Algorithmic architecture co-design and exploration

DWPI Title: Method for matching neural network layouts to hardware architectures by using computer, involves generating neural network algorithm tailored to target hardware architecture according to optimal combination of neural network parameters and hardware architecture parameters
Abstract: A method for matching neural network layouts to hardware architectures is provided. The method comprises iteratively: holding neural network parameters constant while changing a hardware architecture parameters, calculating a first loss value for a combination of the neural network parameters and hardware architecture parameters according to a gradient-based differentiable function within specified resource constraints, holding the hardware architecture parameters constant while changing the neural network parameters, calculating a second loss value for a new combination of parameters within the specified resource constraints, and combining the first loss value and the second loss value to calculate a combined loss value. The above iterative steps are stopped when the combined loss value reaches a specified threshold, and an optimal combination of neural network parameters and hardware architecture parameters is determined according to the combined loss value.
Use: Method for matching neural network layouts to hardware architectures by using a computer.
Advantage: The method enables simplifying optimal design of the neural network depends on the target dataset and the set of primitive operations such as convolutional filters, skip-connections, nonlinearity functions and pooling, how the primitive operations are divided into a neural architecture and optimized and resource constraints such as cost, accuracy and latency. The method enables avoiding more expensive primitive operations more heavily than less expensive operations, when the gradient of the expected resource cost is calculated, so that the penalty can be balanced by how much the primitive operation contributes to classification accuracy.
Novelty: The method (700) involves holding a number of neural network parameters constant while changing a number of hardware architecture parameters (702). A first loss value for a combination of the neural network parameters and hardware architecture parameters is calculated (706) according to a gradient-based differentiable function i.e. resource aware progressive differentiable search function within specified resource constraints i.e. weight. The hardware architecture parameters are hold (708). A second loss value for a new combination of neural network parameters and hardware architecture parameters is calculated (712) according to the gradient-based differentiable function. The first loss value and the second loss value are combined (714) to calculate a combined loss value. A neural network algorithm tailored to a target hardware architecture is generated according to optimal combination of neural network parameters and hardware architecture parameters.
Filed: 8/30/2021
Application Number: US17461847A
Tech ID: SD 15137.1
This invention was made with Government support under Contract No. DE-NA0003525 awarded by the United States Department of Energy/National Nuclear Security Administration. The Government has certain rights in the invention.
Data from Derwent World Patents Index, provided by Clarivate
All rights reserved. Republication or redistribution of Clarivate content, including by framing or similar means, is prohibited without the prior written consent of Clarivate. Clarivate and its logo, as well as all other trademarks used herein are trademarks of their respective owners and used under license.