PaperlayerView exact revision on arXiv

Source-linked AI summary

MIMO Channel Information Feedback Using Deep Recurrent Network

Chao Lu, Wei Xu, Hong Shen, Jun Zhu, Kezhi Wang

arXiv:1811.07535v1cs.ITmath.IT

TL;DR

Accurate transmitter-side CSI is important for MIMO performance, but codebook feedback can have high overhead and complexity. The paper proposes recurrent compression and uncompression modules that exploit channel correlation, alongside a parameter-reduced design. The authors report improved performance over existing methods, including gains for reduced-parameter PR-RecCsiNet over CsiNet.

  • Problem

    Massive MIMO needs accurate transmitter-side CSI, while codebook-based feedback can incur high overhead and its complexity grows exponentially with quantization bits.

    Source: §I · p. 1

  • Method

    The proposed network uses recurrent compression and uncompression modules to exploit temporal channel correlation, with a projection-based variant that reduces parameters.

    Source: §I · p. 1, §II · p. 2, B. Recurrent Compression and… · p. 3

  • Results

    Across compression ratios, RecCsiNet shows noticeable advantages, while PR-RecCsiNet sharply reduces parameter size and beats CsiNet; quantized proposed networks also outperform conventional methods under mmWave channels.

    Source: §IV · p. 4

  • Takeaways & Limitations

    PR-RecCsiNet is described as a promising option for applications limited by memory and computation.

    Source: §IV · p. 4

    The work focuses on compression and uncompression modules; feature-extraction and recovery-module optimization and a compact integral multi-user design remain for further investigation.

    Source: §V · p. 4

Abstract

from arXiv · show

In a multiple-input multiple-output (MIMO) system, the availability of channel state information (CSI) at the transmitter is essential for performance improvement. Recent convolutional neural network (NN) based techniques show competitive ability in realizing CSI compression and feedback. By introducing a new NN architecture, we enhance the accuracy of quantized CSI feedback in MIMO communications. The proposed NN architecture invokes a module named long short-term memory (LSTM) which admits the NN to benefit from exploiting temporal and frequency correlations of wireless channels. Compromising performance with complexity, we further modify the NN architecture with a significantly reduced number of parameters to be trained. Finally, experiments show that the proposed NN architectures achieve better performance in terms of both CSI compression and recovery accuracy.

I. INTRODUCTION

Massive MIMO depends on accurate transmitter-side CSI, while conventional codebook feedback can impose high overhead and complexity. This work introduces recurrent compression and uncompression modules and a parameter-reduction method for CSI feedback.

I. INTRODUCTION

  • Accurate transmitter-side CSI is critical to realize massive MIMO’s system-capacity, energy-efficiency, and anti-interference gains, while downlink pilot and uplink feedback overhead can be high.Codebook quantization may also fail to scale performance linearly with quantization bits because its complexity grows exponentially.Source: §I · p. 1
  • Unlike parallel work [6], which enhances channel recovery, this work focuses on feature compression and uncompression, where fully connected networks may not track temporal correlations effectively.The compression and uncompression modules in and [6] use linear fully connected networks.Source: §I · p. 1
  • The proposed network uses memory-based modules for CSI compression and feedback, with performance reported as significantly improved over existing methods.The authors also propose reducing the network’s training complexity while retaining its performance advantage.Source: §I · p. 1

II. CSI COMPRESSION AND FEEDBACK

The section describes an FDD MIMO CSI-feedback pipeline and motivates neural compression as an alternative to codebooks. It also introduces temporal channel correlation as a basis for the proposed memory-based design.

II. CSI COMPRESSION AND FEEDBACK

  • The system model considers an FDD downlink with Nt base-station antennas and a single antenna per user equipment, with the received signal described at each subcarrier.The surrounding model identifies channel response, precoding vector, data symbol, and additive noise.Source: §II · p. 1
  • Codebook quantization can miss channel sparsity and requires exponentially growing storage and computation as quantization bits increase; neural networks offer another route to CSI compression and feedback.At the user equipment, an estimated channel is quantized and the information is sent to the base station for CSI recovery.Source: §II · p. 1
  • The channel is transformed to the angular-delay domain, split into real and imaginary parts, normalized, and compressed by an encoder for feedback before a decoder recovers it.The encoder output and decoder input each have size M; γ = M/N denotes the channel-information compression ratio.Source: §II · p. 2
  • The baseline CsiNet encoder uses convolution and a linear fully connected layer, while its decoder uses a fully connected layer, two RefineNets, and a convolutional layer.The RefineNet uses three convolution operations and a jump connection.Source: §II · p. 2
  • Because channels vary slowly in many massive-MIMO applications, adjacent CSI inputs can be correlated and that correlation can be exploited for more efficient compression.The section represents a channel sequence using a model in which α represents correlation between adjacent CSI inputs.Source: §II · p. 2

III. PROPOSED CSI NEURAL NETWORK WITH MEMORY

This section introduces the proposed neural-network design through compression and uncompression modules with memory.

III. PROPOSED CSI NEURAL NETWORK WITH MEMORY

  • The section presents design details for proposed compression and uncompression modules that include memory.The associated figure is identified as an LSTM cell structure.Source: §III · p. 2, Figure 2 · p. 2

A. Recurrent Compression and Uncompression Modules

The proposed Encoder and Decoder use LSTM-based compression and uncompression modules to exploit temporal correlations in channel sequences. Their memory mechanism retains information across inputs, while parallel linear FCN paths provide jump connections.

A. Recurrent Compression and Uncompression Modules

  • LSTM-based compression and uncompression modules replace linear FCNs to exploit temporal correlations in channel sequences.The compression module uses parallel LSTM and linear FCN flows, with the LSTM learning residual features; the BS uncompression module likewise uses LSTM and FCN flows.Source: A. Recurrent Compression and… · p. 2
  • A linear FCN jump connection accompanies the compression LSTM, and the two flows are added at the output because both project between N and M dimensions.The paper says the FCN jump connection can accelerate convergence and reduce the vanishing-gradient problem.Source: A. Recurrent Compression and… · p. 2
  • At the BS, uncompression uses linear FCN and LSTM flows with symmetric input and output sizes relative to compression.The Encoder separates feature extraction from compression, while the Decoder separates feature uncompression from channel-matrix recovery.Source: A. Recurrent Compression and… · p. 2
  • The LSTM memory mechanism carries forward a previous cell state and combines it with the present input, allowing it to compress temporal redundancy.Learned parameters determine how much information is kept or forgotten.Source: A. Recurrent Compression and… · p. 3
  • The BS compression and uncompression modules account for (NM) + (4N^2 + 4NM + 4N) trainable parameters, while the UE has the same order of magnitude.The paper notes that LSTM can be replaced by GRU without much additional change to the architecture.Source: A. Recurrent Compression and… · p. 3

B. Recurrent Compression and Uncompression Modules with Reduced Complexity

PR-RecCsiNet reduces recurrent-network parameters by projecting features before the LSTM and using serial FCN–LSTM connections. At compression ratio 1/16, its parameter count falls from about 19M for RecCsiNet to 0.8M, with possible information loss relative to RecCsiNet.

B. Recurrent Compression and Uncompression Modules with Reduced Complexity

  • PR-RecCsiNet uses serial FCN–LSTM connections, unlike RecCsiNet's parallel connections, and reduces the LSTM input size with an FCN projection.The projection allows the LSTM input and output to have the same size, so they can be linked directly without dimension transformation before addition.Source: B. Recurrent Compression and… · p. 3
  • At compression ratio 1/16, PR-RecCsiNet has 0.8M parameters versus about 19M for RecCsiNet.The passage says PR-RecCsiNet's parameter size remains comparable with CsiNet and CsiNet-LSTM; Table I is identified as reporting parameter counts.Source: B. Recurrent Compression and… · p. 3, Table I · p. 3
  • The FCN projection can cause information loss compared with RecCsiNet because it compresses extracted channel features to a lower dimension.The reduced LSTM input size is the stated basis for the parameter reduction.Source: B. Recurrent Compression and… · p. 3

IV. NUMERICAL RESULTS

Experiments evaluate recurrent CSI-feedback networks on COST2100 and mmWave channels, including quantized and unquantized settings. The proposed networks outperform conventional methods in quantized mmWave tests, while temporal correlation and parameter count distinguish the architectures.

IV. NUMERICAL RESULTS

  • Experiments use COST2100 and mmWave channels, four-step sequences, and end-to-end MSE training; the networks use compression ratios 1/64 and 1/32.The setup uses 32 BS antennas and 1024 subcarriers, with 100,000 training, 30,000 validation, and 20,000 testing sequences.Source: §IV · pp. 3–4
  • The quantized proposed networks outperform conventional methods on mmWave channels.Table II reports NMSE (dB) for MmWave Urban Micro at α = 0.1.Source: Table II · p. 4
  • As α increases, temporal correlation decreases and NMSE grows; with temporal correlation, NMSE drops noticeably before convergence across time steps.The correlation-coefficient comparison uses RecCsiNet at compression ratio 1/16; the time-step comparison uses α = 0.1.Source: §IV · p. 4, Figure 4 · p. 4, Figure 5 · p. 4
  • RecCsiNet shows advantages across COST2100 compression ratios, while PR-RecCsiNet uses sharply fewer parameters and beats CsiNet.CsiNet-LSTM has similar gains at some compression ratios but almost five times RecCsiNet’s parameter size.Source: §IV · p. 4
  • At compression ratio 1/64, AoD codebook processing takes 320 ms, versus 0.1–1.1 ms for the tested neural networks.The networks were tested on a GTX 1080Ti GPU and the codebook method on an i7-6800k CPU.Source: §IV · p. 4

V. CONCLUSION

The conclusion presents recurrent compression and uncompression modules for time-varying channel features, alongside a method for reducing parameter size. It identifies module optimization and a more compact multi-user design as future work.

V. CONCLUSION

  • The proposed recurrent compression and uncompression modules exploit time-varying features, and a separate method reduces network parameter size.The paper limits its design to the compression and uncompression modules.Source: §V · p. 4
  • Further optimizing feature extraction and channel recovery is expected to improve encoder-decoder performance.For multi-user MIMO, the BS would need a Decoder Network for each user; a more compact integral design needs further investigation.Source: §V · p. 4
Loading 1811.07535v1…