Scalable architecture for implementing maximization algorithms with resistive devices

Patent metadata

Field	Value
Publication number	US-10380485-B2
Application number	US-201514963667-A
Country	US
Kind code	B2
Filing date	Dec 9, 2015
Priority date	Sep 29, 2015
Publication date	Aug 13, 2019
Grant date	Aug 13, 2019

Abstract

In some aspects, a method may include initializing a first array and a second array with a random voltage value, passing a forward pass by pulsing an input voltage value from an input of the first array and an input of the second array, and reading output voltage values at an output of the first array and an output of the second array. The method may further include passing a backward pass into the inputs of both of the first and second arrays, and reading voltage values at the inputs of the first and second arrays. The method may further include updating, with the first array, a first matrix update on the first array, updating, with the second array, a first matrix update on the second, and updating, with the second array, a second matrix update on the second array.

First claim

What is claimed is:

1. A method for performing matrix operations, the method comprising:
providing a plurality of resistive devices arranged in a network of resistive device arrays, the network comprising a first array and a second array, wherein each resistive device located in a diagonal of the first array is disconnected, and wherein each resistive device located in an off-diagonal of the second array is disconnected;
initializing the first array and the second array with a random voltage;
sending, to a first resistive device in the first array, a first voltage;
receiving, from the first resistive device, a first output current, wherein the first output current is based at least in part on a first conductance of the first resistive device;
sending, to a second resistive device in the second array, a second voltage;
receiving, from the second resistive device, a second output current, wherein the second output current is based at least in part on a second conductance of the second resistive device;
updating the first conductance of the first resistive device; and
updating the second conductance of the second resistive device.

2. The method of claim 1 , wherein the first conductance and the second conductance are updated in parallel at a constant time independent of a size of the first array or the second array.

3. The method of claim 1 , wherein the first array is comprised of resistive devices functioning as weighted connections between neurons.

4. The method of claim 1 , wherein
the first array is configured to hold a plurality of off-diagonal terms of a weight matrix; and
the second array is configured to hold a plurality of diagonal terms of the weight matrix.

Assignees

• Int Business Machines Int
• IBM

Inventors

• Gokmen Tayfun
• Kim Seyoung

Classifications

• G06N3/065

  Analogue means · CPC title

• G06N3/084 Primary

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

• G06N3/049 Primary

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title

• G06N20/00

  Machine learning · CPC title

• G06N3/08

  Learning methods · CPC title