Neural Network Inference Error Protection via Checksum Verification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Neural network inference calculations on hardware platforms are susceptible to transient errors due to increased integration density, leading to potential data corruption in external working memory and communication links.

Innovation Solution

A method is implemented where input data and redundancy information are read into the internal working memory, checked for integrity, and then used for inference calculations. The work results and redundancy information are written to the external working memory, allowing for error detection and correction, thereby reducing the impact of transient errors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If integration density of hardware platform is increased to reduce costs and energy consumption, then computing power efficiency is improved, but probability of transient computational errors increases

Engineering Contradiction:

Improveenergy consumption per unit of computing powerVSAvoidprobability of transient computational errors

Core Design Contradiction:

Use of energy by moving objectVSReliability

Solution Approach 1:

The patent applies preliminary action by calculating redundancy information (checksums) for input data before processing and storing it in external working memory. This preparatory step enables subsequent error detection without affecting the main computation flow, allowing the system to benefit from high integration density while maintaining reliability through pre-established verification mechanisms.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback by continuously verifying data integrity through checksum comparison. When data is read from external working memory, the calculated redundancy information is compared against stored checksums, and any discrepancies trigger error correction procedures. This closed-loop feedback mechanism allows the system to operate at high integration densities while automatically detecting and correcting transient errors.

Inventive Principle:

Principle #23Feedback

2Quantity of substance

If external working memory is used to store large neural network data, then memory capacity is improved, but data integrity against transient errors deteriorates

Engineering Contradiction:

Improvememory capacity for neural network dataVSAvoiddata integrity in external memory

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The patent applies copying by creating and storing redundancy information (checksums) as a separate copy alongside the main data in external working memory. This copy serves as a verification reference that can detect errors in the original data without requiring additional physical memory capacity for the neural network itself, thus maintaining large memory capacity while ensuring data integrity.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent uses preliminary action by pre-calculating and storing checksums for all input data before neural network processing begins. This advance preparation enables rapid error detection during inference without impacting processing speed, allowing the system to utilize large external memory capacities while maintaining reliability through pre-established verification data.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If redundancy information is calculated and stored for all input data, then error detection capability is improved, but memory requirements and processing time increase

Engineering Contradiction:

Improveerror detection capabilityVSAvoidprocessing time for redundancy calculation

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent applies partial action by calculating redundancy information selectively for input data that requires verification, rather than uniformly for all data. The system determines which data portions need checksum calculation based on their importance and error sensitivity, thereby reducing overall processing overhead while maintaining adequate error detection capability for critical neural network operations.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS12253918B2Inference calculation for neural networks with protection against memory errors

Publication Date: 2025.03.18 ROBERT BOSCH GMBH

AI summary

A method for operating a hardware platform for the inference calculation of a layered neural network. In the method: a first portion of input data which are required for the inference calculation of a first layer of the neural network and redundancy information relating to the input data are read in from an external working memory into an internal working memory of the computing unit; the integrity of the input data is checked based on the redundancy information; in response to the input data here being identified as error-free, the computing unit carries out at least part of the first-layer inference calculation for the input data to obtain a work result; redundancy information for the work result is determined, based which the integrity of the work result can be verified; the work result and the redundancy information are written to the external working memory.