Battery-based neural network weights

Patent metadata

Field	Value
Publication number	US-11106966-B2
Application number	US-201715457379-A
Country	US
Kind code	B2
Filing date	Mar 13, 2017
Priority date	Mar 13, 2017
Publication date	Aug 31, 2021
Grant date	Aug 31, 2021

Abstract

A controllable resistive element and methods for controlling the resistance of the same include a resistor layer formed in contact with a shared read/write electrode and a read electrode, the resistor layer having a resistivity that depends on a concentration of charge carrier ions. An electrolyte layer is formed on the resistor layer. A reservoir layer is formed on the electrolyte layer and in contact with a write electrode.

First claim

The invention claimed is:

1. A controllable resistive element, comprising:
a resistor layer formed in direct contact with a shared read/write electrode, that is used for write operation and read operation, and in direct contact with a read electrode, which forms a first conductive path with the shared read/write electrode, the resistor layer having a resistivity that depends on a concentration of charge carrier ions;
an electrolyte layer formed on the resistor layer; and
a reservoir layer formed on the electrolyte layer and in contact with a write electrode, which forms a second conductive path with the shared read-write electrode.

2. The controllable resistive element of claim 1 , wherein the resistor layer, the electrolyte layer, and the reservoir layer form a thin-film lithium-ion battery.

3. The controllable resistive element of claim 1 , wherein a current pulse between the write electrode and the shared read/write electrode causes a change in the electrical resistivity of the resistor layer.

4. The controllable resistive element of claim 1 , wherein a current pulse on the write electrode that reduces the charge carrier ion concentration in the resistor layer causes an decrease in the electrical resistivity of the resistor layer.

5. The controllable resistive element of claim 1 , wherein the read electrode is formed in contact with a sidewall of the resistor layer.

6. The controllable resistive element of claim 5 , wherein the write electrode is formed above the resistor layer and wherein the shared read/write electrode is formed below the resistor layer.

7. The controllable resistive element of claim 1 , wherein the resistor layer and the reservoir layer comprise a lithium-based battery material.

8. The controllable resistive element of claim 1 , wherein a voltage applied to the read electrode of each weight provides a current at the shared read/write electrode of said weight that is proportional to an electrical resistivity of the resistor layer of said weight.

9. A neural network processing system, comprising:
a battery-based neural network, comprising:
a set of one or more input neurons;
one or more sets of hidden neurons, each set of hidden neurons comprising one or more hidden neurons;
a set of one or more output neurons; and

a plurality of weight arrays, each weight array being disposed between sets of neurons and each weight comprising:
a resistor layer formed in direct contact with a shared read/write electrode, that is used for write operation and read operation, and in direct contact with a read electrode, which forms a first conductive path with the shared read/write electrode, the resistor layer having a resistivity that depends on a concentration of a charge carrier;
an electrolyte layer formed on the resistor layer; and
a reservoir layer formed on the electrolyte layer and in contact with a write electrode, which forms a second conductive path with the shared read-write electrode; and

a weight control module configured to provide a voltage applied to the respective weights of the weight arrays.

10. The neural network processing system of claim 9 , wherein the resistor layer, the electrolyte layer, and the reservoir layer form a thin-film lithium-ion battery.

11. The neural network processing system of claim 9 , wherein the weight control module applies a current pulse between the write electrode and the shared read/write electrode that decreases the lithium concentration in the resistor layer to decrease the electrical resistivity of the resistor layer.

12. The neural network processing system of claim 9 , wherein the weight control module applies a current pulse between on the write electrode and the shared read/write electrode that increases the lithium concentration in the resistor layer to increase the electrical resistivity of the resistor layer.

13. The neural network processing system of claim 9 , wherein a voltage applied to the read electrode of each weight provides a current at the shared read/write electrode of said weight that is proportional to an electrical resistivity of the resistor layer of said weight.

14. The neural network processing system of claim 9 , wherein each read electrode is formed on a sidewall of a respective resistor layer, each write electrode is formed above a respective resistor layer, and each shared read/write electrode is formed below a respective resistor layer.

15. The neural network processing system of claim 9 , wherein the weight control module is further configured to wait a predetermined amount of time before applying a read voltage to the respective weights to allow charge carrier ions in the resistor layer to settle.

16. A neural network processing system, comprising:
a battery-based neural network, comprising:
a set of one or more input neurons;
one or more sets of hidden neurons, each set of hidden neurons comprising one or more hidden neurons;
a set of one or more output neurons; and

a plurality of weight arrays, each weight array being disposed between sets of neurons and each weight comprising:
a resistor layer formed in direct contact with a shared read/write electrode, that is used for write operation and read operation, and in direct contact with a read electrode, which forms a first conductive path with the shared read/write electrode, the resistor layer having a resistivity that depends on a concentration of a charge carrier;
an electrolyte layer formed on the resistor layer; and
a reservoir layer formed on the electrolyte layer and in contact with a write electrode, which forms a second conductive path with the shared read-write electrode; and

a weight control module configured to provide a voltage applied to the respective weights of the weight arrays and to wait a predetermined amount of time before applying a read voltage to the respective weights to allow charge carrier ions in the resistor layer to settle.

Assignees

• IBM

Inventors

• De Souza Joel P
• Lee Yun Seog
• Li Ning
• Sadana Devendra K

Classifications

• G06N3/065

  Analogue means · CPC title

• G06N3/063 Primary

  using electronic means · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/084 Primary

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