Early-Exit Neural Network Radar Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Processing radar measurement frames on resource-constrained devices is computationally expensive, limiting the implementation of radar processing algorithms on low-power embedded compute circuitry.

Innovation Solution

Implementing an early-exit neural network architecture that includes multiple layers coupled with early-exit branches, allowing for selective abortion of further processing and reducing computational resources by using early-exit estimates for radar measurement frames.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If full processing pipeline is used for all radar inputs, then measurement precision is improved, but use of energy increases and productivity decreases

Engineering Contradiction:

Improvetarget observable estimate accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system dynamically adjusts the processing depth for each radar input based on early-exit branch predictions. When the early-exit estimate indicates sufficient accuracy, processing is terminated early, reducing energy consumption. When higher precision is needed, the full processing pipeline executes. This dynamic adaptation resolves the contradiction between maintaining measurement precision and reducing power consumption.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes the processing parameter (processing depth) based on the confidence level of early-exit estimates. By monitoring the evolution of early-exit estimates across sequential inputs and comparing similarity scores, the system adjusts whether to use early-exit processing or full pipeline processing, thereby optimizing the balance between accuracy and energy usage.

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If full processing pipeline is used for all radar inputs, then measurement precision is improved, but productivity decreases

Engineering Contradiction:

Improvetarget observable estimate accuracyVSAvoidprocessing throughput

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The system dynamically selects processing depth based on real-time assessment of input similarity and early-exit estimate quality. For inputs similar to previously processed ones with adequate early-exit estimates, processing is shortened to maintain productivity. For dissimilar or complex inputs requiring higher precision, full processing pipeline executes, ensuring measurement accuracy while optimizing overall throughput.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The processing parameter (pipeline depth) is changed based on similarity score thresholds and early-exit confidence levels. When similarity exceeds thresholds and early-exit estimates are sufficient, reduced processing depth is applied to boost productivity. When precision requirements demand, full depth processing is applied, resolving the contradiction between productivity and measurement precision.

Inventive Principle:

Principle #35Parameter changes

3Use of energy by moving object

If early-exit processing is used, then use of energy is reduced, but measurement precision may deteriorate

Engineering Contradiction:

Improvepower consumptionVSAvoidtarget observable estimate accuracy

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The system uses feedback from early-exit branch outputs to determine whether full processing is necessary. By monitoring the evolution of early-exit estimates and comparing similarity scores between sequential inputs, the system feedback-controls the processing depth. When early-exit estimates show sufficient accuracy and high similarity to previous inputs, energy-efficient early-exit processing is selected. When precision thresholds are not met, feedback triggers full pipeline processing, resolving the contradiction between energy reduction and measurement precision.

Inventive Principle:

Principle #23Feedback

4Productivity

If selective processing is implemented, then productivity is improved, but device complexity increases

Engineering Contradiction:

Improveprocessing throughputVSAvoidnetwork architecture complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The processing network is segmented into multiple independent early-exit branches coupled to different layers of the processing pipeline. Each branch can independently provide estimates and control processing continuation. This segmentation enables selective processing for different inputs based on their specific characteristics, improving productivity while managing complexity through modular, independent processing paths.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20240310485A1Early-exit neural networks for radar processing

Publication Date: 2024.09.19 INFINEON TECHNOLOGIES AG

AI summary

In accordance with an embodiment, a method, includes: obtaining a plurality of radar measurement frames; and processing, in a deep neural network, inputs to the deep neural network, the inputs being based on the plurality of radar measurement frames, The processing includes: providing an estimate of a target observable using a processing pipeline of the deep neural network, where the processing pipeline comprises a plurality of layers; and providing early-exit estimates of the target observable using respective early-exit branches of the deep neural network, where two or more layers of the plurality of layers are coupled with the respective early-exit branches of the deep neural network.