Neural Network Layer Operation Stop Signal Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing devices, particularly in-vehicle ECUs, face challenges in reducing power consumption due to the need for numerous devices to determine and stop operations in convolutional neural networks (CNNs), leading to wasteful power usage and increased circuit scale.
Innovation Solution
A computing device with a numerical value operation unit, rounding operation unit, and operation stop signal generation unit that determines whether to stop operations based on input data and omitted bits, allowing for independent stop signals for each layer, reducing unnecessary computations and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rounding operation is performed to reduce bit length after product-sum operation, then bit length is reduced and expressible range is limited, but power consumption increases due to numerous devices needed for operation stop determination
Solution Approach 1:
The patent performs preliminary determination of whether rounding will cause all significant bits to become zero before executing the product-sum operation. By checking in advance whether the omitted bit count exceeds the significant bit count, the system avoids performing unnecessary computations and rounding operations that would consume power, thus resolving the contradiction between maintaining data precision and reducing power consumption.
2Use of energy by moving object
If operation stop determination devices are increased to control power consumption, then power consumption is reduced, but circuit scale increases
Solution Approach 1:
The patent merges the operation stop determination function into the existing rounding operation unit. Instead of adding separate determination devices for each computing unit, the rounding unit itself determines whether to stop operations by checking if omitted bits exceed significant bits. This integration reduces circuit scale while maintaining power consumption benefits, as the same hardware component performs both rounding and operation stop determination.
Solution Approach 2:
The rounding operation unit performs self-determination of whether subsequent operations are necessary. By autonomously checking whether all significant bits will become zero after rounding and stopping operations accordingly, the system eliminates the need for external control devices, thereby reducing circuit scale while achieving power consumption reduction.
3Measurement precision
If bit length is increased to maintain data precision, then data expression capability is improved, but power consumption increases due to more extensive computations
Solution Approach 1:
The patent applies partial action by performing rounding and operation stop determination only when necessary. Instead of continuously maintaining full bit length precision throughout all operations, the system selectively applies rounding with operation stop only when the determined conditions are met (when omitted bits exceed significant bits), thus reducing power consumption without permanently compromising data expression capability.
Data Source
AI summary
The present invention aims to reduce power consumption in the operation based upon a recognition device (1000) including, in a neural network of multiple layers that output type of an object and existing coordinates based on external environment information, a selector (103) that selects input data to convolution operation units (107-1 to L) from external environment information; convolution operation units (107-1 to L) configured by a plurality of layers connected in cascade; and a parameter storage unit (109) that stores a weight parameter of each layer, a cumulative addition count of each layer, and an omitting bit number of each layer. The recognition device includes operation stop signal generation units (116-1 to L) that transmit one or more stop signals for stopping some or all of the computing units of the convolution operation units (107-1 to L) to the convolution operation units (107-1 to L) for each layer.