Machine Learning Matrix Operation Selector

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing optimization techniques for sparse matrix operations are inefficient as they require manual selection based on specific matrix characteristics and hardware environments, leading to suboptimal performance due to fixed settings that do not adapt dynamically to varying sparse matrix structures.

Innovation Solution

A machine learning-based system that extracts features from input matrices to dynamically select optimization techniques, such as compute kernels and hyperparameters, using a Deep Q-Network model to maximize speedup in matrix operations like sparse-dense matrix multiplication, allowing for adaptive selection of optimal settings.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If manual selection of optimization techniques is used based on specific matrix characteristics, then the system is simple to operate, but the performance is suboptimal due to fixed settings that do not adapt dynamically

Engineering Contradiction:

Improveadaptability to varying sparse matrix structuresVSAvoidcomplexity of the selection system

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements dynamic selection of optimization techniques by training a machine learning model to automatically choose the best compute kernel and hyperparameters based on the input matrix characteristics. The system transitions from static manual selection to dynamic adaptive selection, where the model learns optimal configurations through reinforcement learning and adjusts selections based on observed performance rewards.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system enables self-service by allowing the machine learning model to autonomously select optimization techniques without requiring manual intervention. The model extracts features from the input matrix, predicts the best compute kernel and hyperparameters, and applies the selected configuration automatically, making the system self-configuring and adaptive to different matrix types.

Inventive Principle:

Principle #25Self-service

2Productivity

If fixed optimization settings are used, then the system is easy to implement, but the computation time is increased due to suboptimal performance

Engineering Contradiction:

Improvecomputation speed of matrix operationsVSAvoidtime for model training and selection

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the machine learning model on a dataset of matrix operations and their optimal configurations before actual use. This offline training phase allows the model to learn optimal compute kernel selections and hyperparameter settings in advance, so that during runtime, the model can quickly select pre-learned optimal configurations without extensive real-time computation, reducing the time loss during actual matrix operations.

Inventive Principle:

Principle #10Preliminary action

3Adaptability or versatility

If deep reinforcement learning is used to select compute kernels and hyperparameters, then the adaptability is improved, but the device complexity increases

Engineering Contradiction:

Improveability to select optimal optimization techniquesVSAvoidcomplexity of the machine learning system

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent introduces a machine learning model as an intermediary between the input matrix and the compute kernel selection process. This intermediary model extracts relevant features from the matrix, predicts the optimal compute kernel and hyperparameters, and bridges the gap between raw input data and optimization decisions, simplifying the overall system architecture while maintaining high adaptability.

Inventive Principle:

Principle #24Intermediary (Mediator)

4Manufacturing precision

If manual optimization selection is used, then the ease of operation is maintained, but the manufacturing precision of the optimization process deteriorates

Engineering Contradiction:

Improveprecision of optimization technique selectionVSAvoidease of implementing adaptive selection

Core Design Contradiction:

Manufacturing precisionVSEase of operation

Solution Approach 1:

The patent replaces the mechanical/manual system of optimization selection with an automated machine learning-based system. Instead of manually analyzing matrix characteristics and selecting compute kernels, the system uses an trained model that automatically extracts features, predicts optimal configurations, and applies the best optimization techniques, thereby improving selection precision while eliminating the need for manual intervention.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS20240211533A1Systems and methods for matrix operation selector based on machine learning

Publication Date: 2024.06.27 SAMSUNG ELECTRONICS CO LTD

AI summary

Systems and methods for matrix operation selector are disclosed. A selection engine receives a matrix as an input and extracts one or more features from the matrix. A machine learning model selects an action based on the one or more features. The action is for performing a matrix operation based on the matrix, and is predicted to satisfy a criterion with respect to a reward. The action is applied for the matrix operation, and a reward is computed based on the applying of the action. The machine learning model is retrained based on the reward.