Spatio-Temporal Representation with Deep Neural Recurrent Network in MIMO CSI Feedback

Huaming Wu; Xiangyi Li

REVIEW 4 major objections 5 minor 14 references

Spatio-Temporal Representation with Deep Neural Recurrent Network in MIMO CSI Feedback

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that adding recurrent memory and decoupled spatio-temporal convolutions to a CSI feedback autoencoder improves MIMO channel recovery at every tested compression ratio.

desk verdict A plausible incremental improvement to RNN-based CSI feedback whose central claim is not yet established because the closest baseline, CsiNet-LSTM, is never evaluated. read the letter →

arxiv 1908.07934 v2 pith:A5B7JRDG submitted 2019-08-04 eess.SP cs.LG

Xiangyi Li , Huaming Wu This is my paper · ORCID

classification eess.SPcs.LG

keywords CSIfeedbackmassiveMIMOfrequencydivisionduplexrecurrentneuralnetworkConvLSTMpseudo-3Dconvolutiondepthwiseseparablechannelstateinformationcompression

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that an autoencoder for MIMO channel-state-information (CSI) feedback gains accuracy by explicitly modelling temporal correlation with recurrent layers, not just spatial structure. It proposes ConvlstmCsiNet, which adds a ConvLSTM memory layer in feature extraction, replaces standard convolutions with depthwise separable convolutions in the recovery module, and decorates the encoder with P3D blocks that decouple spatial and temporal filtering. On the indoor COST2100 scenario, the best variant ConvlstmCsiNet-A reaches log-NMSE of $-28.4$ dB at compression ratio 1/4, versus $-21.5$ dB for RecCsiNet and $-17.5$ dB for CsiNet, and keeps the lead at 1/8, 1/16, and 1/32. If the claim holds, recurrent temporal modelling is a practical way to cut CSI feedback overhead in FDD massive MIMO without sacrificing reconstruction quality.

What carries the argument

Two named mechanisms carry the argument. ConvLSTM is a convolutional long short-term memory layer that lets the encoder accumulate temporal context while keeping local spatial detail; it sits in the feature-extraction module. P3D blocks are the paper's main refinement: a 3D convolution of size $T_d\times S_d\times S_d$ is decoupled into a $1\times S_d\times S_d$ spatial filter and a $T_d\times 1\times 1$ temporal filter, arranged in three ways (A cascaded, B parallel, C combined), with the cascaded P3D-A giving the best results. Depthwise separable convolution in the recovery module splits each Conv3D into a per-channel depthwise step and a $1\times1\times1$ pointwise step, shrinking parameters and enabling cross-channel information mixing. These pieces work together to make the autoencoder both smaller and more accurate than its predecessors.

What would settle it

Train and test ConvlstmCsiNet-A and RecCsiNet on the same COST2100 channels but with time evolution generated by a Jakes Doppler spectrum (or by measured channel traces) instead of Eq. (3), at CR 1/4 and 1/32; if ConvlstmCsiNet-A no longer beats RecCsiNet by the reported margins, the claim is bound to the AR(1) data model.

Watch

Extended reading notes

Core claim

The paper's central claim is that a CSI feedback autoencoder that separates spatial and temporal feature extraction and injects recurrent memory at the encoder recovers time-varying MIMO channels more accurately than prior deep-learning feedback networks. ConvlstmCsiNet-A, the variant with cascaded P3D-A blocks, reports the best normalized mean-square error and cosine similarity at every tested compression ratio, with the largest gain relative to CsiNet appearing at the most aggressive ratio (114.9% NMSE improvement at CR 1/32). The improvement over RecCsiNet is smaller but still positive at all ratios, and reconstruction quality degrades gracefully as the temporal correlation parameter $\alpha$ rises, confirming that the network is exploiting the assumed time-sequence structure.

Load-bearing premise

The load-bearing premise is that successive channel matrices follow a first-order autoregressive law $H_{t+1}=(1-\alpha^2)H_t+\alpha^2 u_t$ with a single correlation parameter $\alpha=0.1$; if real massive MIMO channels move differently, the temporal-correlation gains reported here may not transfer.

Editorial extensions

If this is right

At all four compression ratios (1/4, 1/8, 1/16, 1/32), the proposed ConvlstmCsiNet-A reports lower NMSE and higher cosine similarity than CsiNet and RecCsiNet on the indoor COST2100 scenario.
The relative gain over CsiNet grows monotonically as compression tightens, so recurrent temporal modelling is most valuable exactly when the feedback codeword is shortest.
Decoupling spatial and temporal filters (P3D) improves on a single 3D convolution, so the architectural change is independently useful beyond the recurrent layer.
Replacing standard Conv3D with depthwise separable convolution reduces parameter count and computation, making the accuracy gains compatible with practical deployment constraints.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the temporal model in Eq. (3) uses one scalar $\alpha$ for the whole channel; a direct extension would train on measured time sequences with per-subcarrier or per-path Doppler and test whether ConvLSTM still buys the same margin.
Beyond the paper: because P3D and depthwise separable convolution are modular substitutions, they could be combined with quantization-aware training or attention mechanisms to push CSI feedback toward real bit-rate constraints, something this letter does not evaluate.
Beyond the paper: the reported gains are in reconstruction NMSE and cosine similarity, not in end-to-end throughput; whether the better reconstruction translates into spectral-efficiency gains under actual precoding is an open testable question.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: this is a sensible engineering paper with a clear architecture and internally consistent results, but the central claim is broader than the evidence. The architecture combines published ideas — ConvLSTM, depthwise separable convolution, P3D blocks — in a new way for CSI feedback. That is a legitimate extension of RecCsiNet, and the reported gains over CsiNet and RecCsiNet are large and consistent at all four compression ratios. The CR 1/4 margin of roughly 7 dB NMSE over RecCsiNet is meaningful if it holds. The paper deserves credit for a clean experiment and a clear figure.

The soft spots are real. The most important is the missing CsiNet-LSTM baseline. The paper cites [2] as an RNN-based extension of CsiNet that exploits temporal correlation, and the introduction says it outperforms CsiNet. Yet the experiments compare only against CsiNet and RecCsiNet. The ConvLSTM layer is precisely the temporal-correlation mechanism, so without CsiNet-LSTM you cannot tell how much of the gain is new and how much is just having any recurrent structure. The headline claim 'outperforms existing DL-based methods' does not follow from the tested set.

Second, the temporal channel model in Eq. (3) is an AR(1) process with F=(1-alpha^2)I, G=alpha^2 I, alpha=0.1. This imposes exactly the temporal correlation the architecture is designed to exploit, and the same model generates training and test data. That is a limitation, though not fatal — it is standard practice in this line, and the authors do examine alpha sensitivity in Fig. 6. Still, the conclusion says the architecture has potential for real MIMO systems, and that inference is weak without a more realistic channel model.

Third, there are no error bars or repeated runs, no parameter counts despite the claim that depthwise separable convolution shrinks the model, and no code or data. These are fixable and would make the comparison much more convincing.

Who is this for? The paper will be useful to people working on DL-based CSI feedback, particularly those building on RecCsiNet. It is readable and the architecture is well presented. It deserves a serious referee, but the referee should ask for the CsiNet-LSTM baseline, parameter counts, and ideally confidence intervals before the central claim is accepted.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes ConvlstmCsiNet, an autoencoder for massive MIMO CSI feedback that extends RecCsiNet by adding a ConvLSTM feature-extraction layer, replacing standard 3D convolutions in the RefineNet recovery module with depthwise separable convolutions, and substituting P3D blocks for the feature-extraction convolution. The paper reports NMSE and cosine-similarity results on the COST2100 indoor scenario at four compression ratios, with channel time sequences generated by the first-order state-space model in Eq. (3), and claims that the proposed ConvlstmCsiNet-A outperforms existing DL-based CSI feedback methods.

Significance. If the claims hold, the architecture would be a useful improvement: all proposed variants outperform CsiNet and RecCsiNet on the tested scenario, with the largest gain at CR 1/4 (log-NMSE -28.4 dB versus -21.5 dB for RecCsiNet), and the qualitative reconstruction figure supports the numerical result. The paper also clearly describes the training protocol and data preprocessing. However, the significance is tempered by the absence of the closest published RNN baseline, CsiNet-LSTM, and by the fact that the temporal correlation is injected by the synthetic AR(1) model rather than measured from a time-varying channel; the general claim 'outperforms existing DL-based methods' is therefore broader than the tested comparison set.

major comments (4)

[Section IV, Tables I and II] CsiNet-LSTM [2], which the Introduction explicitly identifies as an RNN-based extension of CsiNet that exploits temporal correlation, is never evaluated. Since the proposed architecture's main novelty is temporal-correlation modeling and it builds directly on RecCsiNet, the claimed superiority over 'existing DL-based methods' is not established against the most relevant published baseline. Please add CsiNet-LSTM to Tables I and II and discuss the margins over it for all compression ratios.
[Section II, Eq. (3), and Section IV data generation] The temporal sequence is generated from the state-space model H_{t+1} = F H_t + G u_t with F = (1-alpha^2)I, G = alpha^2 I, and alpha = 0.1, and the same model generates the training and test data. The ConvLSTM and RNN components are precisely designed to exploit this injected temporal correlation, so the benchmark is tailored to the method. Please validate on measured or independently simulated time-varying channels (for example, true COST2100 time evolution or channel traces) and report whether the reported gains persist under more realistic temporal dynamics.
[Section III.A.2] One of the two stated contributions is that depthwise separable convolution 'shrink[s] the model,' but no parameter count, FLOPs, or runtime comparison is reported for any model. Without these numbers, the model-compression claim is unverified. Please report parameter counts for CsiNet, RecCsiNet, and the proposed variants, together with the corresponding NMSE results.
[Tables I and II] No error bars, multiple independent runs, or statistical significance test are reported. This matters because some improvements are small: for example, ConvlstmCsiNet versus RecCsiNet at CR 1/32 is reported as a 0.75% improvement in Table II, which corresponds to roughly 0.1 dB in log-NMSE. The authors should report mean and standard deviation over repeated training runs, or otherwise indicate which small-margin entries are reproducible.

minor comments (5)

[Section IV, Table II] The 'percentage improvement' values in Table II appear to be computed on the log-NMSE values in Table I (for example, 64.0% matches (-28.4 + 17.5)/(-17.5) rather than a ratio of linear NMSE). Please define this metric explicitly; reporting linear NMSE ratios would be more physically meaningful.
[Section IV] There is a typo in 'Tabel II,' and 'spacial' in the Conclusion should be 'spatial.'
[Section IV, Figure 6] Figure 6 is referenced in the text but the caption and axis labels are not described in the manuscript body; please add them and indicate whether the same trend is observed for the baselines.
[Section III.B] The sentence stating that P3D blocks 'replace the convolution in feature compression module' seems to contradict the surrounding text and Figure 1, which place P3D in the feature extraction module; please clarify.
[Section III.A, Eq. (5)] The displayed loss divides by MT but not by the number of spatial elements Nt*Nc, so it is not exactly the mean squared error as written; this constant scaling does not change the training dynamics, but the formula should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are experimental architecture comparisons, with no fitted parameter relabeled as a prediction and no load-bearing self-citation chain.

full rationale

This is an empirical architecture paper rather than a derivation chain, so there is no analytical result that could reduce to its own inputs. The only potentially concerning choice is that training and test channels are generated from the state-space model in Eq. (3), H_{t+1} = F H_t + G u_t with F=(1-alpha^2)I and G=alpha^2 I, which injects the temporal correlation that the RNN/ConvLSTM components are designed to exploit. That is experimental design, not circularity: the paper does not fit a parameter from the test set and then present that fit as a prediction, and the synthetic channel model is explicitly attributed to an external reference ([9]) rather than to the authors' own prior results. All baselines compared in Tables I and II, namely CsiNet [6] and RecCsiNet [7], are external works, and the cited CsiNet-LSTM [2] is also from other authors; there are no self-citations, uniqueness theorems imported from the authors, or ansatzes smuggled in via the authors' prior work. The omission of CsiNet-LSTM from the experimental comparison is a legitimate concern about the breadth of the 'outperforms existing DL-based methods' claim, but that is a baseline-completeness/fairness issue, not a circular derivation. No circular step can be exhibited with a specific equation-to-equation reduction, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a synthetic temporal channel model (Eq. 3) that the RNN architecture is designed to exploit, plus the standard DFT sparsity assumption and the COST2100 simulator. No new physical entities are introduced. The three free parameters (alpha, sigma_u, T) are hand-chosen data-generation settings that shape the reported gains.

free parameters (3)

alpha = 0.1 for main results; varied in Fig. 6
Temporal correlation coefficient in H_{t+1} = F H_t + G u_t, Eq. (3). Chosen by hand; it determines the strength of temporal correlation the RNN is meant to exploit and directly influences all reported NMSE values.
sigma_u = 1e-3
Standard deviation of the additive noise u_t in Eq. (3), chosen as 'tiny white Gauss noise' in Section IV.
T = 4
Number of recurrent steps in each CSI sequence, set 'for convenience' in Section IV; defines the temporal context available to the RNN.

assumptions (4)

domain assumption The angular-delay domain CSI matrix is sparse after 2D DFT, and only the first Nc columns contain significant energy.
Invoked in Section II to truncate H from 1024x32 to 32x32 before feedback; inherited from prior CsiNet work and not validated here.
ad hoc to paper The time-varying channel follows H_{t+1} = F H_t + G u_t, with F = (1-alpha^2)I and G = alpha^2 I (Eq. 3).
This synthetic model generates the training and test temporal sequences. The RNN's advantage is contingent on this structure; real channels may not match it.
domain assumption COST2100 with the default indoor picocellular settings at 5.3 GHz is a realistic channel model.
All static CSI examples come from the COST2100 simulator; no measured field data is used to validate the reported gains.
domain assumption The feedback channel transmits the compressed codeword without loss or quantization error.
Stated in Section III-A: 'the feedback channel is assumed to be perfect enough to transmit the compressed codeword without any damage or loss.' This idealization removes feedback-link impairments from the evaluation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spatio-Temporal Representation with Deep Neural Recurrent Network in MIMO CSI Feedback." pith.science (2026). https://pith.science/paper/A5B7JRDG

@misc{pith2026190807934,
  author       = {Pith},
  title        = {Pith review of: Spatio-Temporal Representation with Deep Neural Recurrent Network in MIMO CSI Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A5B7JRDG}},
  note         = {Machine review of arXiv:1908.07934}
}

read the original abstract

In multiple-input multiple-output (MIMO) systems, it is crucial of utilizing the available channel state information (CSI) at the transmitter for precoding to improve the performance of frequency division duplex (FDD) networks. One of the mainchallenges is to compress a large amount of CSI in CSI feedback transmission in massive MIMO systems. In this paper, we propose a deep learning (DL)-based approach that uses a deep recurrent neural network (RNN) to learn temporal correlation and adopts depthwise separable convolution to shrink the model. The feature extraction module is also elaborately devised by studyingdecoupled spatio-temporal feature representations in different structures. Experimental results demonstrate that the proposed approach outperforms existing DL-based methods in terms of recovery quality and accuracy, which can also achieve remarkable robustness at low compression ratio (CR).

Figures

Figures reproduced from arXiv: 1908.07934 by the authors.

**Figure 2.** The structure of three gates in ConvLSTM [10] [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Three designs of P3D block Considering whether the temporal and spatial filters should directly or indirectly influence each other or the final output, [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: NMSE of the proposed NN at CR=1/4 in different correlation [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 5.** Figure 5: Absolute value of original (α = 0.1) and reconstructed CSI images at different CRs [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

[2]

Deep learning-based CSI feedback approach for time-varying massive MIMO channels,

T. Wang, C.-K. Wen, S. Jin, and G. Y . Li, “Deep learning-based CSI feedback approach for time-varying massive MIMO channels,” IEEE Wireless Communications Letters, 2018

work page 2018
[1]

Deep learning based MIMO communications,

T. J. O’Shea, T. Erpek, and T. C. Clancy, “Deep learning based MIMO communications,” arXiv preprint arXiv:1707.07980 , 2017

arXiv 2017
[3]

Deep learning for wireless physical layer: Opportunities and challenges,

T. Wang, C. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, “Deep learning for wireless physical layer: Opportunities and challenges,” China Communications, vol. 14, pp. 92–111, Nov 2017

work page 2017
[4]

Downlink training techniques for FDD massive MIMO systems: Open-loop and closed-loop training with memory,

J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for FDD massive MIMO systems: Open-loop and closed-loop training with memory,” IEEE Journal of Selected Topics in Signal Processing , vol. 8, no. 5, pp. 802–814, 2014

work page 2014
[5]

On capacity of large- scale MIMO multiple access channels with distributed sets of correlated antennas,

J. Zhang, C. Wen, S. Jin, X. Gao, and K. Wong, “On capacity of large- scale MIMO multiple access channels with distributed sets of correlated antennas,” IEEE Journal on Selected Areas in Communications , vol. 31, pp. 133–148, February 2013

work page 2013
[6]

Deep learning for massive MIMO CSI feedback,

C. Wen, W. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Communications Letters , vol. 7, pp. 748–751, Oct 2018

work page 2018
[7]

MIMO channel information feedback using deep recurrent network,

C. Lu, W. Xu, H. Shen, J. Zhu, and K. Wang, “MIMO channel information feedback using deep recurrent network,” IEEE Communications Letters , vol. 23, no. 1, pp. 188–191, 2018

work page 2018
[8]

Mobilenetv2: Inverted residuals and linear bottlenecks,

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp. 4510–4520, 2018

work page 2018

Show all 14 references

[9]

An EM-based forward-backward kalman ﬁlter for the estimation of time-variant channels in OFDM,

T. Y . Al-Naffouri, “An EM-based forward-backward kalman ﬁlter for the estimation of time-variant channels in OFDM,” IEEE Transactions on Signal Processing, vol. 55, no. 7, pp. 3924–3930, 2007

2007
[10]

Convolutional LSTM network: A machine learning approach for precipitation nowcasting,

S. Xingjian, Z. Chen, H. Wang, D.-Y . Yeung, W.-K. Wong, and W.-c. Woo, “Convolutional LSTM network: A machine learning approach for precipitation nowcasting,” in Advances in neural information processing systems, pp. 802–810, 2015

2015
[11]

Learning spatio-temporal representation with pseudo-3d residual networks,

Z. Qiu, T. Yao, and T. Mei, “Learning spatio-temporal representation with pseudo-3d residual networks,” in proceedings of the IEEE International Conference on Computer Vision , pp. 5533–5541, 2017

2017
[12]

Inception-v4, inception-resnet and the impact of residual connections on learning,

C. Szegedy, S. Ioffe, and V . Vanhoucke, “Inception-v4, inception-resnet and the impact of residual connections on learning,” AAAI Conference on Artiﬁcial Intelligence, 02 2016

2016
[13]

Identity mappings in deep residual networks,

K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” in European conference on computer vision , pp. 630–645, Springer, 2016

2016
[14]

The COST 2100 MIMO channel model,

L. Liu, C. Oestges, J. Poutanen, K. Haneda, P. Vainikainen, F. Quitin, F. Tufvesson, and P. De Doncker, “The COST 2100 MIMO channel model,” IEEE Wireless Communications, vol. 19, no. 6, pp. 92–99, 2012

2012

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [2]

Deep learning-based CSI feedback approach for time-varying massive MIMO channels,

T. Wang, C.-K. Wen, S. Jin, and G. Y . Li, “Deep learning-based CSI feedback approach for time-varying massive MIMO channels,” IEEE Wireless Communications Letters, 2018

work page 2018

[2] [1]

Deep learning based MIMO communications,

T. J. O’Shea, T. Erpek, and T. C. Clancy, “Deep learning based MIMO communications,” arXiv preprint arXiv:1707.07980 , 2017

arXiv 2017

[3] [3]

Deep learning for wireless physical layer: Opportunities and challenges,

T. Wang, C. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, “Deep learning for wireless physical layer: Opportunities and challenges,” China Communications, vol. 14, pp. 92–111, Nov 2017

work page 2017

[4] [4]

Downlink training techniques for FDD massive MIMO systems: Open-loop and closed-loop training with memory,

J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for FDD massive MIMO systems: Open-loop and closed-loop training with memory,” IEEE Journal of Selected Topics in Signal Processing , vol. 8, no. 5, pp. 802–814, 2014

work page 2014

[5] [5]

On capacity of large- scale MIMO multiple access channels with distributed sets of correlated antennas,

J. Zhang, C. Wen, S. Jin, X. Gao, and K. Wong, “On capacity of large- scale MIMO multiple access channels with distributed sets of correlated antennas,” IEEE Journal on Selected Areas in Communications , vol. 31, pp. 133–148, February 2013

work page 2013

[6] [6]

Deep learning for massive MIMO CSI feedback,

C. Wen, W. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Communications Letters , vol. 7, pp. 748–751, Oct 2018

work page 2018

[7] [7]

MIMO channel information feedback using deep recurrent network,

C. Lu, W. Xu, H. Shen, J. Zhu, and K. Wang, “MIMO channel information feedback using deep recurrent network,” IEEE Communications Letters , vol. 23, no. 1, pp. 188–191, 2018

work page 2018

[8] [8]

Mobilenetv2: Inverted residuals and linear bottlenecks,

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp. 4510–4520, 2018

work page 2018