Deep neural network perforance analysis on shared memory accelerator systems

Patent metadata

Field	Value
Publication number	US-2019080232-A1
Application number	US-201715699320-A
Country	US
Kind code	A1
Filing date	Sep 8, 2017
Priority date	Sep 8, 2017
Publication date	Mar 14, 2019
Grant date	—

Abstract

A Deep Neural Networks (DNN) analysis method, system, and computer program product include characterizing a space of possible configurations for a DNN, evaluating a metric-of-interest for a configuration of the possible configurations, and searching the space to identify a configuration of the possible configurations that maximizes the metric-of-interest.

First claim

What is claimed is:

1 . A computer-implemented Deep Neural Networks analysis method, the method comprising:
characterizing a space of possible configurations for a deep neural network (DNN); evaluating a metric-of-interest for a configuration of the possible configurations; and searching the space to identify a configuration of the possible configurations that maximizes the metric-of-interest.

2 . The computer-implemented method of claim 1 , wherein the characterizing characterizes the space based on an input description of the DNN.

3 . The computer-implemented method of claim 1 , wherein the DNN comprises a sequence of convolutional and fully connected layers whose parameters are represented using a tuple <in,out,ij,mb,kij>.

4 . The computer-implemented method of claim 1 , wherein, for each combination of spatial work division across cores of the computer, the configuration with a best record is used, and passed on to a next stage.

5 . The computer-implemented method of claim 1 , wherein the possible configurations comprise at least one of a computation configuration and a data-partitioning configuration.

6 . The computer-implemented method of claim 1 , wherein the evaluating evaluates the metric-of-interest based on a predetermined hardware specification.

7 . The computer-implemented method of claim 6 , wherein the predetermined hardware specification comprises:
a plurality of processor cores; a plurality of memory elements; and a plurality of data-links.

8 . The computer-implemented method of claim 7 , further comprising periodically sending control information and receiving status updates to and from the cores and a memory enabling the system to realize all computations in the DNN.

9 . The computer-implemented method of claim 1 , embodied in a cloud-computing environment.

10 . A computer program product for Deep Neural Networks analysis, the computer program product comprising a computer-readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform:
characterizing a space of possible configurations for a deep neural network (DNN); evaluating a metric-of-interest for a configuration of the possible configurations; and searching the space to identify a configuration of the possible configurations that maximizes the metric-of-interest.

11 . The computer program product of claim 10 , wherein the characterizing characterizes the space based on an input description of the DNN.

12 . The computer program product of claim 10 , wherein the DNN comprises a sequence of convolutional and fully connected layers whose parameters are represented using a tuple <in,out,ij,mb,kij>.

13 . The computer program product of claim 10 , wherein, for each combination of spatial work division across cores, the configuration with a best record is used, and passed on to a next stage.

14 . The computer program product of claim 10 , wherein the possible configurations comprise at least one of a computation configuration and a data-partitioning configuration.

15 . The computer program product of claim 10 , wherein the evaluating evaluates the metric-of-interest based on a predetermined hardware specification.

16 . The computer program product of claim 15 , wherein the predetermined hardware specification comprises:
a plurality of processor cores; a plurality of memory elements; and a plurality of data-links.

17 . The computer program product of claim 16 , wherein the computer program product further stores instructions to cause the computer to perform:
periodically sending control information and receiving status updates to and from the cores and a memory enabling the system to realize all computations in the DNN.

18 . A Deep Neural Networks analysis system, said system comprising:
a processor; and a memory, the memory storing instructions to cause the processor to perform:
characterizing a space of possible configurations for a deep neural network (DNN);
evaluating a metric-of-interest for a configuration of the possible configurations; and
searching the space to identify a configuration of the possible configurations that maximizes the metric-of-interest.

19 . The system of claim 18 , wherein the characterizing characterizes the space based on an input description of the DNN.

20 . The system of claim 18 , embodied in a cloud-computing environment.

Assignees

• IBM

Inventors

• Choi Jungwook
• Srinivasan Vijayalakshmi
• Venkataramani Swagath

Classifications

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/105

  Shells for specifying net layout · CPC title

• G06N3/063

  using electronic means · CPC title