# US 2021/0279133 A1 · Memory system

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0279133 A1
- Document ID: US20210279133A1
- Status: Published application
- Kind: A1
- Application number: 17/184,166
- Filed: 2021-02-24
- Published: 2021-09-09
- Earliest claimed: 2020-03-05
- Assignees: Kioxia Corporation
- Applicants: Kioxia Corporation
- Inventors: Yuma YOSHINAGA; Tomoya KODAMA; Osamu TORII; Kenichiro FURUTA; Ryota YOSHIZAWA
- Web page: https://patentgrep.com/patent/US20210279133A1

## CPC

- G06F 11/1068: {in sector programmable memories, e.g. flash disk (G06F11/1072 takes precedence)}
- H03M 13/1111: {Soft-decision decoding, e.g. by means of message passing or belief propagation algorithms}
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- H03M 13/458: {by updating bit probabilities or hard decisions in an iterative fashion for convergence to a final decoding result}
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- H03M: CODING; DECODING; CODE CONVERSION IN GENERAL (using fluidic means F15C4/00; optical analogue/digital converters G02F7/00; coding, decoding or code conversion, specially adapted for particular applications, see the relevant subclasses, e.g. G01D, G01R, G06F, G06T, G09G, G10L, G11B, G11C, H04B, H04L, H04M, H04N; ciphering or deciphering for cryptography or other purposes involving the need for secrecy G09C)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A memory system includes a non-volatile memory and a memory controller. The memory controller is configured to read a received word from the non-volatile memory, estimate noise by using a plurality of different models for estimating the noise included in the received word to obtain a plurality of noise estimation values, select one noise estimation value from the plurality of noise estimation values, update the received word by using a value obtained by subtracting the selected noise estimation value from the read received word, and decode the updated received word by using a belief-propagation method.

## Claims

### Claim 1

A memory system comprising:
a non-volatile memory; and
a memory controller configured to:
read a received word from the non-volatile memory,
estimate noise based on a plurality of different models for estimating the noise in the received word to provide a plurality of noise estimation values,
select one noise estimation value from the plurality of noise estimation values,
update the received word using a value obtained by subtracting the selected noise estimation value from the read received word, and
decode the updated received word using a belief-propagation method.

### Claim 2 (depends on 1)

The memory system according to claim 1,
wherein the memory controller is configured to:
decode the read received word using the belief-propagation method,
obtain the plurality of noise estimation values when using the plurality of models when decoding by the belief-propagation method fails,
select one noise estimation value from the plurality of noise estimation values,
update the received word by using a value based on subtracting the selected noise estimation value from the read received word, and
decode the updated received word using a belief-propagation method.

### Claim 3 (depends on 1)

The memory system according to claim 1,
wherein the memory controller is configured to:
read the received word from the non-volatile memory as a soft decision value,
obtain the plurality of noise estimation values based on the soft decision value and a hard decision value, the hard decision value calculated based on the soft decision value,
select one noise estimation value from among the plurality of obtained noise estimation values,
update the received word by a value based on subtracting the selected noise estimation value from the received word,
decode the updated received word using the belief-propagation method, and
obtain the plurality of noise estimation values using the plurality of models when decoding by the belief-propagation method fails.

### Claim 4 (depends on 1)

The memory system according to claim 1,
wherein the memory controller is configured to select one noise estimation value among the plurality of noise estimation values, the selection being based on distance between a value obtained by subtracting the noise estimation value from the read received word and a most likely decoding word being less than the other noise estimation values.

### Claim 5 (depends on 1)

The memory system according to claim 1,
wherein the memory controller is configured to select one noise estimation value that is smaller than the other noise estimation values among the plurality of noise estimation values.

### Claim 6 (depends on 1)

The memory system according to claim 1,
wherein each of the plurality of models is a convolutional neural network (CNN).

### Claim 7 (depends on 1)

The memory system according to claim 1,
wherein each of the plurality of models is learned based on different pieces of attribute information indicating attributes of the non-volatile memory.

### Claim 8 (depends on 7)

The memory system according to claim 7,
wherein the attribute information includes at least one of the number of word lines, the number of blocks, or the number of pages of the non-volatile memory.

### Claim 9 (depends on 1)

The memory system according to claim 1,
wherein each of the plurality of models is learned based on a decoding result when a hard error is generated.

### Claim 10

A method of operating a memory system including a non-volatile memory, the method comprising:
reading a received word from the non-volatile memory;
estimating noise based on a plurality of different models for estimating the noise in the received word to provide a plurality of noise estimation values;
selecting one noise estimation value from the plurality of noise estimation values,
updating the received word using a value obtained by subtracting the selected noise estimation value from the read received word, and
decoding the updated received word using a belief-propagation method.

### Claim 11 (depends on 10)

The method according to claim 10, further comprising:
decoding the read received word using the belief-propagation method;
obtaining the plurality of noise estimation values using the plurality of models when decoding by the belief-propagation method fails;
selecting one noise estimation value from the plurality of noise estimation values;
updating the received word by using a value based on subtracting the selected noise estimation value from the read received word; and
decoding the updated received word using a belief-propagation method.

### Claim 12 (depends on 10)

The method according to claim 10, further comprising:
reading the received word from the non-volatile memory as a soft decision value;
obtaining the plurality of noise estimation values based on the soft decision value and a hard decision value, the hard decision value calculated based on the soft decision value;
selecting one noise estimation value from among the plurality of obtained noise estimation values;
updating the received word by a value based on subtracting the selected noise estimation value from the received word;
decoding the updated received word using the belief-propagation method, and
obtaining the plurality of noise estimation values using the plurality of models when decoding by the belief-propagation method fails.

### Claim 13 (depends on 10)

The method according to claim 10, further comprising selecting one noise estimation value among the plurality of noise estimation values, the selection being based on a distance between a value obtained by subtracting the noise estimation value from the read received word and a most likely decoding word being less than the other noise estimation values.

### Claim 14 (depends on 10)

The method according to claim 10, further comprising selecting one noise estimation value that is smaller than the other noise estimation values, among the plurality of noise estimation values.

### Claim 15 (depends on 10)

The method according to claim 10,
wherein each of the plurality of models is a convolutional neural network (CNN).

### Claim 16 (depends on 10)

The method according to claim 10, wherein each of the plurality of models is learned based on different pieces of attribute information indicating attributes of the non-volatile memory.

### Claim 17 (depends on 16)

The method according to claim 16, wherein the attribute information includes at least one of the number of word lines, the number of blocks, or the number of pages of the non-volatile memory.

### Claim 18 (depends on 10)

The method according to claim 10, wherein each of the plurality of models is learned based on a decoding result when a hard error is generated.
