Neural Network Data Processing Device Chip Area Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In neural network data processing, the integration of digital and analog circuits to perform feedforward and feedback computing, such as back propagation, leads to increased circuit size due to the need for frequent digital-to-analog and analog-to-digital conversions, which occupies more chip area.
Innovation Solution
A data processing device with a neural network structure that includes a hidden layer comprising a digital-to-analog converter, a first neuron circuit, a second neuron circuit, and a comparator, where the first and second neuron circuits have potential holding circuits connected to the same bit line, allowing for product-sum operations without explicit digital-to-analog or analog-to-digital conversions, thereby reducing the need for conversion circuits and power sources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If digital and analog circuits are integrated to perform feedforward and feedback computing in neural networks, then computing functionality is improved, but circuit size increases due to the need for digital-to-analog and analog-to-digital converters
Solution Approach 1:
The patent merges digital and analog computing circuits into a unified neural network processing architecture. Digital circuits perform feedforward computing while analog circuits handle feedback computing (backpropagation), eliminating the need for separate digital-to-analog and analog-to-digital converters. This integration reduces circuit size while maintaining full computing functionality for both forward propagation and backward error propagation.
Solution Approach 2:
The patent creates a multi-functional circuit architecture where the same hardware components serve multiple purposes. The digital circuit performs feedforward computing and the analog circuit performs feedback computing, allowing a single integrated system to handle both essential neural network operations without requiring additional conversion circuits, thereby reducing overall circuit size.
2Adaptability or versatility
If digital-to-analog and analog-to-digital converters are added for signal conversion, then signal compatibility is improved, but occupied chip area increases
Solution Approach 1:
The patent extracts and eliminates the unnecessary digital-to-analog and analog-to-digital converter components from the neural network architecture. By carefully designing the system to use digital circuits for feedforward computing and analog circuits for feedback computing, the invention removes these conversion components entirely, reducing chip area while maintaining signal compatibility through direct digital-analog cooperation between circuit stages.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This configuration suppresses the increase in circuit size by eliminating the need for digital-to-analog and analog-to-digital converters and their power sources, allowing for efficient data processing without frequent signal conversions, thus optimizing chip area usage.
Implementation Method 1
Each of the first potential holding circuit and the second potential holding circuit comprises a first transistor, a second transistor, and a third transistor
Implementation Method 2
analog signal can be held and thus analog computing is suitable... a leakage current (off-state current) of a transistor including an oxide semiconductor (OS transistor) in the off-state is extremely small
Data Source
AI summary
To provide a data processing device using a neural network that can suppress increase in the occupied area of a chip. A product-sum operation circuit is formed using a transistor including an oxide semiconductor having an extremely small off-state current. Signals are input to and output from the product-sum operation circuits included in a plurality of hidden layers through comparators. The outputs of the comparators are used as digital signals to be input signals for the next-stage hidden layer. The combination of a digital circuit and an analog circuit can eliminate the need for an analog-to-digital converter or a digital-to-analog converter which occupies a large area of a chip.