Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2305.04347 · IEEE International Symposium on Information Theory (ISIT)

Interpreting Training Aspects of Deep-Learned Error-Correcting Codes

Natasha Devroye, Abhijeet Mulgund, Raj Shekhar, György Turán, Miloš Žefran, Yingyao Zhou

Published Jun 25, 2023• 4 citations
arXivHTMLPDF

Abstract

As new deep-learned error-correcting codes continue to be introduced, it is important to develop tools to interpret the designed codes and understand the training process. Prior work focusing on the deep-learned TurboAE has both interpreted the learned encoders post-hoc by mapping these onto nearby "interpretable" encoders, and experimentally evaluated the performance of these interpretable encoders with various decoders. Here we look at developing tools for interpreting the training process for deep-learned error-correcting codes, focusing on: 1) using the Goldreich-Levin algorithm to quickly interpret the learned encoder; 2) using Fourier coefficients as a tool for understanding the training dynamics and the loss landscape; 3) reformulating the training loss, the binary cross entropy, by relating it to encoder and decoder parameters, and the bit error rate (BER); 4) using these insights to formulate and study a new training procedure. All tools are demonstrated on TurboAE, but are applicable to other deep-learned forward error correcting codes (without feedback).

DOI
10.1109/isit54713.2023.10206783

Related papers

ProductAE: Toward Deep Learning Driven Error-Correction Codes of Large Dimensions

Mohammad Vahid Jamali, Hamid Saber, Homayoon Hatami +1 more

While decades of theoretical research have led to the invention of several classes of error-correction codes, the design of such codes is an extremely challenging task, mostly driven by human ingenuity. Recent studies demonstrate that such designs can be effectively automated and accelerated via tools from machine learning (ML), thus enabling ML-driven classes of error-correction codes with promising performance gains compared to classical designs. A fundamental challenge, however, is that it is prohibitively complex, if not impossible, to design and train fully ML-driven encoder and decoder pairs for large code dimensions. In this paper, we propose Product Autoencoder (ProductAE) -- a computationally-efficient family of deep learning driven (encoder, decoder) pairs -- aimed at enabling the training of relatively large codes (both encoder and decoder) with a manageable training complexity. We build upon ideas from classical product codes and propose constructing large neural codes using smaller code components. ProductAE boils down the complex problem of training the encoder and decoder for a large code dimension $k$ and blocklength $n$ to less-complex sub-problems of training encoders and decoders for smaller dimensions and blocklengths. Our training results show successful training of ProductAEs of dimensions as large as $k = 300$ bits with meaningful performance gains compared to state-of-the-art classical and neural designs. Moreover, we demonstrate excellent robustness and adaptivity of ProductAEs to channel models different than the ones used for training.

arXiv:2303.16424•Mar 29, 2023•5 citations

Higher-order Interpretations of Deepcode, a Learned Feedback Code

Yaoyu Zhou, Natasha Devroye, Gy. Turán +1 more

We present an interpretation of Deepcode, a learned feedback code that showcases higher-order error correction relative to an earlier interpretable model [1]. By interpretation, we mean succinct analytical encoder and decoder expressions (albeit with learned parameters) in which the role of feedback in achieving error correction is easy to understand. By higher-order, we mean that longer sequences of large noise values are acted upon by the encoder (which has access to these through the feedback) and used in error correction at the decoder in a two-stage decoding process.11This work was supported by NSF under awards 1900911, 2217023, and 2240532, and by the AI National Laboratory Program (RRF-2.3.1-21-2022-00004). The contents of this article are solely the responsibility of the authors and do not necessarily represent the official views of the NSF.

arXiv:2408.11907•Allerton Conference on Communication, Control, and Computing•Sep 24, 2024•3 citations

Deep code: representation and learning algorithms for neural networks & their applications to communication codes

Ashok Vardhan Makkuva

Codes are the backbone of modern information age.Codes, composed of encoder and decoder pairs, are the basic mathematical objects that enable reliable communication.Landmark codes include convolutional, Reed-Muller, turbo, LDPC, and polar: each is linear and represents a mathematical breakthrough.Their impact on humanity is huge; each of these codes has been used in global communication standards over the past six decades.On the other hand, designing codes is a challenging task, mostly driven by human ingenuity.Befittingly, historically, the progress in discovery of codes has been sporadic.In this thesis, we present a new paradigm to invent codes via harnessing tools from deep-learning.Our major result is the invention of KO codes, a computationally efficient family of deep-learning driven codes that outperform the state-of-the-art RM and polar codes, in the challenging short-to-medium block length regime.The key technical innovation behind KO codes is the design of a novel family of neural architectures inspired by the computation tree of the Kronecker Operation (KO) central to RM and polar codes.These architectures pave the way for discovery of a much richer class of hitherto unexplored non-linear codes.This design technique can be viewed as an instantiation of the classical neural augmentation principle.In the process, we also study a popular neural network model called Mixture-of-Experts (MoE) that realizes this principle.We provide the first set of consistent and efficient algorithms with global learning guarantees for learning the parameters in a MoE which has been an open problem for more than two decades.B.1 Comparison of SGD on our losses (L 4 , L log ) vs. ℓ 2 and the EM algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . .106 C.2 KO code achieves a significant gain over the Polar(64, 7) code in BLER when trained on AWGN channel.KO decoder also matches the optimal MAP decoder. . . . . . . . . . . . .108 xii C.6 KO(6, 1) code.Left: KO(6, 1) code achieves significant gain over RM(6, 1) code (with Dumer) when trained on AWGN channel.Right: Under the optimal MAP decoding, KO(6, 1) and RM(6, 1) codes achieve the same performance.Error rates for a random Gaussian codebook are also plotted as a baseline. . . . . . . . . . . . . . . . . . . . . . . . . . . .120 C.1 Plotkin trees for the Polar(64, 7) encoder and our neural KO encoder counterpart.Both codes have dimension k = 7 and blocklength n = 64. . . . . . . . . . . . . . . . . . . . . .121 C.3 The same KO(8,2) encoder and decoder as those used in Figure 4.4 achieve a significant gain (without any retraining or fine-tuning) when list decoding is used together with the KO decoder.The magnitude of the gain is comparable to the gain achieved by the same list decoding technique on the successive cancellation decoder of the RM(8,2) code.We used the list decoding from [1] but without the permutation technique. . . . . . . . . . . . . . . . . . . . . . . . . . . . .122 C.4 Plotkin trees for RM(8,2) and KO(8,2) encoders.Leaves are highlighted in green.Both codes have dimension k = 37 and blocklength n = 256. . . . . . . . . . . . . . . . . . . . . .123 C.5 Plotkin trees for the RM(8, 2) and KO(8, 2) decoders.Red arrows indicate the bit decoding order. . . . . . . .

•Jul 15, 2022•0 citations

On deep learning-based channel decoding

Tobias Gruber, Sebastian Cammerer, Jakob Hoydis +1 more

We revisit the idea of using deep neural networks for one-shot decoding of random and structured codes, such as polar codes. Although it is possible to achieve maximum a posteriori (MAP) bit error rate (BER) performance for both code families and for short codeword lengths, we observe that (i) structured codes are easier to learn and (ii) the neural network is able to generalize to codewords that it has never seen during training for structured, but not for random codes. These results provide some evidence that neural networks can learn a form of decoding algorithm, rather than only a simple classifier. We introduce the metric normalized validation error (NVE) in order to further investigate the potential and limitations of deep learning-based decoding with respect to performance and complexity.

Conference on Information Sciences and Systems•Mar 1, 2017•594 citations

Machine Learning Optimal Quantum Error Correction Thresholds

Dominik Seip, Luis Colmenarez, Markus Schmitt +1 more

As quantum computers remain susceptible to noise, QEC is essential for preserving logical information during computations. However, the performance of QEC codes breaks down beyond certain noise thresholds, revealing fundamental limits on their ability to protect quantum information. These limits can be characterized using information-theoretic measures such as the coherent information, which quantifies the maximum rate at which logical information can be reliably transmitted through a noisy quantum channel. In this work, we establish a direct connection between the CI and the binary cross-entropy loss used when training neural network decoders. Specifically, we show that the CI constitutes a sharp lower bound on the achievable loss for decoders that track logical operators across noisy channels. To this end, we develop a transformer-based neural network model based on maximum likelihood decoding. We train this network to estimate the CI and evaluate its performance on the surface code under three noise models: code capacity, phenomenological, and circuit-level noise. Our results demonstrate that the network accurately predicts CI and yields threshold estimates that closely match known theoretical limits. When used as a decoder, the network significantly outperforms the minimum weight perfect matching decoder in terms of logical error rate. We also introduce a novel soft post-selection scheme that independently treats uncertainty in both logical operators and relies on confidence-based filtering of the network's output. We prove that such post-selection strategies, based on the MLD cosets, are optimal, and demonstrate their scalability in terms of both logical error rate and abort probability. These findings establish transformer-based architectures as powerful tools for QEC and provide the first numerical evidence supporting the optimality and scalability of MLD-based post-selection.

quant-pharXiv:2606.22194•Jun 23, 2026•3 citations

On Deep Learning-Based Channel Decoding

Tobias Gruber, Sebastian Cammerer, Jakob Hoydis +1 more

We revisit the idea of using deep neural networks for one-shot decoding of random and structured codes, such as polar codes. Although it is possible to achieve maximum a posteriori (MAP) bit error rate (BER) performance for both code families and for short codeword lengths, we observe that (i) structured codes are easier to learn and (ii) the neural network is able to generalize to codewords that it has never seen during training for structured, but not for random codes. These results provide some evidence that neural networks can learn a form of decoding algorithm, rather than only a simple classifier. We introduce the metric normalized validation error (NVE) in order to further investigate the potential and limitations of deep learning-based decoding with respect to performance and complexity.

arXiv:1701.07738•Jan 26, 2017•60 citations
As new deep-learned error-correcting codes continue to be introduced, it is important to develop tools to interpret the designed codes and understand the training process. Prior work focusing on the deep-learned TurboAE has both interpreted the learned encoders post-hoc by mapping these onto nearby "interpretable" encoders, and experimentally evaluated the performance of these interpretable encoders with various decoders. Here we look at developing tools for interpreting the training process for deep-learned error-correcting codes, focusing on: 1) using the Goldreich-Levin algorithm to quickly interpret the learned encoder; 2) using Fourier coefficients as a tool for understanding the training dynamics and the loss landscape; 3) reformulating the training loss, the binary cross entropy, by relating it to encoder and decoder parameters, and the bit error rate (BER); 4) using these insights to formulate and study a new training procedure. All tools are demonstrated on TurboAE, but are applicable to other deep-learned forward error correcting codes (without feedback).
While decades of theoretical research have led to the invention of several classes of error-correction codes, the design of such codes is an extremely challenging task, mostly driven by human ingenuity. Recent studies demonstrate that such designs can be effectively automated and accelerated via tools from machine learning (ML), thus enabling ML-driven classes of error-correction codes with promising performance gains compared to classical designs. A fundamental challenge, however, is that it is prohibitively complex, if not impossible, to design and train fully ML-driven encoder and decoder pairs for large code dimensions. In this paper, we propose Product Autoencoder (ProductAE) -- a computationally-efficient family of deep learning driven (encoder, decoder) pairs -- aimed at enabling the training of relatively large codes (both encoder and decoder) with a manageable training complexity. We build upon ideas from classical product codes and propose constructing large neural codes using smaller code components. ProductAE boils down the complex problem of training the encoder and decoder for a large code dimension kkk and blocklength nnn to less-complex sub-problems of training encoders and decoders for smaller dimensions and blocklengths. Our training results show successful training of ProductAEs of dimensions as large as k=300k = 300k=300 bits with meaningful performance gains compared to state-of-the-art classical and neural designs. Moreover, we demonstrate excellent robustness and adaptivity of ProductAEs to channel models different than the ones used for training.
We present an interpretation of Deepcode, a learned feedback code that showcases higher-order error correction relative to an earlier interpretable model [1]. By interpretation, we mean succinct analytical encoder and decoder expressions (albeit with learned parameters) in which the role of feedback in achieving error correction is easy to understand. By higher-order, we mean that longer sequences of large noise values are acted upon by the encoder (which has access to these through the feedback) and used in error correction at the decoder in a two-stage decoding process.11This work was supported by NSF under awards 1900911, 2217023, and 2240532, and by the AI National Laboratory Program (RRF-2.3.1-21-2022-00004). The contents of this article are solely the responsibility of the authors and do not necessarily represent the official views of the NSF.
Codes are the backbone of modern information age.Codes, composed of encoder and decoder pairs, are the basic mathematical objects that enable reliable communication.Landmark codes include convolutional, Reed-Muller, turbo, LDPC, and polar: each is linear and represents a mathematical breakthrough.Their impact on humanity is huge; each of these codes has been used in global communication standards over the past six decades.On the other hand, designing codes is a challenging task, mostly driven by human ingenuity.Befittingly, historically, the progress in discovery of codes has been sporadic.In this thesis, we present a new paradigm to invent codes via harnessing tools from deep-learning.Our major result is the invention of KO codes, a computationally efficient family of deep-learning driven codes that outperform the state-of-the-art RM and polar codes, in the challenging short-to-medium block length regime.The key technical innovation behind KO codes is the design of a novel family of neural architectures inspired by the computation tree of the Kronecker Operation (KO) central to RM and polar codes.These architectures pave the way for discovery of a much richer class of hitherto unexplored non-linear codes.This design technique can be viewed as an instantiation of the classical neural augmentation principle.In the process, we also study a popular neural network model called Mixture-of-Experts (MoE) that realizes this principle.We provide the first set of consistent and efficient algorithms with global learning guarantees for learning the parameters in a MoE which has been an open problem for more than two decades.B.1 Comparison of SGD on our losses (L 4 , L log ) vs. ℓ 2 and the EM algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . .106 C.2 KO code achieves a significant gain over the Polar(64, 7) code in BLER when trained on AWGN channel.KO decoder also matches the optimal MAP decoder. . . . . . . . . . . . .108 xii C.6 KO(6, 1) code.Left: KO(6, 1) code achieves significant gain over RM(6, 1) code (with Dumer) when trained on AWGN channel.Right: Under the optimal MAP decoding, KO(6, 1) and RM(6, 1) codes achieve the same performance.Error rates for a random Gaussian codebook are also plotted as a baseline. . . . . . . . . . . . . . . . . . . . . . . . . . . .120 C.1 Plotkin trees for the Polar(64, 7) encoder and our neural KO encoder counterpart.Both codes have dimension k = 7 and blocklength n = 64. . . . . . . . . . . . . . . . . . . . . .121 C.3 The same KO(8,2) encoder and decoder as those used in Figure 4.4 achieve a significant gain (without any retraining or fine-tuning) when list decoding is used together with the KO decoder.The magnitude of the gain is comparable to the gain achieved by the same list decoding technique on the successive cancellation decoder of the RM(8,2) code.We used the list decoding from [1] but without the permutation technique. . . . . . . . . . . . . . . . . . . . . . . . . . . . .122 C.4 Plotkin trees for RM(8,2) and KO(8,2) encoders.Leaves are highlighted in green.Both codes have dimension k = 37 and blocklength n = 256. . . . . . . . . . . . . . . . . . . . . .123 C.5 Plotkin trees for the RM(8, 2) and KO(8, 2) decoders.Red arrows indicate the bit decoding order. . . . . . . .
We revisit the idea of using deep neural networks for one-shot decoding of random and structured codes, such as polar codes. Although it is possible to achieve maximum a posteriori (MAP) bit error rate (BER) performance for both code families and for short codeword lengths, we observe that (i) structured codes are easier to learn and (ii) the neural network is able to generalize to codewords that it has never seen during training for structured, but not for random codes. These results provide some evidence that neural networks can learn a form of decoding algorithm, rather than only a simple classifier. We introduce the metric normalized validation error (NVE) in order to further investigate the potential and limitations of deep learning-based decoding with respect to performance and complexity.
As quantum computers remain susceptible to noise, QEC is essential for preserving logical information during computations. However, the performance of QEC codes breaks down beyond certain noise thresholds, revealing fundamental limits on their ability to protect quantum information. These limits can be characterized using information-theoretic measures such as the coherent information, which quantifies the maximum rate at which logical information can be reliably transmitted through a noisy quantum channel. In this work, we establish a direct connection between the CI and the binary cross-entropy loss used when training neural network decoders. Specifically, we show that the CI constitutes a sharp lower bound on the achievable loss for decoders that track logical operators across noisy channels. To this end, we develop a transformer-based neural network model based on maximum likelihood decoding. We train this network to estimate the CI and evaluate its performance on the surface code under three noise models: code capacity, phenomenological, and circuit-level noise. Our results demonstrate that the network accurately predicts CI and yields threshold estimates that closely match known theoretical limits. When used as a decoder, the network significantly outperforms the minimum weight perfect matching decoder in terms of logical error rate. We also introduce a novel soft post-selection scheme that independently treats uncertainty in both logical operators and relies on confidence-based filtering of the network's output. We prove that such post-selection strategies, based on the MLD cosets, are optimal, and demonstrate their scalability in terms of both logical error rate and abort probability. These findings establish transformer-based architectures as powerful tools for QEC and provide the first numerical evidence supporting the optimality and scalability of MLD-based post-selection.
We revisit the idea of using deep neural networks for one-shot decoding of random and structured codes, such as polar codes. Although it is possible to achieve maximum a posteriori (MAP) bit error rate (BER) performance for both code families and for short codeword lengths, we observe that (i) structured codes are easier to learn and (ii) the neural network is able to generalize to codewords that it has never seen during training for structured, but not for random codes. These results provide some evidence that neural networks can learn a form of decoding algorithm, rather than only a simple classifier. We introduce the metric normalized validation error (NVE) in order to further investigate the potential and limitations of deep learning-based decoding with respect to performance and complexity.