Tunneling Devices Replace Op-Amps in Neural Network Circuits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hardware-based artificial neural networks face challenges in reducing the size and cost of electric circuits due to the use of operational amplifiers, which also consume significant power.
Innovation Solution
The implementation of nonlinear tunneling devices, such as Zener diodes, in electric circuits to perform activation and neuronal functions, reducing the need for operational amplifiers and allowing for simpler, more efficient circuit designs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If operational amplifiers are used in hardware-based artificial neural networks, then the circuit can perform neural network functions, but the size, cost, and power consumption increase
Solution Approach 1:
The patent extracts and removes operational amplifiers from the neural network circuit, replacing them with simple nonlinear tunneling devices. This extraction eliminates the high power consumption and large size associated with operational amplifiers while retaining the essential neural network computation capabilities through the tunneling devices' inherent nonlinear characteristics.
Solution Approach 2:
The patent employs inexpensive nonlinear tunneling devices that can be easily replaced or reconfigured, substituting for expensive and power-hungry operational amplifiers. These tunneling devices provide the necessary nonlinear activation functions at minimal cost and power consumption, enabling practical hardware neural network implementations.
2Reliability
If operational amplifiers are used in hardware-based artificial neural networks, then the circuit can perform neural network functions, but the circuit size and cost increase
Solution Approach 1:
The patent removes operational amplifiers from the circuit architecture, replacing them with compact nonlinear tunneling devices. This extraction dramatically reduces the circuit footprint while maintaining neural network computational functionality through the tunneling devices' nonlinear current-voltage characteristics that naturally implement activation functions.
Solution Approach 2:
The patent uses small, inexpensive nonlinear tunneling devices that occupy minimal circuit area compared to operational amplifiers. These devices provide the necessary nonlinear processing in a compact form factor, enabling dense integration of neural network layers without requiring large operational amplifier circuits.
3Reliability
If operational amplifiers are used in hardware-based artificial neural networks, then the circuit can perform neural network functions, but the manufacturing complexity and cost increase
Solution Approach 1:
The patent extracts operational amplifiers from the neural network implementation, replacing them with simple nonlinear tunneling devices that have fewer components and simpler fabrication requirements. This extraction reduces manufacturing complexity and cost while preserving the essential neural network functions through the tunneling devices' inherent nonlinear behavior.
Solution Approach 2:
The patent employs inexpensive, easily manufactured nonlinear tunneling devices that can be produced using standard semiconductor fabrication processes. These devices replace costly and complex operational amplifiers, significantly reducing manufacturing costs and simplifying production while maintaining neural network computational capabilities.
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 approach reduces the size and cost of neural network circuits, lowers power consumption, and enables easier modification and replacement of components, making hardware-based neural networks more practical and efficient.
Implementation Method 1
a plurality of non-linear devices formed in each of a plurality of column wires configured to receive an input signal, wherein at least one of the non-linear devices has a characteristic of activation function
Implementation Method 2
at least one of the non-linear device has a characteristic of neuronal function... the characteristic of the neuronal function includes a sigmoid or a rectifying linear unit (ReLU)
Data Source
AI summary
Systems and methods for mitigating defects in a crossbar-based computing environment are disclosed. In some implementations, an apparatus comprises: a plurality of row wires; a plurality of column wires connecting between the plurality of row wires; a plurality of non-linear devices formed in each of a plurality of column wires configured to receive an input signal, wherein at least one of the non-linear devices has a characteristic of activation function and at least one of the non-linear devices has a characteristic of neuronal function.