The technology relates to bandwidth constrained communication systems with neural network based detection. In some embodiments, a bandwidth constrained equalized transport (BCET) communication system comprises: a transmitter comprising an error control code encoder, a pulse-shaping filter, and a first interleaver; a communication channel; and a receiver comprising a neural network processing block that processes a received signal. The error control code encoder can append redundant information onto the signal. The pulse-shaping filter can intentionally introduce memory into the signal in the form of inter-symbol interference. The first interleaver can change a temporal order of the symbols in the signal. The error control code encoder can be a low-density parity-check (LDPC) error control code encoder. The neural network can be trained with positive mappings between transmitted and decoded training signals, or negative mappings between training signals and a null space of an LDPC generation matrix.

Neural Normalized MinSum (N-NMS) decoding for improving frame error rate (FER) performance on linear block codes over conventional normalized MinSum (NMS). Dynamic multiplicative weights are assigned to each check-to-variable message in each iteration to efficiently provide training parameters of N-NMS that support N-NMS for longer block lengths. Embodiment are described for neural two-dimensional normalized MinSum (N-2D-NMS) decoders requiring fewer training parameters. The N-2D-NMS approaches for example use the same weight for edges with the same check and/or variable node degree. Simulation results indicate that that this LDPC decoding performs similarly to previous techniques while substantially reducing the amount of training necessary.

Message faulting is a critical unsolved problem for 5G and 6G. Disclosed herein is a method for combining an AI-based analysis of the waveform data of each message element, plus the constraint of an associated error-detection code (such as a CRC or parity construct of the correct message) to localize and, in many cases, correct a limited number of faults per message, without a retransmission. For example, the waveform data may include a deviation of the amplitude or phase of a particular message element, relative to an average of the amplitudes or phases of the other message elements that have the same demodulation value. The outliers are thereby exposed as the most likely faulted message elements. In addition, using the error-detection code, the AI model can determine the most likely corrected message, thereby avoiding retransmission delays and power usage and other costs.

A memory system, a bit-flipping (BF) low-density parity check (LDPC) decoder that may be included in the memory system and operating methods thereof in which such decoder or decoding has a reduced error floor. Such a BF LDPC decoder is configured using a deep learning framework of trained and training neural networks and data separation that exploits the degree distribution information of the constructed LDPC codes.

A receiver may use a trained AI model to recover a faulted 5G/6G message by interpreting the meaning or intent of the message by correlating the message content with one of the expected message types. For example, the AI model may consider changes to the message, for consistency with an associated error-detection code, thereby producing a series of candidate messages. The AI model can then determine a likelihood that each of the candidate messages is correct, in the context of the receiver (such as an action or condition of the receiver, or a planned activity of the receiver) or is commonly received in that context. For example, the AI model can be trained to recognize the expected messages or message types, and thereby indicate which candidate message has the highest likelihood of being correct. The AI model may also consider waveform parameters to identify likely faults.

An analog error correction circuit is disclosed that implements an analog error correction code. The analog circuit includes a crossbar array of memristors or other non-volatile tunable resistive memory devices. The crossbar array includes a first crossbar array portion programmed with values of a target computation matrix and a second crossbar array portion programmed with values of an encoder matrix for correcting computation errors in the matrix multiplication of an input vector with the computation matrix. The first and second crossbar array portions share the same row lines and are connected to a third crossbar array portion that is programmed with values of a decoder matrix, thereby enabling single-cycle error detection. A computation error is detected based on output of the decoder matrix circuitry and a location of the error is determined via an inverse matrix multiplication operation whereby the decoder matrix output is fed back to the decoder matrix.

The technology relates to bandwidth constrained communication systems with neural network based detection. In some embodiments, a bandwidth constrained equalized transport (BCET) communication system comprises: a transmitter comprising an error control code encoder, a pulse-shaping filter, and a first interleaver; a communication channel; and a receiver comprising a neural network processing block that processes a received signal. The error control code encoder can append redundant information onto the signal. The pulse-shaping filter can intentionally introduce memory into the signal in the form of inter-symbol interference. The first interleaver can change a temporal order of the symbols in the signal. The neural network can be trained with positive mappings between transmitted and decoded training signals, or negative mappings between training signals and erroneous decoded signals that are known to contain errors.

A receiver may use a trained AI model to recover a faulted 5G/6G message by interpreting the meaning or intent of the message by correlating the message content with one of the expected message types. For example, the AI model may consider changes to the message, for consistency with an associated error-detection code, thereby producing a series of candidate messages. The AI model can then determine a likelihood that each of the candidate messages is correct, in the context of the receiver (such as an action or condition of the receiver, or a planned activity of the receiver) or is commonly received in that context. For example, the AI model can be trained to recognize the expected messages or message types, and thereby indicate which candidate message has the highest likelihood of being correct. The AI model may also consider waveform parameters to identify likely faults.

A method of decoding a linear block code transmitted over a transmission channel subject to noise, comprising receiving, over a transmission channel, a linear block code corresponding to a parity check matrix, propagating the received code through a neural network of one or more decoders, the neural network having an input layer, an output layer and a plurality of hidden layers comprising a plurality of nodes corresponding to transmitted messages over a plurality of edges of a bipartite graph representation of the encoded code and a plurality of edges connecting the plurality of nodes, each edge having source node and destination nodes is assigned with a weight calculated during a training session of the neural network, the propagation follows a propagation path through the neural network dictated by respective weights of the edges and outputting a recovered version of the code according to a final output of the neural network.

Techniques for improving the bit error rate (BER) performance of an error correction system are described. In an example, the error correction system implements low-density parity-check (LDPC) decoding that uses bit flipping. In a decoding iteration, a feature map is generated for a bit of an LDPC codeword. The bit corresponds to a variable node. The feature map is input to a neural network that is trained to determine whether bits should be flipped based on corresponding feature maps. An output of the neural network is accessed. The output indicates that the bit should be flipped based on the feature map. The bit is flipped in the decoding iteration based on the output of the neural network.