Neural network system

• US 8,694,451 B2
• Filed: 09/15/2011
• Issued: 04/08/2014
• Est. Priority Date: 03/18/2009
• Status: Active Grant

• Current Assignee: Piece Future Pte., Ltd.
• Original Assignee: Panasonic Corporation (Panasonic Holdings Corporation)
• Inventors: Sasagawa, Yukihiro
• Primary Examiner(s): Chaki, Kakali
• Assistant Examiner(s): Seck, Ababacar
• Application Number:
  US13/233,196
• Publication Number:
  US 20120005141A1
• Time in Patent Office: 936 Days
• Field of Search: 706/25
• US Class Current: 706/25
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A neural network system that can minimize circuit resources for constituting a self-learning mechanism and be reconfigured into network configurations suitable for various purposes includes a neural network engine that operates in a first and a second operation mode and performs an operation representing a characteristic determined by setting network configuration information and weight information with respect to the network configuration, and a von Neumann-type microprocessor that is connected to the neural network engine and performs a cooperative operation in accordance with the first or the second operation mode together with the neural network engine. The von Neumann-type microprocessor recalculates the weight information or remakes the configuration information as a cooperative operation according to the first operation mode, and sets or updates the configuration information or the weight information set in the neural network engine, as a cooperative operation according to the second operation mode.

Claims

1. A neural network system comprising:

a neural network engine that operates in a first operation mode and a second operation mode and performs an operation representing a characteristic determined by setting network configuration information indicating a network configuration to be formed and weight information indicating a weight with respect to the network configuration; and

a von Neumann-type microprocessor that performs a cooperative operation in accordance with the first operation mode or the second operation mode together with said neural network engine, said von Neumann-type microprocessor being connected to said neural network engine,

wherein said neural network engine includes;

a neural processing element that performs neural signal processing;

a routing switch;

a memory containing control information of said neural processing element;

a memory containing control information of said routing switch;

an interconnect, and

said von Neumann-type microprocessor recalculates the weight information or remake the network configuration information as a cooperative operation according to the first operation mode; and

sets or updates the network configuration information or the weight information set in said neural network engine, as a cooperative operation according to the second operation mode.

2. The neural network system according to claim 1, wherein said von Neumann-type microprocessor executes a program of emulating an error propagation process of said neural network engine as a cooperative operation according to the first operation mode, to thereby recalculate the weight information or remake the network configuration information.

7. The neural network system according to claim 1, wherein said neural network engine includes a plurality of said neural processing elements, a plurality of said routing switches and a plurality of said interconnects, said plurality of routing switches and plurality of interconnects allowing for connection of said plurality of neural processing elements to one another.

8. The neural network system according to claim 7, wherein said plurality of neural processing elements and said plurality of routing switches are alternately arranged in a checkerboard pattern.

9. The neural network system according to claim 7, wherein any one of said plurality of neural processing elements can be connected to any other one of said plurality of neural processing elements via said plurality of routing switches and said plurality of interconnects.

3. An operation method to be performed by a neural network system including a neural network engine that operates in a first operation mode and a second operation mode and performs an operation representing a characteristic determined by setting network configuration information and weight information with respect to the network configuration, and a von Neumann-type microprocessor connected to the neural network engine for performing a cooperative operation in accordance with the first operation mode or the second operation mode together with the neural network engine, said method comprising causing the von Neumann-type microprocessor to:

perform a first cooperative operation including recalculating the weight information or remaking the network configuration information as a cooperative operation according to the first operation mode; and

perform a second cooperative operation including setting or updating the network configuration information or the weight information set in the neural network engine, as a cooperative operation according to the second operation mod;

wherein said neural network engine includes neural processing elements, routing switches and interconnects, said routing switches and interconnects allowing for connection of said neural processing elements to one another.

4. The method according to claim 3, wherein said causing the von Neumann-type microprocessor to perform a second cooperative operation includes causing the von Neumann-type microprocessor to:

perform a first operation mode change including changing the operation mode of the neural network engine to the second operation mode;

set or update the network configuration information or the weight information set in the neural network engine, as a cooperative operation according to the second operation mode; and

perform a second operation mode change including changing the operation mode of the neural network engine to an operation mode different from the second operation mode.

5. The method according to claim 3, further comprising causing the von Neumann-type microprocessor to decide whether the operation mode of the neural network engine is the first operation mode,

wherein said causing the von Neumann-type microprocessor to perform a first cooperative operation includes causing the von Neumann-type microprocessor to:

execute emulation of the error propagation process as a cooperative operation according to the first operation mode, in the case where the von Neumann-type microprocessor has decided that the operation mode of the neural network engine is the first operation mode; and

recalculate the weight information or remake the network configuration information on the basis of a result of the emulation.

6. The method according to claim 5, further comprising causing the von Neumann-type microprocessor to execute a program in the case where the von Neumann-type microprocessor has decided that the operation mode of the neural network engine is different from the first operation mode and the second operation mode.

10. The method according to claim 3, wherein said neural processing elements and said routing switches are alternately arranged in a checkerboard pattern.

11. The method according to claim 3, wherein any one of said neural processing elements can be connected to any other one of said neural processing elements via said routing switches and said interconnects.