Robot Control Model Scaling for Accuracy-Cost Balance

Solution Overview

Problem

Existing robot systems face a trade-off between inference accuracy and calculation cost, where achieving higher accuracy often requires larger models with more parameters and multiplications, increasing computational costs.

Innovation Solution

A robot system that uses tensor decomposition to adjust the width of weight matrices in machine learning models, allowing for dynamic model size adjustment based on desired hardware performance, such as power consumption and memory capacity, while maintaining inference accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a model with a large number of parameters and multiplications is used to achieve more accurate inference processing, then inference accuracy is improved, but calculation cost increases

Engineering Contradiction:

Improveinference accuracyVSAvoidcalculation cost

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent segments the weight matrix into multiple sub-matrices and applies low-rank decomposition to each segment. This allows the model to maintain high inference accuracy through detailed parameter representation while reducing calculation cost by replacing full-rank matrices with products of lower-rank matrices, thereby resolving the contradiction between accuracy and computational burden.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent dynamically changes the rank parameter of the decomposed weight matrices based on available computational resources. By adjusting the rank parameter, the system can flexibly balance inference accuracy against calculation cost, enabling the model to adapt to different hardware constraints while maintaining optimal performance.

Inventive Principle:

Principle #35Parameter changes

2Quantity of substance

If the calculation cost is reduced by using a smaller model, then hardware performance is improved, but inference accuracy deteriorates

Engineering Contradiction:

Improvecalculation costVSAvoidinference accuracy

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent segments the weight matrix into multiple sub-matrices and applies low-rank decomposition to each segment. This allows the model to maintain high inference accuracy through detailed parameter representation while reducing calculation cost by replacing full-rank matrices with products of lower-rank matrices, thereby resolving the contradiction between accuracy and computational burden.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces dynamic adjustment of the decomposition rank based on hardware performance and resource constraints. The system can adaptively select the appropriate rank level to match available computational resources, ensuring optimal balance between calculation cost reduction and maintaining sufficient inference accuracy for the given hardware environment.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS11654553B2Robot system and driving method

Publication Date: 2023.05.23 KK TOSHIBA

AI summary

A robot system according to an embodiment includes one or more processors. The processors acquire first input data predetermined as data affecting an operation of a robot. The processors calculate a calculation cost of inference processing using a machine learning model for inferring control data used for controlling the robot, on the basis of the first input data. The processors infer the control data by the machine learning model set according to the calculation cost. The processors control the robot using the inferred control data.