Performance and robustness of real-world Acoustic Event Classification (AEC) solutions depend on ability to train on diverse data from wide range of end-point devices and acoustic environments. Federated Learning (FL) provides a framework to leverage annotated and non-annotated AEC data from servers and client devices in a privacy preserving manner. In this work we propose a novel Federated Relaxed Pareto Optimization (FedRPO) method for semi-supervised FL with heterogeneous client data. In contrast to federated averaging class of FL algorithms (fedAvg) that perform unconstrained weighted aggregation across all data sources, FedRPO enables special treatment of data with high quality annotations vs. data with pseudo-labels of unknown, varying qualities. In particular, FedRPO computes the updates to the global model solving a constrained linear program, with explicit Pareto constraints to prevent performance degradation on annotated data, and controlled relaxation of the Pareto constraints on pseudo-labeled data to prevent learning of patterns in conflict with the annotated data. We show FedRPO significantly outperforms FedAvg on Amazon internal de-identified dataset on AEC tasks. On supervised learning, FedRPO improved precision by 32.5% over FedAvg when maintaining recall at 90%. Combined with FixMatch [1] for semi-supervised learning, FedRPO outperformed FedAvg on precision by 50.5% at 90% recall.

Noise recycling is a recently proposed method that significantly enhances decoding performance when used for orthogonal channels impacted by correlated noise with only receiver side changes. In this paper, we establish that noise recycling can be applied in a single communication channel that is subject to temporally correlated noise by leveraging a standard matrix interleaver to create the effect of orthogonal channels. The proposed interleaved noise recycling technique works with any code, requires no sender-side alterations, and only minor changes to the receiver architecture. In a hard-detection scenario, we demonstrate noise recycling can enable an accurate estimate of continuous realization of noise without using any soft information, resulting in a gain of more than 2 dB in Block Error Rate (BLER). We use the first hardware implementation of Guessing Random Additive Noise Decoding (GRAND), a universal noise-centric decoder, to illustrate the advantages of noise recycling in hardware performance. At a correlation coefficient of 0.75, Eb/N0of 4 dB, a maximum of 36× decoding energy savings with a 12× reduction in latency is achieved using a BCH(127,113) code when GRAND is equipped with the proposed noise recycling.

Closing the gap between high data rates and low delay in real-time streaming applications is a major challenge in advanced communication systems. While adaptive network coding schemes have the potential of balancing the rate and the delay in real-time, they often rely on a prediction of the channel behavior. In practice, such a prediction is based on delayed feedbacks, making it difficult to acquire causality, particularly when the channel model is unknown. In this work, we propose a deep learning-based noise prediction (DeepNP) algorithm, which augments the recently proposed adaptive and causal random linear network coding scheme with a neural network that learns to carry out noise prediction from data. This neural augmentation is utilized to maximize the throughput while minimizing in-order delivery delay of the coding scheme, and operate in a channel-model-agnostic manner. We numerically show that performance can dramatically increase by the learned prediction of the channel noise rate, demonstrating that DeepNP gains up to a factor of four in mean and maximum delay and a factor of two in throughput compared with statistic-based network coding approaches.

We introduce Iterative GRAND (IGRAND), a universal product code decoder that applies iterative bounded distance decoding and decodes component codes using code-agnostic Guessing Random Additive Noise Decoding (GRAND). We empirically determine its accuracy and, based on GRAND hardware measurements, its complexity, showing gains over alternative algorithms. We prove that the class of product codes with random linear component codes, which IGRAND is capable of decoding, are capacity-achieving in hard-decision channels.

One of the main challenges in containing the Coronoavirus disease 2019 (COVID-19) pandemic stems from the difficulty in carrying out efficient mass diagnosis over large populations. The leading method to test for COVID-19 infection utilizes qualitative polymerase chain reaction, implemented using dedicated machinery which can simultaneously process a limited amount of samples. A candidate method to increase the test throughput is to examine pooled samples comprised of a mixture of samples from different patients. In this work we study pooling-based COVID-19 tests. We identify the specific requirements of COVID-19 testing, including the need to characterize the infection level and to operate in a one-shot fashion, which limit the application of traditional group-testing (GT) methods. We then propose a multi-level GT scheme, designed specifically to meet the unique requirements of COVID-19 tests, while exploiting the strength of GT theory to enable accurate recovery using much fewer tests than patients. Our numerical results demonstrate that multi-level GT reliably and efficiently detects the infection levels, while achieving improved accuracy over previously proposed one-shot COVID-19 pooled-testing methods.

Two main problems arise in the Multiple Access Channel (MAC): interference from different users, and additive noise channel noise. Maximum A-Posteriori (MAP) joint decoding or successive interference cancellation are known to be capacity-achieving for the MAC when paired with appropriate codes. We extend the recently proposed Soft Guessing Random Additive Noise Decoder (SGRAND) to guess, using soft information, the effect of noise on the sum of users' transmitted codewords. Next, we manage interference by applying ZigZag decoding over the resulting putative noiseless MAC to obtain candidate codewords. Guessing continues until the candidate codewords thus obtained pertain to the corresponding users' codebooks. This MAC SGRAND decoder is a MAP decoder that requires no coordination between users, who can use arbitrary moderate redundancy short length codes of different types and rates.

In this paper we introduce Neural Network Coding (NNC), a data-driven approach to joint source and network coding. In NNC, the encoders at each source and intermediate node, as well as the decoder at each destination node, are neural networks which are all trained jointly for the task of communicating correlated sources through a network of noisy point-to-point links. The NNC scheme is application-specific and makes use of a training set of data, instead of making assumptions on the source statistics. In addition, it can adapt to any arbitrary network topology and power constraint. We show empirically that, for the task of transmitting MNIST images over a network, the NNC scheme shows improvement over baseline schemes, especially in the low-SNR regime.

Maximum Likelihood (ML) decoding of forward error correction codes is known to be optimally accurate, but is not used in practice as it proves too challenging to efficiently implement. Here we propose a development of a previously described hard detection ML decoder called Guessing Random Additive Noise Decoding (GRAND). We introduce Soft GRAND (SGRAND), a ML decoder that fully avails of soft detection information and is suitable for use with any arbitrary high-rate, short-length block code. We assess SGRAND's performance on Cyclic Redundancy Check (CRC)-aided Polar (CA-Polar) codes, which will be used for all control channel communication in 5G New Radio (NR), comparing its accuracy with CRC-Aided Successive Cancellation List decoding (CA-SCL), a state-of-the-art soft-information decoder specific to CA-Polar codes.

We introduce Noise Recycling, a method that enhances decoding performance of channels subject to correlated noise without joint decoding. The method can be used with any combination of codes, code-rates and decoding techniques. In the approach, a continuous realization of noise is estimated from a lead channel by subtracting its decoded output from its received signal. This estimate is then used to improve the accuracy of decoding of an orthogonal channel that is experiencing correlated noise. In this design, channels aid each other only through the provision of noise estimates post-decoding. In a Gauss-Markov model of correlated noise, we constructively establish that noise recycling employing a simple successive order enables higher rates than not recycling noise. Simulations illustrate noise recycling can be employed with any code and decoder, and that noise recycling shows Block Error Rate (BLER) benefits when applying the same predetermined order as used to enhance the rate region. Finally, for short codes we establish that an additional BLER improvement is possible through noise recycling with racing, where the lead channel is not pre-determined, but is chosen on the fly based on which decoder completes first.

We construct error-correcting write-once memory (WOM) codes that guarantee correction of any specified number of errors in q-level memories. The constructions use suitably designed short q-ary WOM codes and concatenate them with outer error-correcting codes over different alphabets using suitably designed mappings. With a new storage-efficiency measure, we call EC-rate and show that for common error types the codes save redundancy and implementation complexity over straightforward concatenation. In addition to constructions for guaranteed error correction, we extend the error-correcting WOM scheme to binary multi-level coding for random errors, and toward soft-decision decoding provide an efficient way to extract reliability information without using higher-precision readout.

We construct error-correcting WOM (write-once memory) codes that can correct any specified number of errors in q-Ievel memories. The constructions use suitably designed short q-ary WOM codes and concatenate them with outer error-correcting codes over different alphabets, using suitably designed mappings. With a new storage-efficiency measure we call EC-rate, we show that for common error types the codes save redundancy and implementation complexity over straightforward concatenation.