De-noising using multiple threshold-expert machine learning models
Systems and methods of the present disclosure may be used to improve equalization module architectures for NAND cell read information. For example, embodiments of the present disclosure may provide for de-noising of NAND cell read information using a Multiple Shallow Threshold-Expert Machine Learning Models (MTM) equalizer. An MTM equalizer may include multiple shallow machine learning models, where each machine learning model is trained to specifically solve a classification task (e.g., a binary classification task) corresponding to a weak decision range between two possible read information values for a given NAND cell read operation. Accordingly, during inference, each read sample with a read value within a weak decision range is passed through a corresponding shallow machine learning model (e.g., a corresponding threshold expert) that is associated with (e.g., trained for) the particular weak decision range.
1. A method for flash memory de-noising, comprising:
receiving read information for a memory cell of a memory device;
selecting a threshold expert from a plurality of threshold experts based on the read information, wherein each of the threshold experts corresponds to at least one weak decision voltage range from a plurality of weak decision voltage ranges of the memory cell; and
generating a modified memory channel output from the read information using the selected threshold expert.
2. The method of claim 1 , wherein:
a number of the threshold experts is one less than a number of hard decision thresholds.
3. The method of claim 1 , further comprising:
determining that the read information corresponds to a weak decision voltage range from the plurality of weak decision voltage ranges, wherein the threshold expert is selected based on the determination.
4. The method of claim 1 , further comprising:
transmitting the read information to the selected threshold expert; and
generating an improved value of the channel output.
5. The method of claim 1 , wherein:
each of the weak decision voltage ranges comprises a region between two hard decision thresholds where voltage information for different values of the channel output are likely to overlap.
6. The method of claim 1 , wherein:
the weak decision voltage ranges are separated from each other by a plurality of strong decision voltage ranges, and wherein each of the strong decision voltage ranges corresponds to a value of the channel output.
7. The method of claim 1 , wherein:
each of the threshold experts is trained to operate using read information corresponding to a specific weak decision voltage range.
8. The method of claim 1 , further comprising:
reading a plurality of memory cells including the memory cell;
generating a codeword based on the reading;
performing an error correction code (ECC) operation;
determining that the ECC operation failed; and
generating an updated codeword based on the modified memory channel output.
9. The method of claim 1 , wherein:
the read information comprises one or more of a voltage information of the memory cell, a voltage information of a cell in a neighboring wordline of the memory cell, a voltage of a cell in a same wordline as the memory cell, one or more hard decision (HD) thresholds of a wordline, one or more soft decision (SD) thresholds of a wordline, and a wordline index.
10. The method of claim 1 , further comprising:
generating log likelihood ratio (LLRs) based on the modified memory channel output.
11. The method of claim 1 , further comprising:
receiving second read information for a second memory cell of the memory device;
determining that the second read information corresponds to a second weak decision voltage range from the plurality of weak decision voltage ranges, wherein the second weak decision voltage range is different from the weak decision voltage range;
determining that a second threshold expert corresponds to the second weak decision voltage range; and
selecting the second threshold expert from the plurality of threshold experts, wherein the second threshold expert is different from the threshold expert.
12. The method of claim 1 , further comprising:
receiving a read command from a host device;
reading voltage information from the memory cell in response to the read command, wherein the read information comprises the voltage information; and
transmitting a response to the host device based on a codeword.
13. A system comprising:
a memory device comprising a plurality of memory cells arranged in a plurality of wordlines and a plurality of bitlines; and
a memory controller comprising a processor, wherein the memory controller is configured to:
receive read information for a memory cell of the memory device;
select a threshold expert from a plurality of threshold experts based on the read information, wherein each of the threshold experts corresponds to a different weak decision voltage range from a plurality of weak decision voltage ranges; and
generate an estimate of a channel output from the read information using the selected threshold expert.
14. The system of claim 13 , further comprising:
a host device configured to:
transmit a read command to the memory device, wherein the read information is read by the memory controller based on the read command; and
receive a response from the memory controller based at least in part on the estimate of the channel output.
15. A method for flash memory de-noising, comprising:
reading a memory cell to obtain read information;
generating a codeword based on the reading;
performing an error correction code (ECC) operation on the codeword;
determining that the ECC operation failed;
selecting a threshold expert from a plurality of threshold experts based on the read information and based on the determination that the ECC operation failed, wherein each of the threshold experts corresponds to at least one weak decision voltage range from a plurality of weak decision voltage ranges of the memory cell;
generating a modified memory channel output from the read information using the selected threshold expert; and
generating an updated codeword based on the modified memory channel output.