Reconfigurable Analog Neural Network Circuits for Low-Power Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for configuring analog neural networks face challenges in achieving a balance between high accuracy, low power consumption, and low complexity due to suboptimal neural architectures, mapping errors, and process variations in electronic components, leading to energy overhead and precision loss.
Innovation Solution
A method for configuring an analog neural network using a reconfigurable electronic circuit with switchable connections between resistive and summing elements, evaluating multiple configurations directly on hardware to select an optimal architecture without backpropagation-based training, leveraging process variation for diversity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a larger neural network model is used to improve prediction accuracy, then computational precision is improved, but power consumption increases linearly with the number of electronic computational elements
Solution Approach 1:
The patent changes the fundamental parameter of computation from digital to analog domain. By using analog components (resistors for weights, current mirrors for activation functions) to perform neural network computations directly in the analog domain, the system achieves high prediction accuracy with significantly reduced power consumption compared to digital implementations that require large numbers of computational elements
Solution Approach 2:
The patent replaces the mechanical/digital computing system with an analog electronic system. Instead of using digital processors that mechanically execute algorithms, the system uses analog electronic circuits where computations occur naturally through physical laws (Ohm's law, Kirchhoff's laws), eliminating the need for complex digital logic and reducing power consumption
2Device complexity
If engineer's empirical decision is used to design neural network architecture, then device complexity is reduced, but neural architecture quality becomes suboptimal leading to accuracy loss
Solution Approach 1:
The patent enables the neural network architecture to optimize itself automatically through the analog hardware implementation. The analog components naturally adapt to find optimal architectures through physical processes during training, eliminating the need for complex manual design while achieving superior accuracy compared to empirical approaches
3Manufacturing precision
If process variation in electronic components is present, then manufacturing precision is reduced, but mapping errors increase causing computational precision loss
Solution Approach 1:
The patent converts the harmful effect of process variation into a beneficial feature. By using analog components with inherent variations (resistors, transistors) to represent neural network parameters, the system naturally captures the diversity needed for robust learning. The variations are compensated through training procedures that adapt to the specific hardware instance, turning manufacturing imprecision into a source of architectural diversity and improved generalization
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
Eliminates mapping errors and achieves an optimal neural network configuration with improved accuracy, reduced power consumption, and lower complexity by exploiting process variation for architectural diversity.
Implementation Method 1
resistive elements that are nominally equal and exhibit process variation
Implementation Method 2
summing elements that are nominally equal and exhibit process variation
Data Source
AI summary
A method for configuring a neural network, the neural network implemented using analog components, comprises: for each of a plurality of configurations of a reconfigurable electronic circuit for implementing the neural network, configuring connections between analog components of the reconfigurable electronic circuit according to the respective configuration to adapt an architecture of the neural network, and evaluating an input signal using the neural network implemented by the respective configuration of the reconfigurable electronic circuit to obtain an inference result for the input signal. An electronic component for implementing an analog neural network, e.g. to be configured using the disclosed method, comprises a reconfigurable electronic circuit, which comprises a plurality of resistive elements, a plurality of summing elements, and a plurality of switchable connections, each of the switchable connections being switchable to establish a connection between one of the resistive elements and one of the summing elements.