System and method of deploying an artificial neural network on a target device
AssigneesUniversity of Pittsburgh—Of the Commonwealth System of Higher Education, Hefei University of Technology
InventorsYoutao Zhang, Lei Zhao, Jun Yang, Shuai Ding
Life of this patent
Estimating the term…
Abstract
A method of deploying a neural network on a target device method includes extracting a number of device dependent error masks from the target device, using the number of device dependent error masks in a training phase of the neural network to generate a device dependent weight matrix for the neural network, and storing the device dependent weight matrix in the target device for use by the target device to perform inference tasks using the neural network.
Claims17
A method of deploying a neural network on a target drive, comprising:
- extracting a number of device dependent error masks from the target device;
- using the number of device dependent error masks in a training phase of the neural network to generate a device dependent weight matrix for the neural network;
- and
- storing the device dependent weight matrix in the taarget device for use by the target device to perform inference tasks using the neural network, wherein the target device includes a memory portion, and wherein the extracting the number of device dependent error masks from the target device comprises generating the device error mask using the memory portion, and wherein the memory portion comprises dynamic random access memory (DRAM) including a plurality of cells, wherein the generating the number of device dependent error masks using the memory portion comprises choosing a restore duration time such that a number of the cells will be stuck below threshold, and identifying the number of the cells that will be stuck below threshold based on the chosen restore duration time.
A method of deploying a neural network on a target device, comprising:
- extracting a number of device dependent error masks from the target device;
- using the number of device dependent error masks in a training phase of the neural network to generate a device dependent weight matrix for the neural network;
- and
- storing the device dependent weight matrix in the target device for use by the target device to perform inference tasks using the nueral network, wherein the extracting, the using and the storing steps are performed by a manufacturer of the target device.
An electronic device configured to implement a nueral network, comprising:
- a processing portion;
- and
- a memory portion storing a device dependent weight matrix for use by the procession portion to perform inference tasks using the neural network, wherein the device dependent weight matrix is based on a number of device dependent error masks extracted from the electronic device, wherein the memory portion comprises dynamic random access memory (DRAM) including a plurality of cells, wherein the number of device dependent error masks are based on a restore duration time for the DRAM that causes a number of the cells to be stuck below threshold.
Dependencies are inferred from the claim text.