Neural Network Arithmetic Device Signal Compensation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network arithmetic devices face inefficiencies in processing and compensating output distribution signals, which affects the accuracy and speed of neural network operations.

Innovation Solution

The proposed arithmetic device includes an input distribution signal generation circuit, an output distribution signal generation circuit, and an output distribution signal compensation circuit, which generate and compensate signals based on MAC calculations and activation functions, improving signal processing and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If neural network arithmetic devices use conventional MAC calculation and activation function processing, then basic computation is achieved, but processing speed and accuracy are insufficient

Engineering Contradiction:
Improveprocessing speedVSAvoidaccuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The arithmetic device is divided into distinct functional modules: MAC calculation unit, input distribution signal generation circuit, output distribution signal generation circuit, and output distribution signal compensation circuit. Each module handles a specific aspect of the computation, allowing parallel processing and optimizing both speed and accuracy without requiring a complete redesign of the computational architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The input distribution signal generation circuit performs preliminary processing by generating distribution signals and compensation signals from MAC calculation results before the activation function is applied. This preliminary action prepares the data in an optimized format that enables faster and more accurate subsequent processing by the output distribution signal generation circuit.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If conventional signal processing is used without distribution signal generation and compensation, then device complexity is low, but processing accuracy is insufficient

Engineering Contradiction:
Improveoutput signal accuracyVSAvoidcircuit complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The input distribution signal generation circuit serves multiple functions: it generates input distribution signals for the activation function, generates compensation signals for output distribution signal compensation, and processes MAC calculation results. This multi-functionality increases processing accuracy without requiring separate dedicated circuits for each function, thereby limiting the increase in overall device complexity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The input distribution signal and compensation signal act as intermediaries between the MAC calculation unit and the output distribution signal generation circuit. These intermediate signals enable accurate output signal generation by carrying processed information through the system without requiring direct complex interactions between all components.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If activation function is applied directly to MAC results without distribution signal processing, then processing steps are minimized, but output signal quality is insufficient

Engineering Contradiction:
Improveoutput signal qualityVSAvoidsignal processing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The input distribution signal generation circuit performs preliminary processing of MAC calculation results by generating properly distributed input signals before they are applied to the activation function. This preliminary action ensures that the activation function receives optimally prepared data, improving output signal quality without requiring complex post-processing corrections.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The output distribution signal compensation circuit uses compensation signals generated from MAC results to adjust and improve the quality of output distribution signals. This feedback mechanism continuously refines the output signal quality by comparing expected and actual results and applying corrective adjustments, enhancing reliability without requiring complete redesign of the processing pipeline.

Inventive Principle:
Principle #23Feedback

Data Source

PatentUS20210132953A1Arithmetic devices for neural network
Publication Date: 2021.05.06 SK HYNIX INC
  • US20210132953A1 patent drawing
  • US20210132953A1 patent drawing
  • US20210132953A1 patent drawing

AI summary

An arithmetic device includes an input distribution signal generation circuit, an output distribution signal generation circuit, and an output distribution signal compensation circuit. The input distribution signal generation circuit generates an input distribution signal and a compensation signal based on an arithmetic result signal generated from a result of a multiplying-accumulating (MAC) calculation. The output distribution signal generation circuit applies the input distribution signal to an activation function to generate first and second output distribution signals. The output distribution signal compensation circuit compensates for the first output distribution signal based on the compensation signal, the first output distribution signal, and the second output distribution signal to generate a compensated distribution signal.