MEMS Analog Edge AI Computing Without ADC/DAC Bottlenecks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital computers for edge AI computing are inefficient due to analog-to-digital and digital-to-analog conversion burdens and processing bottlenecks, and in-situ training for edge applications is challenging, particularly with conventional stochastic gradient descent approaches.
Innovation Solution
Analog computing using MEMS devices, specifically MEMS CTRNN, which perform edge AI computations in continuous time, eliminating the need for analog-to-digital conversion and enabling in-situ training through difference target propagation, allowing adaptation to hardware variations and drift.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If digital computers are used for edge AI computing, then computation can be performed, but power consumption is high and processing speed is limited due to ADC/DAC conversion burdens
Solution Approach 1:
The patent replaces the digital computing system with an analog computing system using MEMS devices. Instead of converting analog signals to digital and processing them digitally, the system directly processes analog signals through the inherent capacitive properties of MEMS devices, eliminating ADC/DAC conversion steps and reducing power consumption while increasing processing speed
Solution Approach 2:
The patent changes the computational approach from digital binary operations to analog continuous-time capacitive operations. By utilizing the natural capacitive behavior of MEMS devices and implementing difference target propagation in the analog domain, the system achieves efficient AI computation without the overhead of digital conversion and processing
2Adaptability or versatility
If conventional SGD learning approach is used for in-situ training, then model training can be performed, but implementation in analog hardware is very difficult
Solution Approach 1:
The patent changes the training algorithm from conventional SGD with error backpropagation to difference target propagation (DTP). DTP is specifically designed to be compatible with analog hardware by using local gradient estimation and avoiding the complex global backpropagation required by SGD, thereby enabling in-situ training in analog MEMS devices
Solution Approach 2:
The patent implements a self-training capability where the analog MEMS system can perform in-situ training locally without requiring external digital processing. The difference target propagation algorithm allows the system to adapt its weights directly in the analog domain using local information, eliminating the need for complex external training infrastructure
Data Source
AI summary
An analog system for edge artificial intelligence computing includes a first plurality of analog edge devices configured to receive an input analog signal and to output a first plurality of output analog signals, a second plurality of analog edge devices configured to receive the first plurality of output analog signals and to output a second plurality of output analog signal, and one or more memory devices in communication with the first plurality of analog edge devices and the second plurality of analog edge devices, and configured to store weight parameters, the weight parameters being adjustable based on time constants of the first plurality of analog edge devices or the second plurality of analog devices, or both. The second plurality of output analog signals are multiplied by the weight parameters to obtain a plurality of weighted analog signals.