Calculation error correction device and method applied to resistive memory-based neural network accelerator

Patent metadata

Field	Value
Publication number	US-12400727-B2
Application number	US-202117636895-A
Country	US
Kind code	B2
Filing date	Aug 4, 2021
Priority date	Aug 4, 2021
Publication date	Aug 26, 2025
Grant date	Aug 26, 2025

Abstract

The present disclosure provides a calculation error correction device including a first learning unit that trains an effective weight value prediction model for outputting an effective weight value matrix by using learning data in response to an input of the random weight value matrix, an effective weight value calculation unit that inputs a first weight value matrix into the effective weight value prediction model to derive the effective weight value matrix, a second learning unit that applies a second input vector to a target neural network as an input value, applied the effective weight value matrix as a weight value, and trains the weight value such that an output vector follows a result of multiplication of the second input vector and the first weight value matrix, and a control unit that performs matrix-vector multiplication by mapping the first input vector and the trained weight value matrix to the resistive memory.

First claim

The invention claimed is:

1. A calculation error correction device applied to a resistive memory (ReRAM)-based neural network accelerator for performing matrix-vector multiplication, the calculation error correction device comprising:
a processor configured to implement:
a first learning unit configured to train an effective weight value prediction model for outputting an effective weight value matrix in which a voltage drop of a resistive memory is reflected by using learning data including a set input vector and a random weight value matrix,
wherein the learning data includes a target weight value matrix corresponding to the random weight value matrix, the target weight value matrix being previously estimated by simulating the resistive memory based on specification information of the resistive memory including wire resistance values of interconnect lines,
wherein the first learning unit trains the effective weight value prediction model such that an effective weight value matrix output from the effective weight value prediction model follows the target weight value matrix;
an effective weight value calculation unit configured to input a first weight value matrix into the effective weight value prediction model to derive the effective weight value matrix corresponding thereto;
a second learning unit configured to apply a second input vector to a target neural network as an input value, derive an output vector by applying the derived effective weight value matrix to the target neural network as a weight value, train the weight value of the target neural network such that the output vector matches a result of multiplication of the second input vector and the first weight value matrix, and obtain a trained weight value matrix; and
a control unit configured to perform matrix-vector multiplication by mapping a first input vector and the trained weight value matrix to the resistive memory,
wherein the resistive memory includes n bit lines to which n-dimensional input vectors are applied, n×m resistive elements to which an m×n weight value matrix is mapped, and m word lines for deriving m-dimensional output vectors according to the matrix-vector multiplication.

2. The calculation error correction device of claim 1 , wherein
the effective weight value prediction model is implemented by a row-column network (RCN) that is trained by combining a layer of a row component and a layer of a column component respectively corresponding to a bit line and a word line of the resistive memory.

3. The calculation error correction device of claim 1 , wherein
the first learning unit performs learning by using a mean square error calculated by a following equation,

L
=

1
N

⁢



W
^

e

-

W
e



2
2

where L represents the mean square error and N=n×m.

4. A calculation error correction method for a resistive memory (ReRAM)-based neural network accelerator that performs matrix-vector multiplication, the calculation error correction method comprising:
training an effective weight value prediction model for outputting an effective weight value matrix in which a voltage drop of a resistive memory is reflected by using learning data including a set input vector and a random weight value matrix,
wherein the training comprises
providing the learning data including a random weight value matrix and a corresponding target weight value matrix that is previously estimated by simulating the resistive memory based on specification information of the resistive memory including wire resistance values of interconnect lines, and
updating parameters of the effective weight value prediction model such that an output effective weight value matrix the effective weight value prediction model follows the target weight value matrix;

inputting a first weight value matrix into the effective weight value prediction model to derive the effective weight value matrix corresponding thereto;
applying a second input vector to a target neural network as an input value, deriving an output vector by applying the first derived effective weight value matrix to the target neural network as a weight value, training the weight value of the target neural network such that the output vector matches a result of multiplication of the second input vector and the first weight value matrix, and obtaining a trained weight value matrix; and
performing matrix-vector multiplication by mapping a first input vector and the trained weight value matrix to the resistive memory,
wherein the resistive memory includes n bit lines to which n-dimensional input vectors are applied, n×m resistive elements to which an m×n weight value matrix is mapped, and m word lines for deriving m-dimensional output vectors according to the matrix-vector multiplication.

5. The calculation error correction method of claim 4 , wherein
the effective weight value prediction model is implemented by a row-column network (RCN) that is trained by combining a layer of a row component and a layer of a column component respectively corresponding to a bit line and a word line of the resistive memory.

6. The calculation error correction method of claim 4 , wherein,
in the training of an effective weight value prediction model, learning is performed by using a mean square error calculated by a following equation,

L
=

1
N

⁢



W
^

e

-

W
e



2
2

where L represents the mean square error and N=n×m.

Assignees

• Ulsan Nat Inst Science & Tech Unist

Inventors

• Lee Jongeun
• Lee Sugil

Classifications

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06F17/16 Primary

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title

• G11C2029/0411

  Online error correction · CPC title

• G11C29/52 Primary

  Protection of memory contents; Detection of errors in memory contents · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title