ML Data Pipeline Orchestrator for Training Optimization

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current machine learning data processing pipelines face inefficiencies in optimizing the training of models due to high computational costs and the risk of converging to local minima, particularly when dealing with large datasets and limited time budgets.

Innovation Solution

The implementation of a data processing pipeline with an orchestrator node, preparator node, and executor nodes that pool machine learning trials, sub-sample datasets, and adjust trial parameter sampling to optimize computation time and avoid premature convergence, allowing for efficient model training within time constraints.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning trials are executed individually on large datasets, then model accuracy can be improved, but computational cost and time consumption increase significantly

Engineering Contradiction:

Improvemodel accuracyVSAvoidtraining time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the training process by dividing machine learning trials into pools that can be executed in parallel across multiple executor nodes. The orchestrator node coordinates these pools, allowing simultaneous processing of multiple trials on large datasets without sequentially increasing total training time, thus maintaining model accuracy while reducing time loss.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension of parallel execution by distributing trials across multiple executor nodes simultaneously. This transforms the single-threaded sequential training approach into a multi-dimensional parallel processing system, enabling multiple trials to run concurrently on large datasets without proportionally increasing total computation time.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Reliability

If more machine learning trials are executed to explore hyper-parameter space, then model robustness improves, but computational overhead increases

Engineering Contradiction:

Improvemodel robustnessVSAvoidcomputational overhead

Core Design Contradiction:

ReliabilityVSLoss of energy

Solution Approach 1:

The patent merges multiple individual trials into pooled groups that are executed together on the same executor nodes. This combining approach allows more trials to be run for better hyper-parameter exploration and model robustness, while sharing computational resources across the pool to reduce overall computational overhead and energy loss.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The orchestrator node dynamically adjusts trial parameters and pooling configurations to optimize the balance between exploring hyper-parameter space for model robustness and managing computational overhead. By changing execution parameters such as pool size and distribution, the system can execute more trials when resources are available while maintaining efficiency.

Inventive Principle:

Principle #35Parameter changes

3Productivity

If machine learning trials are distributed across multiple executor nodes, then processing capacity increases, but system complexity increases

Engineering Contradiction:

Improveprocessing capacityVSAvoidsystem complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The orchestrator node serves as an intermediary that manages the complexity of distributing trials across multiple executor nodes. It handles task allocation, coordination, and result aggregation, thereby increasing processing capacity through parallel execution while abstracting away the system complexity from the user and simplifying the management of distributed resources.

Inventive Principle:

Principle #24Intermediary (Mediator)

4Speed

If sub-sampling of training dataset is applied, then training speed increases, but model accuracy may decrease

Engineering Contradiction:

Improvetraining speedVSAvoidmodel accuracy

Core Design Contradiction:

SpeedVSMeasurement precision

Solution Approach 1:

The patent applies partial action by using sub-sampled training datasets for initial machine learning trials to achieve fast training speed and quick results. The orchestrator node then uses results from these partial trials to inform subsequent trials that may use larger or complete datasets, balancing the trade-off between training speed and model accuracy by applying sub-sampling only where sufficient.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS11797885B2Optimizations for machine learning data processing pipeline

Publication Date: 2023.10.24 SAP SE

AI summary

A data processing pipeline may be generated to include an orchestrator node, a preparator node, and an executor node. The preparator node may generate a training dataset. The executor node may execute machine learning trials by applying, to the training dataset, a machine learning model and/or a different set of trial parameters. The orchestrator node may identify, based on a result of the machine learning trials, a machine learning model for performing a task. The execution of the data processing pipeline may be optimized. Examples of optimizations include pooling multiple machine learning trials for execution at a single executor node, executing at least some machine learning trials using a sub-sample of the training dataset, and adjusting a proportion of trial parameters sampled from a uniform distribution to avoid a premature convergence to a local minima within the hyper-parameter space for generating the machine learning model.