Pulse shaping unit cell and array for symmetric updating

Patent metadata

Field	Value
Publication number	US-9934838-B1
Application number	US-201715591834-A
Country	US
Kind code	B1
Filing date	May 10, 2017
Priority date	May 10, 2017
Publication date	Apr 3, 2018
Grant date	Apr 3, 2018

Abstract

A memory unit cell and memory array device are provided. The memory unit cell includes a pulse adjustment circuit for providing an adjusted pulse with symmetric weight updating for a given state update in response to an input pulse and state feedback. The memory unit further includes a synapse element having a memory element with hysteresis for storing one of multiple possible states responsive to the adjusted pulse and for providing the state feedback to the pulse adjustment circuit.

First claim

What is claimed is:

1. A memory unit cell, comprising:
a pulse adjustment circuit for providing an adjusted pulse with symmetric weight updating for a given state update in response to an input pulse and state feedback; and
a synapse element having a memory element with hysteresis for storing one of multiple possible states responsive to the adjusted pulse and for providing the state feedback to the pulse adjustment circuit.

2. The memory unit cell of claim 1 , wherein the given state update is comprised in a backpropagation-based non-spiking neural network that is configured to read and update synapse weights along axons and dendrites.

3. The memory unit cell of claim 1 , wherein the state feedback represents a state of a synapse in a neural network.

4. The memory unit cell of claim 1 , wherein the adjusted pulse has an adjusted pulse width relative to the input pulse.

5. The memory unit cell of claim 1 , wherein the memory element with hysteresis comprises a ferroelectric memory device.

6. The memory unit cell of claim 1 , wherein the memory element with hysteresis comprises a Complimentary Metal Oxide Semiconductor (CMOS) circuit.

7. The memory unit cell of claim 1 , wherein the memory element with hysteresis comprises a Resistive Processing Unit (RPU).

8. The memory unit cell of claim 1 , wherein the memory unit cell is configured to selectively operate in a given mode selected from the group consisting of a read out mode and an update mode.

9. The memory unit cell of claim 1 , wherein the multiple possible states comprise three different states, with two of the three different states being partially based on an input polarity.

10. A memory array device, comprising:
a plurality of memory unit cells, each comprising:
a pulse adjustment circuit for providing an adjusted pulse with symmetric weight updating for a given state update in response to an input pulse and state feedback; and
a synapse element having a memory element with hysteresis for storing one of multiple possible states responsive to the adjusted pulse and for providing the state feedback to the pulse adjustment circuit.

11. The memory array device of claim 10 , further comprising a pair of row lines having a row read line and a row update line and a pair of column lines having a column read line and a column update line.

12. The memory array device of claim 1 , further comprising:
a first set of ReConfigurable Integrator and Drivers (RCIDs), each connected to the row read line; and
a second set of RCIDs, each connected to the column read line.

13. The memory array device of claim 10 , further the first set and the second set of RCIDs are configured to perform different functions during different modes of the memory unit cell.

14. The memory array device of claim 12 , further comprising:
a first random pulse generator connected to the row update line; and
a second random pulse generator connected to the column update line.

15. The memory array device of claim 14 , wherein the first random pulse generator and the second random pulse generator are configured to perform a stochastic update process for a backpropagation-based non-spiking neural network.

16. The memory array device of claim 10 , wherein the given state update is comprised in a backpropagation-based non-spiking neural network that is configured to read and update synapse weights along axons and dendrites.

17. The memory array device of claim 10 , wherein the adjusted pulse has an adjusted pulse width relative to the input pulse.

18. The memory array device of claim 10 , wherein the memory element with hysteresis comprises an item selected from the group consisting of a ferroelectric memory device, a Complimentary Metal Oxide Semiconductor (CMOS) circuit, and a Resistive Processing Unit (RPU).

19. The memory array device of claim 10 , further comprising an enable circuit for controlling the pulse adjustment circuit.

20. The memory array device of claim 10 , wherein the enable circuit is configured to provide an enable signal to the pulse adjustment circuit of a given one of the plurality of memory unit cells response to a pulse state provided from a column input and another pulse state provided from a row input.

21. The memory array device of claim 10 , wherein the memory unit cell is configured to selectively operate in a given mode selected from the group consisting of a read out mode and an update mode, and wherein the memory device comprises a vertical conductor line and a horizontal conductor line connected to each of the plurality of memory unit cells for the read out mode, and another vertical conductor line and another horizontal conductor line connected to each of the plurality of memory unit cells for the update mode.

22. The memory array device of claim 10 , further the first set and the second set of RCIDs are configured to perform different functions during different modes of the memory unit cell.

23. The memory array device of claim 10 , further comprising a set of read lines configured for reuse during update and array operations on the memory unit cell.

24. The memory array device of claim 10 , wherein an output of memory array device is based on an integration operation that includes a number of positive pulses and a number of negative pulses applied to the memory array device during an integration window.

25. The memory array device of claim 10 , wherein an input to the memory array device comprises a set of density modulated pulses, and an output from the memory array device comprises a set of integrated voltages.

Assignees

• IBM

Inventors

• Han Jin-Ping
• Sun Xiao
• Yang Teng

Classifications

• G11C13/004

  Reading or sensing circuits or methods · CPC title

• G11C2013/0052

  Read process characterized by the shape, e.g. form, length, amplitude of the read pulse · CPC title

• G11C11/2273 Primary

  Reading or sensing circuits or methods · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06N3/084

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