Hybrid decoder, operating method thereof, and neural network training method therefor

Abstract

Embodiments relate to a hybrid decoder. The hybrid decoder according to the embodiments comprises a transformer-based decoder and further comprises at least one of a pre-decoder and a post-decoder, wherein the pre-decoder may decode received data and output a codeword of a first decoding, the transformer-based decoder may decode the data or the codeword of the first decoding and output a codeword of a second decoding, and the post-decoder may decode the codeword of the second decoding on the basis of a result of the second decoding.

Description

Hybrid decoder, its operation method, and neural network learning method for the same

[0001] The embodiments relate to a hybrid decoder, a method of operation thereof, and a neural network learning method for the same.

[0002] With the advancement of deep learning, neural decoders based on neural network architectures have been proposed to decode error correction codes. Among them, neural decoders based on transformer structures are attracting attention for their excellent decoding performance.

[0003] A transformer-based decoder takes the magnitude |y| of a received vector y and the syndrome value s(y) as input and outputs an estimated value of the multiplied noise. During the decoding process, the transformer-based decoder utilizes an attention mechanism to leverage the relationships between the information of each bit. In this case, the parity check matrix H is used as a mask matrix for the attention matrix to reflect the information of the code. Such transformer decoders are generally trained using the binary cross-entropy loss function and perform error correction on a bit-by-bit basis.

[0004] Meanwhile, hybrid decoding refers to a method that improves decoding performance by combining two or more decoding techniques. A representative example of a hybrid decoding method is the combination of soft decoding and hard decoding. Soft decoding uses a received soft vector y as input to provide high reliability, but it requires high decoding complexity. On the other hand, hard decoding determines the received soft vector y as a binary value before decoding, resulting in lower decoding complexity and enabling faster processing. By combining these two methods, this hybrid decoding approach can utilize the advantages of both the high accuracy of soft decoding and the processing speed of hard decoding.

[0005] The embodiment proposes a hybrid decoder structure to achieve superior performance compared to a conventional transformer-based error correction decoder, and proposes a method for training a transformer-based decoder for this purpose.

[0006] A hybrid decoder may be provided, comprising a transformer-based decoder and further comprising at least one of a pre-decoder or a post-decoder, wherein the pre-decoder decodes received data and outputs a codeword of a first decoding, the transformer-based decoder decodes the data or the codeword of the first decoding and outputs a codeword of a second decoding, and the post-decoder decodes the codeword of the second decoding based on the result of the second decoding.

[0007] The above-mentioned pre-decoder and post-decoder may correspond to one of a soft decision decoder and a hard decision decoder.

[0008] If the number of errors included in the codeword of the first decoding is less than or equal to a predetermined number, the first decoding result of the data is output as the final decoding result, and if the number of errors exceeds the predetermined number, the data can be input to the transformer-based decoder.

[0009] If the above post-decoder is a hard decision decoder, the codeword of the second decoding can be input to the post-decoder in response to the failure of the above second decoding.

[0010] If the above post-decoder is a soft decision decoder, the probability of an error for a codeword of the above second decoding can be input to the above post-decoder in response to the failure of the above second decoding.

[0011] The value obtained by multiplying the error probability for the codeword of the second decoding by the sign of the data can be input to the post-decoder.

[0012] The above transformer-based decoder can be trained using a cross-entropy-based loss function for a failure dataset in which the decoding of codewords by the prior decoder fails among the training data.

[0013] A method of operation of a hybrid decoder may be provided, comprising: a step of decoding received data in a pre-decoder to output a codeword of a first decoding; a step of decoding the data or the first decoding codeword in a transformer-based decoder based on the result of the first decoding to output a codeword of a second decoding; and a step of decoding the codeword of the second decoding in a post-decoder based on the result of the second decoding.

[0014] A method for training a transformer-based decoder may be provided, comprising: a step of acquiring a training data set; and a step of training the transformer-based decoder using a cross-entropy-based loss function for a failure data set among the training data in which decoding of a codeword by a pre-decoder combined with the transformer-based decoder fails, wherein the transformer-based decoder is configured as a hybrid decoder having a combined structure of the pre-decoder, the transformer decoder, and the post-decoder.

[0015] If the above-mentioned post-decoder is a hard decision decoder, the cross-entropy-based loss function may include the assumption that data in which the error of the codeword of the second decoding exceeds a predetermined number is determined to be a failure in the decoding of the codeword of the second decoding.

[0016] If the above posterior decoder is a soft decision decoder, the cross-entropy-based loss function may include a conditional probability that the decoding of the codeword is determined to have failed based on the error probability of the codeword by the posterior decoder.

[0017] A hybrid decoder may be provided, comprising: a pre-decoder that decodes received data and outputs a codeword of a first decoding; and a transformer-based decoder that decodes the data or the first decoding codeword and outputs a codeword of a second decoding based on the result of the first decoding.

[0018] A hybrid decoder may be provided, comprising: a transformer-based decoder that decodes data and outputs a codeword of a first decoding; and a post-decoder that decodes a codeword by the decoding based on the result of the first decoding.

[0019] According to the embodiment, a hybrid decoder structure is proposed to derive superior performance compared to a conventional transformer-based error correction decoder, and a method for training a transformer-based decoder for this purpose can be provided.

[0020] Figure 1 is a diagram illustrating a hybrid decoder structure in an embodiment.

[0021] FIG. 2 is a flowchart illustrating the operation method of a hybrid decoder in an embodiment.

[0022] FIG. 3 is a flowchart illustrating a method for learning a transformer-based decoder of a hybrid decoder in an embodiment.

[0023] FIG. 4 is a diagram illustrating the configuration of a receiving device for hybrid decoding in an embodiment.

[0024] FIGS. 5a to 5d are drawings illustrating the performance of a hybrid decoder in BCH code decoding in an embodiment.

[0025] FIGS. 6a and 6b are drawings illustrating an example of a two-stage hybrid decoder in an embodiment.

[0026] Hereinafter, embodiments are described in detail with reference to the attached drawings. However, various modifications may be made to the embodiments, and thus the scope of the patent application is not limited or restricted by these embodiments. It should be understood that all modifications, equivalents, and substitutions to the embodiments are included within the scope of the rights.

[0027] The terms used in the embodiments are for illustrative purposes only and should not be interpreted as intended to be limiting. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "comprising" or "having" are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0028] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the embodiments pertain. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0029]

[0030] In addition, when describing with reference to the attached drawings, identical components are assigned the same reference numeral regardless of drawing symbols, and redundant descriptions thereof are omitted. In describing the embodiments, if it is determined that a detailed description of related prior art could unnecessarily obscure the essence of the embodiments, such detailed description is omitted.

[0031] In addition, terms such as first, second, A, B, (a), (b), etc., may be used when describing the components of the embodiments. These terms are intended merely to distinguish the components from other components, and the nature, order, or sequence of the components is not limited by these terms. Where it is stated that a component is "connected," "combined," or "connected" to another component, it should be understood that while the component may be directly connected or connected to the other component, another component may also be "connected," "combined," or "connected" between each component.

[0032] Components included in any one embodiment and components having common functions shall be described using the same names in other embodiments. Unless otherwise stated, the description in any one embodiment may also apply to other embodiments, and specific descriptions shall be omitted to the extent of overlap.

[0033]

[0034] Figure 1 is a diagram illustrating a hybrid decoder structure in an embodiment.

[0035] The hybrid decoder has a structure in which a pre-decoder (110), a transformer-based decoder (120), and a post-decoder (130) are connected. Data y is data received via wired or wireless communication from a data receiving device and is the target for decoding.

[0036] The pre-decoder (110) and the post-decoder (130) are conventional decoders and may correspond to either a hard decision decoder or a soft decision decoder. For example, both the pre-decoder (110) and the post-decoder (130) may be hard decision decoders or soft decision decoders, and each of the pre-decoder (110) and the post-decoder (130) may correspond to either a hard decision decoder or a soft decision decoder.

[0037] The structure according to the embodiment can correct different error patterns from received data by connecting the decoder in three stages, and thus can reduce errors in decoding.

[0038] If decoding is successful in the decoder of the preceding stage and the correct codeword is estimated, the decoding process is terminated, and if decoding fails, decoding can be performed in the decoder of the next stage.

[0039] For example, if the correct codeword is estimated from the pre-decoder (110), the decoding process is terminated, and if decoding fails, the data can be input to the transformer-based decoder (120). The data is decoded in the transformer-based decoder (120), and if decoding is successful, the decoding process is terminated, and if decoding fails, the decoding process can be performed in the post-decoder (130).

[0040] In the following description, the pre-decoder (110) and post-decoder (130) will be described as being hard decision decoders.

[0041] In the pre-decoder (110), data y can be decoded. The pre-decoder (110) can decode binary data containing noise, perform error correction equal to its error correction capability, and estimate a data codeword of 0 or 1. If an error less than or equal to a predetermined number is detected in the codeword estimated by the pre-decoder (110), the decoding is determined to be successful, and the codeword can be output.

[0042] Conversely, if an error exceeding a predetermined number is detected in the codeword estimated by the pre-decoder (110), the decoding is determined to have failed, and the decoding process can be continued in the transformer-based decoder (120) combined with the pre-decoder (110). The transformer-based decoder (120) can acquire and decode data y when decoding fails in the pre-decoder (110). This does not request retransmission of data, but rather inputs the same data that was input into the pre-decoder (110) into the transformer-based decoder (120). Instead of receiving and decoding the retransmitted data, the communication waiting time can be effectively reduced by decoding the same code again using hybrid decoding.

[0043] Since the transformer-based decoder (120) performs decoding only when decoding fails in the pre-decoder (110), in order to improve the error correction performance of the transformer-based decoder (120), it can be trained by considering the data that failed to decode from the pre-decoder (110) combined with the transformer-based decoder (120). This training method will be explained in detail later.

[0044] If decoding is successful in the transformer-based decoder (120), the decoded codeword can be output to terminate the decoding process. On the other hand, if decoding fails, the decoding process can be performed through a post-decoder (130) combined with the transformer-based decoder (120). At this time, the post-decoder (130) can obtain the codeword decoded from the transformer-based decoder (120) and decode the corresponding codeword. Since the codeword decoded from the transformer-based decoder (120) can be considered to have some errors corrected, efficient decoding can be performed by inputting the codeword decoded from the transformer-based decoder (120) into the post-decoder (130).

[0045] In the following description, the pre-decoder (110) and post-decoder (130) will be described as being soft decision decoders.

[0046] In the pre-decoder (110), data y can be decoded. The pre-decoder (110) can decode the input data of the LLR (Log Likelihood Ratio) and output probability information for estimating a data codeword of 0 or 1. If an error of less than or equal to a predetermined number is detected in the codeword estimated by the pre-decoder (110), the decoding is determined to be successful and the codeword can be output.

[0047] If an error exceeding a predetermined number is detected in the codeword estimated by the pre-decoder (110), the decoding is determined to have failed, and the decoding process can be performed in the transformer-based decoder (120) combined with the pre-decoder (110). The transformer-based decoder (120) can acquire and decode data y when decoding fails in the pre-decoder (110). Alternatively, the output of the pre-decoder (110) can be used as the input to the transformer-based decoder (120). The output of the soft decision decoder contains error probability information for the decoded codeword and is output as a real value, so it can be used as the input to the transformer-based decoder (120).

[0048] As previously explained, the transformer-based decoder (120) can be trained by taking into account the data that failed to decode from the pre-decoder (110) combined with the transformer-based decoder (120) in order to improve the performance for error correction of the transformer-based decoder (120).

[0049] If decoding is successful in the transformer-based decoder (120), the decoded codeword can be output to terminate the decoding process. On the other hand, if decoding fails, the decoding process can be performed through a post-decoder (130) combined with the transformer-based decoder (120). At this time, since the post-decoder (130) is a soft decision decoder and cannot receive binary output input, it receives probability information for each codeword decoded from the transformer-based decoder (120). At this time, it can receive probability information obtained by multiplying the output of the transformer-based decoder by the sign of data y.

[0050] The probability of error (LLR) can be calculated by multiplying the output of the transformer-based decoder (120) by the sign of the data. Through this calculation, the input of the soft decision decoder can be derived. Accordingly, if decoding fails in the transformer-based decoder (120), the probability obtained by multiplying the output of the transformer-based decoder (120) by the sign of the data can be input to the post-decoder (130) to perform decoding.

[0051]

[0052] FIG. 2 is a flowchart illustrating the operation method of a hybrid decoder in an embodiment.

[0053] In the following embodiments, each step may be performed sequentially, but is not necessarily performed sequentially. For example, the order of each step may be changed, and at least two steps may be performed in parallel. The hybrid decoder includes a form combining a pre-decoder, a transformer-based decoder, and a post-decoder, and may include at least one processor for controlling the following operations.

[0054] In step (210), the hybrid decoder decodes the received data through a pre-decoder and outputs a codeword of the first decoding.

[0055] The pre-decoder is a conventional decoder and may correspond to either a hard decision decoder or a soft decision decoder. The decoding process performed in the pre-decoder is referred to as the first decoding.

[0056] If the pre-decoder is a hard decision decoder, it can decode data received by the receiving device to perform error correction and estimate a data codeword of 0 or 1. Additionally, if the pre-decoder is a soft decision decoder, since it decodes soft information, it can input information indicating the probability that noisy data is 0 or 1 into the pre-decoder and perform error correction by decoding it. The soft decision decoder can estimate the data codeword based on probability information regarding the decoded codeword.

[0057] In step (220), the hybrid decoder can determine the result of the first decoding.

[0058] Based on the result of the first decoding, it can be determined whether the decoding was successful or failed.

[0059] For example, the success of the first decoding can be determined by determining whether the error included in the codeword estimated by the first decoding is less than or equal to a predetermined number or exceeds it.

[0060] If an error of less than or equal to a predetermined number is detected in the codeword estimated by the pre-decoder, the decoding is determined to be successful, and the decoding process can be terminated to output the codeword.

[0061] If an error exceeding a predetermined number is detected in the codeword estimated in the pre-decoder, in step (230), the hybrid decoder decodes the data or the first decoding codeword in the transformer-based decoder and outputs the codeword of the second decoding.

[0062] The decoding of a transformer-based decoder shall be referred to as second decoding.

[0063] If it is determined that the first decoding has failed, the decoding process can be continued in a transformer-based decoder combined with a pre-decoder. If the pre-decoder is a hard decision decoder, the transformer-based decoder can acquire data received from a receiving device and perform a second decoding if decoding fails in the pre-decoder. This does not request data retransmission, but rather inputs the same data that was input to the pre-decoder into the transformer-based decoder. Additionally, if the pre-decoder is a soft decision decoder, data may be input into the transformer-based decoder, or the output of the soft decision decoder may be input.

[0064] A transformer-based decoder may correspond to a decoder trained by considering the operation of a pre-decoder to improve the performance of the hybrid decoder's operation.

[0065] In step (240), the hybrid decoder can determine the result of the second decoding.

[0066] Based on the result of the second decoding, it can be determined whether the decoding was successful or failed. If the codeword estimated by the second decoding is determined to be successfully decoded, the decoding process can be terminated to output the corresponding codeword.

[0067] In step (250), the hybrid decoder decodes the codeword of the second decoding in the post-decoder based on the result of the second decoding.

[0068] If, as a result of the second decoding, it is determined that the decoding failed, the decoding process can be continued with a post-decoder connected to a transformer-based decoder.

[0069] If the post-decoder is a hard decision decoder, the post-decoder can obtain a codeword decoded from a transformer-based decoder and decode the corresponding codeword. Efficient decoding can be performed by inputting the codeword decoded by the transformer-based decoder into the post-decoder, considering that the codeword has at least partially corrected errors from the transformer-based decoder.

[0070] When the posterior decoder is a soft decision decoder, since the posterior decoder cannot accept binary output as input, it receives probability information for each codeword decoded from a transformer-based decoder. In this case, probability information obtained by multiplying the output of the transformer-based decoder by the sign of the input data can be received. The LLR can be calculated by multiplying the output of the transformer-based decoder by the sign of the data, and the derived LLR can be used as an input to the soft decision decoder. Refer to the explanation in FIG. 3 for the derivation of this.

[0071] In this way, an operation method for error correction in a hybrid decoder can be provided.

[0072]

[0073] FIG. 3 is a flowchart illustrating a method for learning a transformer-based decoder of a hybrid decoder in an embodiment.

[0074] In step (310), the learning device acquires a learning data set.

[0075] The training data set may include data having encoding and decoding relationships corresponding to the training purpose.

[0076] In step (320), the learning device learns a transformer-based decoder using a cross-entropy loss function for a failed dataset in which the decoding of codewords by the pre-decoder fails.

[0077] A transformer-based decoder can be trained for situations where the number of errors in the first decoding result exceeds a predetermined number.

[0078] The pre- and post-decoders can be trained based on different loss functions depending on whether they are hard decision decoders or soft decision decoders.

[0079] The loss function for training a transformer-based decoder corresponds to the cross-entropy loss function and can be expressed by the following mathematical formula.

[0080]

[0081] y represents the received data, x represents the correct codeword, and represents the output of a transformer-based decoder. represents binarized multiplicative noise. The finally estimated codeword is ...is. Here, sign(α) is the sign function, which takes a value of 1 if α≥ 0 and 0 otherwise, and bin(α) is the binarization function, defined as 0.5(1 - α). The bin function is a built-in function that converts an input value into a binary string.

[0082] When the pre- and post-decoders are hard decision decoders, a transformer-based decoder can be trained based on a final loss function such as Equation 2 below using these cross-entropy loss functions.

[0083]

[0084] u(x) is a step function, with u(x)=0 when x≤0 and u(x)=1 when x>0. t represents the error correction capability of the hard decision decoder. Since decoding is completed in the pre-decoder when an error less than or equal to t occurs, this condition is taken into account during training.

[0085] In the implementation, the conditional expectation is derived from Monte Carlo sampling. H It can be calculated by taking only samples (x,y) such that (x,y) > t. Since the derivative of the step function is zero in almost all regions, STE (Straight through estimator) is used to enable optimization using gradient descent.

[0086] The corresponding loss function indicates that the success of decoding in the post-decoder is determined based on the error of the codeword output from the transformer-based decoder.

[0087] Meanwhile, when the pre- and post-decoders are soft decision decoders, the transformer-based decoder can be trained using a different type of loss function than when the pre- and post-decoders are hard decision decoders.

[0088] Since transformer-based decoders use the cross-entropy loss function as the loss function, the value obtained by applying the sigmoid function to the output of the transformer-based decoder is the probability It can be interpreted as. `bin` is a built-in function that converts an input value into a binary string. Also, through this, the expression for LLR It can induce.

[0089] Accordingly, a transformer-based decoder combined with a soft decision decoder as a pre- and post-decoder can be trained based on the following loss function.

[0090]

[0091] Here, Pi is It represents the probability of error correction failure in the posterior decoder when there is error i in the output of the transformer-based decoder. Since the decoding success or failure of the soft decision decoder is not entirely determined by the number of errors, the loss is P rather than a step function consisting of 0s or 1s.i Composed of Multiply by . Accordingly, it can be approximated to be optimized using gradient descent.

[0092] = As such, it refers to a process that includes the decoding result of the pre-decoder and whether the decoding was successful. In implementation, It can be calculated in a way that ≠x, that is, the pre-decoder takes only samples where error correction failed.

[0093] The embodiment enables effective decoding of data that failed to be decoded in the pre-decoder combined with the transformer-based decoder by training the transformer-based decoder using the loss function as described above.

[0094]

[0095] FIG. 4 is a diagram illustrating the configuration of a receiving device for hybrid decoding in an embodiment.

[0096] Referring to FIG. 4, a receiving device (400) according to one embodiment may include a communication interface (410), a processor (430), and a memory (450). The communication interface (410), the processor (430), and the memory (450) may communicate with each other through a communication bus (405).

[0097] The communication interface (410) receives data. The data corresponds to the decoding target.

[0098] The processor (430) can perform three-stage decoding on data received through the communication interface (410) using the structure of a hybrid decoder. The processor (430) determines whether decoding is successful for each decoder constituting the hybrid decoder, and depending on the determination, it can terminate the decoding or allow decoding to be performed at the next decoder.

[0099] The memory (450) can store various information generated during the processing of the processor (430) described above. In addition, the memory (450) can store various data and programs. The memory (450) may include volatile memory or non-volatile memory. The memory (450) can store various data by providing a large-capacity storage medium such as a hard disk.

[0100] Additionally, the processor (430) may perform at least one method or an algorithm corresponding to at least one method described above through FIGS. 1 to 3. The processor (430) may be a data processing device implemented in hardware having a circuit having a physical structure for executing desired operations. For example, the desired operations may include code or instructions included in a program. The processor (430) may be composed of, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a NPU (Neural Network Processing Unit). For example, the prediction device (500) implemented in hardware may include a microprocessor, a central processing unit, a processor core, a multi-core processor, a multiprocessor, an ASIC (Application-Specific Integrated Circuit), or a FPGA (Field Programmable Gate Array).

[0101] The processor (430) can execute a program and control the receiving device (400). The program code executed by the processor (430) can be stored in memory (450).

[0102]

[0103] FIGS. 5a to 5d are drawings illustrating the performance of a hybrid decoder in BCH code decoding in an embodiment.

[0104] Experimental results of the hybrid decoder are illustrated. To evaluate the effectiveness of the hybrid decoder, the frame error rate (FER) performance was evaluated for four BCH codes ((31, 16), (63, 36), (63, 45), (63, 51)) in Figures 5a to 5d, respectively. These are codes with t=3, t=5, t=3, and t=2, respectively.

[0105] The graphs in Figures 5a to 5d show that the decoding results by the hybrid decoder are effective in error correction performance.

[0106] Compared to conventional transformer-based decoders, the hybrid decoder can significantly improve decoding performance in both scenarios, combining hard decision decoders and soft decision decoders. The results indicate that the proposed hybrid decoder outperforms other decoders across all SNR values.

[0107] In other words, by combining two different decoders and training with a loss function designed for a transformer-based decoder, an effective decoding structure can be provided.

[0108]

[0109] FIGS. 6a and 6b are drawings illustrating an example of a hybrid decoder composed of two stages in an embodiment.

[0110] A hybrid decoder can be proposed in which a conventional decoder is connected as a pre-decoder of a transformer-based decoder as in Fig. 6a, or a conventional decoder is connected as a post-decoder of a transformer-based decoder as in Fig. 6b.

[0111] The hybrid decoder of FIG. 6a performs decoding of the received data in a conventional decoder, and determines whether to decode in a transformer-based decoder based on the success or failure of the decoding. If decoding is successful in the conventional decoder, the decoding of the hybrid decoder is terminated and the decoded codeword can be output. If it is determined that decoding failed in the conventional decoder, the data can be decoded in a transformer-based decoder. The result decoded in the transformer-based decoder can be output as the final decoded codeword.

[0112] The hybrid decoder of FIG. 6b first processes the received data in a transformer-based decoder and can be decoded in the transformer-based decoder based on whether the decoding is successful. If the decoding in the transformer-based decoder is successful, the decoding process is terminated and the codeword decoded in the transformer-based decoder can be output. Conversely, if it is determined that the decoding in the transformer-based decoder has failed, the decoding result of the transformer-based decoder can be input to a conventional decoder to perform decoding.

[0113] If the conventional decoder is a hard decision decoder, it can receive the codewords produced by the decoding of a transformer-based decoder directly. On the other hand, if it is a soft decision decoder, the codewords cannot be received directly. Therefore, the form of the LLR that can be input to the soft decision decoder can be calculated by multiplying the codeword estimate from the transformer-based decoder's decoding—for example, the estimate of multiplication noise—by the sign of the data. By multiplying the codeword estimate by the data sign, the LLR can be calculated, and the conventional decoder can receive the LLR and perform decoding.

[0114]

[0115] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0116]

[0117] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0118]

[0119] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based on the above. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0120] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

1. In a hybrid decoder, Transformer-based decoder Includes, It further includes at least one of a pre-decoder or a post-decoder, and The above-mentioned pre-decoder decodes the received data and outputs a codeword of the first decoding, the above-mentioned transformer-based decoder decodes the data or the codeword of the first decoding and outputs a codeword of the second decoding, and the above-mentioned post-decoder decodes the codeword of the second decoding based on the result of the second decoding. Hybrid decoder.

2. In Paragraph 1, The above-mentioned pre-decoder and post-decoder are, Corresponding to one of a Soft Decision decoder and a Hard Decision decoder, Hybrid decoder.

3. In Paragraph 1, If the number of errors included in the codeword of the first decoding is less than or equal to a predetermined number, the first decoding result of the data is output as the final decoding result, and If the number of the above errors exceeds the above predetermined number, the data is input to the transformer-based decoder. Hybrid decoder.

4. In Paragraph 1, If the above post-decoder is a hard decision decoder, In response to the failure of the second decoding above, the codeword of the second decoding above is input to the post-decoder, Hybrid decoder.

5. In Paragraph 1, If the above post-decoder is a soft decision decoder, In response to the failure of the second decoding, the error probability for the codeword of the second decoding is input to the posterior decoder. Hybrid decoder.

6. In Paragraph 5, Calculate the Log Likelihood Ratio (LR), which is the probability of the error, by multiplying the estimated value for the codeword of the second decoding above by the sign of the data, and Inputting the above calculated LLR into the above post-decoder, Hybrid decoder.

7. In Paragraph 1, The above transformer-based decoder is, Among the training data, trained using a cross-entropy-based loss function for a failure dataset where codeword decoding by a pre-decoder fails, Operation method of a hybrid decoder.

8. In the method of operation of a hybrid decoder, A step of decoding the received data in a pre-decoder and outputting the codeword of the first decoding; Based on the result of the first decoding, a step of decoding the data or the first decoding codeword in a transformer-based decoder to output a codeword of the second decoding; and Based on the result of the second decoding, the step of decoding the codeword of the second decoding in the post-decoder including, Operation method of a hybrid decoder.

9. In Paragraph 8, The above-mentioned pre-decoder and post-decoder are, Corresponding to one of a Soft Decision decoder and a Hard Decision decoder, Operation method of a hybrid decoder.

10. In Paragraph 8, Based on the result of the first decoding, the step of decoding the data or the first decoding codeword in a transformer-based decoder and outputting the codeword of the second decoding is: A step of outputting the first decoding result of the data as the final decoding result when the number of errors included in the codeword of the first decoding is less than or equal to a predetermined number; and If the number of errors exceeds the predetermined number, the step of inputting the data into the transformer-based decoder including, Operation method of a hybrid decoder.

11. In Paragraph 8, Based on the result of the second decoding above, the step of decoding the codeword of the second decoding in the post-decoder is, If the above post-decoder is a hard decision decoder, the step of inputting the codeword of the second decoding into the post-decoder in response to the failure of the above second decoding. including, Operation method of a hybrid decoder.

12. In Paragraph 8, Based on the result of the second decoding above, the step of decoding the codeword of the second decoding in the post-decoder is, If the above posterior decoder is a soft decision decoder, the step of inputting the error probability for the codeword of the second decoding to the posterior decoder in response to the failure of the second decoding. including, Operation method of a hybrid decoder.

13. In Paragraph 12, The step of inputting the error probability for the codeword of the second decoding to the posterior decoder is: Calculating LLR by multiplying the estimated value for the codeword of the second decoding by the sign of the data, and inputting the calculated LLR to the post-decoder. including, Operation method of a hybrid decoder.

14. In Paragraph 8, The above transformer-based decoder is, Among the training data, trained using a cross-entropy-based loss function for a failure dataset where codeword decoding by a pre-decoder fails, Operation method of a hybrid decoder.

15. In the training method of a transformer-based decoder, Step of acquiring a training data set; and A step of training the transformer-based decoder using a cross-entropy-based loss function for a failure dataset among the above training data, in which the decoding of codewords by a pre-decoder combined with the transformer-based decoder fails. Includes, The above transformer-based decoder is, A hybrid decoder configured in a combined structure of the above-mentioned pre-decoder, the above-mentioned transformer decoder, and the above-mentioned post-decoder, Learning method for a transformer-based decoder.

16. In Paragraph 15, If the above post-decoder is a hard decision decoder, The above cross-entropy-based loss function includes the product of a step function and cross-entropy, and includes the assumption that data in which the error of a codeword by the transformer-based decoder exceeds a predetermined number is determined to be a failure in decoding of the codeword. Learning method for a transformer-based decoder.

17. In Paragraph 15, If the above post-decoder is a soft decision decoder, The above cross-entropy-based loss function includes a conditional probability that determines that the decoding of the codeword has failed based on the error probability of the codeword by the posterior decoder. Learning method for a transformer-based decoder.

18. In Paragraph 15, The above hybrid decoder is, A pre-decoder connected to the input terminal of the above-mentioned transformer-based decoder, which decodes received data and outputs a codeword of the first decoding; A transformer-based decoder that decodes the data or the first decoding codeword based on the result of the first decoding and outputs a codeword of the second decoding; and A post-decoder connected to the output terminal of the transformer-based decoder and decoding the codeword of the second decoding based on the result of the second decoding including Learning method for a transformer-based decoder.

19. In Paragraph 15, The above-mentioned pre-decoder and post-decoder are, Corresponding to one of a Soft Decision decoder and a Hard Decision decoder, Learning method for a transformer-based decoder.

20. A computer-readable recording medium storing a program that, when executed by a computer, enables the method according to paragraph 8 to be performed in combination with hardware.

Images