Resistive processing unit

Patent metadata

Field	Value
Publication number	US-10248907-B2
Application number	US-201514887564-A
Country	US
Kind code	B2
Filing date	Oct 20, 2015
Priority date	Oct 20, 2015
Publication date	Apr 2, 2019
Grant date	Apr 2, 2019

Abstract

Embodiments are directed to a two-terminal resistive processing unit (RPU) having a first terminal, a second terminal and an active region. The active region effects a non-linear change in a conduction state of the active region based on at least one first encoded signal applied to the first terminal and at least one second encoded signal applied to the second terminal. The active region is configured to locally perform a data storage operation of a training methodology based at least in part on the non-linear change in the conduction state. The active region is further configured to locally perform a data processing operation of the training methodology based at least in part on the non-linear change in the conduction state.

First claim

What is claimed is:

1. A two-terminal resistive processing unit (RPU) comprising:
a first terminal;
a second terminal; and
an active region having a resistance;
wherein the RPU represents a neuron in a neural network, and the resistance of the active region represents a weight of the neuron according to a training methodology applied to the neural network;
wherein the active region is configured to change the resistance in a stochastic manner wherein the change in the conduction state comprises a non-linear change based on a first encoded voltage signal applied to the first terminal and a second encoded voltage signal applied to the second terminal; and
wherein the active region is further configured to locally perform a data processing operation of the training methodology using the changed resistance.

2. The two-terminal RPU of claim 1 , wherein the change in the resistance is based at least in part on a result of the data processing operation.

3. The two-terminal RPU of claim 1 , wherein:
the active region is further configured to locally perform a data storage operation of the training methodology based at least in part on the non-linear change in the resistance; and
the active region is further configured to locally perform the data processing operation of the training methodology based at least in part on the non-linear change in the resistance.

4. The two-terminal RPU of claim 1 , wherein the training methodology comprises at least one of: an online neural network training; a matrix inversion; and a matrix decomposition.

5. A two-terminal resistive processing unit (RPU) comprising:
a first terminal;
a second terminal; and
an active region having a resistance, the RPU represents a neuron in a neural network, and the resistance of the active region represents a weight of the neuron according to a training methodology applied to the neural network;
wherein the active region is configured to effect a non-linear change in the resistance in a stochastic manner based on at least one first encoded signal applied to the first terminal and at least one second encoded signal applied to the second terminal;
wherein the active region is further configured to locally perform a data storage operation of a training methodology based at least in part on the non-linear change in the resistance; and
wherein the active region is further configured to locally perform a data processing operation of the training methodology based at least in part on the non-linear change in the resistance.

6. The two-terminal RPU of claim 5 , wherein the encoding of the at least one first encoded signal and the at least one second encoded signal comprises a stochastic sequence of pulses.

7. The two-terminal RPU of claim 5 , wherein the encoding of the at least one first encoded signal and the at least one second encoded signal comprise s a magnitude modulation.

8. The two-terminal RPU of claim 5 , wherein the non-linear change comprises a rectifying non-linear change or a saturating non-linear change.

9. The two-terminal RPU of claim 5 , wherein the non-linear change comprises an exponential non-linear change.

10. A trainable crossbar array comprising:
a set of conductive row wires;
a set of conductive column wires configured to form a plurality of crosspoints at intersections between the set of conductive row wires and the set of conductive column wires; and
a two-terminal resistive processing unit (RPU) at each of the plurality of crosspoints, the RPU represents a neuron in a neural network represented by the trainable crossbar array, wherein a resistance of an active region of the RPU represents a weight of the neuron;
wherein the RPU is configured to locally perform a data storage operation of a training methodology applied to the neural network represented by the trainable crossbar array, the data storage operation performed by a change in the resistance based on a first encoded voltage signal applied to a first terminal and a second encoded voltage signal applied to a second terminal;
wherein the RPU is further configured to locally perform a data processing operation of the training methodology using the changed resistance.

11. The array of claim 10 , wherein:
the data storage operation comprises a change in the resistance that is based at least in part on a result of the data processing operation; and
the change in the resistance comprises a non-linear change based on at least one first encoded signal applied to the first terminal and at least one second encoded signal applied to the second terminal.

12. The array of claim 11 , wherein:
the active region is further configured to locally perform the data storage operation of the training methodology based at least in part on the non-linear change in the resistance; and
the active region is further configured to locally perform the data processing operation of the training methodology based at least in part on the non-linear change in the resistance.

13. The array of claim 10 , wherein the training methodology comprises at least one of:
an online neural network training;
a matrix inversion; and
a matrix decomposition.

Assignees

• IBM

Inventors

• Gokmen Tayfun
• Kim Seyoung
• Vlasov Yurii

Classifications

• G06N3/065 Primary

  Analogue means · CPC title

• G06N3/084 Primary

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

• G06N3/063 Primary

  using electronic means · CPC title

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/0635

  Physics · mapped topic