential processing graph definitions, while the platform-specific configuration logic is separated into the interpretation and implementation layer. This extraction allows users to work with simplified declarative syntax while the complex translation to platform-specific configurations is handled automatically by the system
Data Source
AI summary
Systems, methods, and computer program products for configuring labelers, including machine learning labelers are provided. A declarative model describes a processing graph of labelers for a use case at logical level. The declarative model defines a configuration for each labeler in the processing graph of labelers in a declarative language. Each labeler in the processing graph of labelers can represent a wrapper on executable code. The declarative model is interpreted to implement the processing graph of labelers, which is executed to label a set of data records.