Neural Network Convolution Overhead Reduction via Template Storage

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional dedicated processing apparatuses for neural networks are inefficient in processing convolution operations between CNN feature amounts due to overhead from setting up data other than coefficients.

Innovation Solution

An information processing apparatus with a coefficient storage unit, feature storage unit, and convolution operation unit that stores filter coefficients and feature data, allowing for efficient convolution operations between feature data and filter coefficients, as well as between feature data and template feature data to compute correlation data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional dedicated processing apparatus is applied to correlation operation between CNN feature amounts, then the apparatus can perform convolution operations, but processing efficiency is reduced due to overhead of setting up data other than coefficients

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidoverhead of setting up data

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The coefficient storage unit is designed to store both filter coefficients and template feature data, allowing the same storage resource to serve multiple functions. This eliminates the need for separate data setup structures and reduces overhead when performing correlation operations between CNN feature amounts.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

Template feature data is pre-stored in the coefficient storage unit before correlation operations are performed. This preliminary storage action eliminates the need for dynamic data setup during operation, reducing overhead and improving processing efficiency.

Inventive Principle:

Principle #10Preliminary action

2Speed

If conventional dedicated processing apparatus is applied to correlation operation between CNN feature amounts, then the apparatus can compute correlation values, but processing speed is reduced due to data setup overhead

Engineering Contradiction:

Improveprocessing speedVSAvoidtime for data setup overhead

Core Design Contradiction:

SpeedVSLoss of time

Solution Approach 1:

Template feature data is pre-loaded into the coefficient storage unit during initialization or preprocessing stages. This preliminary action ensures that when correlation operations are executed, the data is already in place, eliminating setup time during the critical computation phase and improving processing speed.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The coefficient storage unit serves itself by storing both filter coefficients and template feature data in the same location. This self-service capability eliminates the need for separate data setup operations and reduces the time required to prepare data for correlation computations.

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS20220392207A1Information processing apparatus, information processing method, and non-transitory computer-readable storage medium

Publication Date: 2022.12.08 CANON KK

AI summary

An information processing apparatus operable to perform computation processing in a neural network comprises a coefficient storage unit configured to store filter coefficients of the neural network, a feature storage unit configured to store feature data, a storage control unit configured to store in the coefficient storage unit a part of previously obtained feature data as template feature data, a convolution operation unit configured to compute new feature data by a convolution operation between feature data stored in the feature storage unit and filter coefficients stored in the coefficient storage unit, and compute, by a convolution operation between feature data stored in the feature storage unit and the template feature data stored in the coefficient storage unit, correlation data between the feature data stored in the feature storage unit and the template feature data.