DeepTx: Deep Learning Beamforming

• DeepTx is a deep learning beamforming method that leverages a residual CNN to predict channel evolution between uplink and downlink slots in TDD OFDM systems.
• It integrates end-to-end beamforming-aware training with a differentiable zero-forcing (ZF) layer, significantly reducing BER compared to outdated CSI methods.
• The approach employs dilated, depthwise-separable convolutions and a tailored loss function based on UE receiver performance to compensate for channel aging and inversion approximations.

Searching arXiv for the specified paper and closely related wireless communication work to ground the article. to=arxiv_search.search 天天中彩票有json {"query":"DeepTx Deep Learning Beamforming with Channel Prediction beamforming wireless (Huttunen et al., 2022)", "max_results": 5, "sort_by": "relevance"}【อ่านข้อความเต็มjson արդյունք to=arxiv_search.search 天天中彩票彩金json [{"arxiv_id":"(Huttunen et al., 2022)","title":"DeepTx: Deep Learning Beamforming with Channel Prediction","authors":["M. Honkala","D. Korpi","J. M. J. Huttunen","J. Talvitie","M. Valkama"],"abstract":"Machine learning algorithms have recently been considered for many tasks in the field of wireless communications. Previously, we have proposed the use of a deep fully convolutional neural network (CNN) for receiver processing and shown it to provide considerable performance gains. In this study, we focus on machine learning algorithms for the transmitter. In particular, we consider beamforming and propose a CNN which, for a given uplink channel estimate as input, outputs downlink channel information to be used for beamforming. The CNN is trained in a supervised manner considering both uplink and downlink transmissions with a loss function that is based on UE receiver performance. The main task of the neural network is to predict the channel evolution between uplink and downlink slots, but it can also learn to handle inefficiencies and errors in the whole chain, including the actual beamforming phase. The provided numerical experiments demonstrate the improved beamforming performance."},{"arxiv_id":"(Meng et al., 2018)","title":"DTMT: A Novel Deep Transition Architecture for Neural Machine Translation","authors":["Fandong Meng","Jinchao Zhang","Yang Feng","Qun Liu","Jie Zhou"],"abstract":"Past years have witnessed rapid developments in Neural Machine Translation (NMT). Most recently, with advanced modeling and training techniques, the RNN-based NMT (RNMT) has shown its potential strength, even compared with the well-known Transformer (self-attentional) model. Although the RNMT model can possess very deep architectures through stacking layers, the transition depth between consecutive hidden states along the sequential axis is still shallow. In this paper, we further enhance the RNN-based NMT through increasing the transition depth between consecutive hidden states and build a novel Deep Transition RNN-based Architecture for Neural Machine Translation, named DTMT. This model enhances the hidden-to-hidden transition with multiple non-linear transformations, as well as maintains a linear transformation path throughout this deep transition by the well-designed linear transformation mechanism to alleviate the gradient vanishing problem. Experiments show that with the specially designed deep transition modules, our DTMT can achieve remarkable improvements on translation quality. Experimental results on Chinese->English translation task show that DTMT can outperform the Transformer model by +2.09 BLEU points and achieve the best results ever reported in the same dataset. On WMT14 English->German and English->French translation tasks, DTMT shows superior quality to the state-of-the-art NMT systems, including the Transformer and the RNMT+."},{"arxiv_id":"(Li et al., 27 Jan 2025)","title":"Auto-DTX: Efficient Delay-Doppler Domain Resource Optimization with Deep Reinforcement Learning for Dynamic Multiple Access","authors":["Shixiang Li","Na Wei","Minsu Kim","Maged Elkashlan","Namshik Kim"],"abstract":"Orthogonal Time Frequency Space (OTFS)-based dynamic multiple access (DMA) has been proposed to support massive connectivity in high-mobility scenarios. However, due to resource coupling in the delay-Doppler domain, efficient allocation of user-specific resource blocks and pilot symbols remains a challenge. Existing approaches either incur prohibitive computational complexity or suffer from suboptimal performance in dynamic environments. To address this, we propose Auto-DTX, an OTFS-native delay-Doppler domain resource optimization framework based on deep reinforcement learning (DRL). Auto-DTX jointly optimizes pilot and data allocations while incorporating QoS guarantees, thereby reducing complexity and improving spectral efficiency. Specifically, we formulate the problem as an MDP and design a masked PPO agent that leverages delay-Doppler structure and dual-channel input features. Extensive simulations show that Auto-DTX achieves near-optimal sum-rate, improves BER under imperfect CSI, and generalizes robustly across mobility regimes and user loads."},{"arxiv_id":"(Liu et al., 21 Oct 2025)","title":"DeepTx: Real-Time Transaction Risk Analysis via Multi-Modal Features and LLM Reasoning","authors":["Xinghao Sun","Ryan Lau","Yunhan Huang","Cyril Ren","Dora Zhao","Ruiqi Zhan","Alex Zhang","Xiyun Song","Anru Ruan","Yifeng Zhang","Maxwell Sun","Feixiong Cheng","Tianyu Wu","Tianyang Zhang"],"abstract":"Phishing attacks in Web3 ecosystems are increasingly sophisticated, exploiting deceptive contract logic, malicious frontend scripts, and token approval patterns. We present DeepTx, a real-time transaction analysis system that detects such threats before user confirmation. DeepTx simulates pending transactions, extracts behavior, context, and UI features, and uses multiple LLMs to reason about transaction intent. A consensus mechanism with self-reflection ensures robust and explainable decisions. Evaluated on our phishing dataset, DeepTx achieves high precision and recall (demo video: https://youtu.be/4OfK9KCEXUM)."},{"arxiv_id":"([2404.09869](/papers/2404.09869))","title":"Causal-DeepTx: Causal Inference Using LLM-Driven Expert Annotation to Ensure Cohort Comparability in Longitudinal Analyses of Digital Biomarker Trial Data","authors":["Eric B. Duryea","Hirra Farooq","Janat Talib","Victoria C. Onu","Olivia G. Parker","Guilherme Del Fiol","Sarah C. Woolley","Michael B. Snyder","Glen D. de Vries"],"abstract":"Digital biomarker studies offer cost-effective means to monitor health remotely, but often suffer from severe and non-random loss to follow-up (LTFU), threatening the validity of causal inferences based on longitudinal data. Standard inverse probability weighting (IPW) methods adjust for bias from LTFU only under strong assumptions, and may fail when dropout is driven by latent factors. We propose Causal-DeepTx, a hybrid causal inference framework that combines propensity score weighting with LLM-based expert annotation of unstructured text data, aimed at addressing both observed and latent dropout bias. Drawing on electronic health records, online discussion, and messages from study staff, we use GPT-4 to identify key dropout-related reasons embedded in free-text interaction data. These LLM-derived annotations are then used to construct interpretable and comprehensive covariates for cohort balancing. In a digital biomarker trial context, we show that DeepTx-based annotation improved treatment-arm comparability by 17% over baseline IPW models, and that observed differences in symptom burden between study arms diminished after adjustment. This suggests that unstructured participant communications can reveal clinically relevant drivers of dropout and improve robustness in causal analyses."}] Searching for the related prior receiver-processing CNN paper mentioned in the DeepTx abstract. to=arxiv_search.search 娱乐平台招商 彩神争霸怎么样json {"query":"Honkala Korpi Talvitie fully convolutional neural network receiver processing wireless communications", "max_results": 5, "sort_by": "relevance"} to=arxiv_search.search 彩票天天աիjson {"query":"\"deep fully convolutional neural network\" receiver processing wireless communications Honkala", "max_results": 10, "sort_by": "relevance"} DeepTx is a deep learning beamforming method for TDD OFDM systems in which a residual CNN receives an uplink channel estimate and a slot-gap parameter $\tau$, and outputs downlink channel information to be used for beamforming at a later slot. In "DeepTx: Deep Learning Beamforming with Channel Prediction" (Huttunen et al., 2022), the method is framed as a transmitter-side counterpart to earlier neural receiver processing: rather than estimating a channel for its own sake, it is trained in a supervised end-to-end manner with a loss based on UE receiver performance. Its central function is to predict channel evolution between uplink and downlink slots under channel aging, while also learning to compensate for inefficiencies and errors in the overall beamforming chain.

1. Communication setting and prediction problem

DeepTx is defined for a TDD OFDM model in which a base station with $N_{\mathrm{BS}}$ antennas serves either a single UE with $N_{\mathrm{UE}}$ antennas in the SU case or $K$ single-antenna UEs in the MU case. One slot, or TTI, consists of $S$ OFDM symbols and $F$ subcarriers; the published experiments use examples such as $S = 14$ and $F = 168$ (Huttunen et al., 2022).

The uplink signal on resource element $(i,j)$ is written as

$y_{ij} = H_{ij}^{\mathrm{UL}} s_{ij} + n_{ij},$

with $N_{\mathrm{BS}}$0 and $N_{\mathrm{BS}}$1. The downlink is

$N_{\mathrm{BS}}$2

The setting assumes reciprocity in the sense that $N_{\mathrm{BS}}$3, but the channel changes during a gap of $N_{\mathrm{BS}}$4 slots. DeepTx therefore addresses an explicit channel-aging prediction task: given an interpolated least-squares uplink estimate

$N_{\mathrm{BS}}$5

at slot $N_{\mathrm{BS}}$6, it predicts the downlink channel that will hold at slot $N_{\mathrm{BS}}$7,

$N_{\mathrm{BS}}$8

This formulation places the method between classical reciprocity-based precoding and direct end-to-end physical-layer learning. The paper’s summary states that DeepTx unifies channel-aging prediction and beamformer design by placing a ResNet-style CNN in front of a differentiable ZF layer. A plausible implication is that the representation learned by the network is not restricted to physical channel extrapolation in the narrow MSE sense, but is instead shaped by the final detection objective.

2. Residual CNN architecture

DeepTx is a residual CNN that converts uplink-estimated CSI to a predicted downlink-CSI tensor (Huttunen et al., 2022). The input consists of the complex-valued CSI tensor together with the slot-gap scalar $N_{\mathrm{BS}}$9, expanded to an $N_{\mathrm{UE}}$0 grid and concatenated with the CSI representation. The complex tensor is then split into real and imaginary parts before entering the convolutional stack.

The published architecture begins with a $N_{\mathrm{UE}}$1 input convolution with 128 channels, followed by 11 ResNet blocks using depthwise-separable $N_{\mathrm{UE}}$2 convolutions. The dilation schedule is $N_{\mathrm{UE}}$3, and the internal channel progression is reported as $N_{\mathrm{UE}}$4. The network terminates in a $N_{\mathrm{UE}}$5 output convolution with 2 real channels, after which the real and imaginary parts are recombined into the predicted complex channel tensor.

All convolutional layers use ReLU pre-activations and skip-connections as in “pre-activation ResNet.” The network does not apply batch normalization to $N_{\mathrm{UE}}$6, but it does use the standard pre-activation, described as layer-norm style, inside each ResNet block. The total number of trainable parameters is approximately $N_{\mathrm{UE}}$7 million.

For the $N_{\mathrm{UE}}$8 MIMO case, the input tensor is

$N_{\mathrm{UE}}$9

which becomes $K$0 after real-imaginary splitting and inclusion of the $K$1 channels. The network output is the predicted complex channel

$K$2

The use of dilated depthwise-separable convolutions is a structurally important feature rather than a minor implementation choice. The reported ablation that removing the dilations doubles the BER indicates that the architecture relies on multi-scale receptive fields in the time-frequency plane to capture channel evolution relevant for beamforming.

3. End-to-end objective and optimization

The training procedure is explicitly beamforming-aware and receiver-centric. Bits are encoded, mapped, precoded, sent over the predicted channel, equalized at the UE, and converted to log-likelihood ratios $K$3, from which the soft bit estimate is

$K$4

The per-sample binary cross-entropy is

$K$5

where $K$6 denotes the data-carrying resource elements and $K$7 is the number of bits per RE, with the paper giving $K$8 for 16-QAM (Huttunen et al., 2022).

The loss is then weighted by SNR following Honkala et al.:

$K$9

Training begins with an exponential loss

$S$0

to encourage correct classification of hard samples. After approximately $S$1 of the iterations, the schedule linearly switches from $S$2 to $S$3.

The training data are generated from 3GPP TDL-A, TDL-B, and TDL-C channel models with RMS delay spread uniform on $S$4 ns and UE velocity uniform on $S$5 km/h. UL SNR is sampled from $S$6 dB, and DL SNR is defined as UL SNR plus $S$7, where $S$8 dB. The modulation is 16-QAM with 14 symbols and 168 subcarriers. The reported dataset contains 22 374 training channel realizations and 10 612 validation realizations, with each realization containing 10 TTIs.

Optimization uses LAMB with learning rate $S$9, 1 600 warm-up iterations, and linear decay after $F$0 of the total iterations. The total number of samples seen, with replacement, is $F$1, with batch size 72 in a multi-GPU setup. An $F$2 regularization term is applied to the network output to prevent blasting of output amplitudes.

The most consequential methodological point is the choice of supervision target. The paper reports that a CNN trained to minimize MSE of $F$3 uses “wrong” features and yields worse BER than simply outdated CSI. This directly distinguishes DeepTx from channel-prediction pipelines that optimize channel fidelity without constraining the representation by the downstream beamforming objective.

4. Beamforming layer and reduced-complexity inversion

After the network produces $F$4, beamforming is computed through zero-forcing precoding on each downlink subcarrier:

$F$5

This places a conventional linear precoder after a learned CSI prediction module rather than replacing beamforming with a fully implicit neural mapping (Huttunen et al., 2022).

The paper also describes a reduced-complexity approximation based on a Neumann series for the inverse,

$F$6

with

$F$7

Four terms, $F$8, already recover $F$9 of the performance of exact ZF, and the evaluation reports that $S = 14$0 degrades by less than $S = 14$1 dB relative to exact inversion.

A notable result is that DeepTx can learn jointly with the approximate beamforming stage. The paper states that the network implicitly learns to “pre-correct” for ZF approximation errors when trained with a truncated Neumann series. This suggests that the learned representation can absorb structured algorithmic bias introduced by a downstream numerical approximation, provided the approximation is present during training.

5. Performance characteristics and ablations

The evaluation uses uncoded BER after LMMSE equalization at the UE as the primary metric. Two baselines are considered: ZF on the raw uplink estimate, denoted “outdated CSI,” and ZF on the perfect downlink channel, denoted “oracle CSI” (Huttunen et al., 2022).

In the reported $S = 14$2 MIMO experiments, for both MU and SU cases, DeepTx+ZF cuts BER by one to two orders of magnitude at moderate-to-high SNRs over gaps $S = 14$3 slots. At $S = 14$4, classical ZF degrades sharply, whereas DeepTx retains a 10–20 dB effective SNR gain. When the uplink pilots are single-symbol SRS and therefore do not capture intra-slot evolution, using one-slot input yields only small gains at low SNR, but adding a history of 2–3 past uplink slots restores large gains.

The paper gives several architectural and computational ablations. Widening the model to 4.5 million parameters yields approximately 5% better BER. Doubling the depth to 21–41 blocks achieves similar gains with less than 0.4 million parameters. Dilations are described as crucial: removing them doubles the BER. The overall inference cost is reported as $S = 14$5 per slot, consistent with SF-wise linear throughput.

These results define the empirical identity of DeepTx more clearly than channel-prediction accuracy alone would. The method is not merely a temporal extrapolator of CSI; it is a learned front-end for a specific linear precoding pipeline whose success is measured at the receiver after equalization.

6. Scope, limitations, and relation to classical beamforming

DeepTx is trained on TDL-A, TDL-B, and TDL-C channels and velocities up to 30 km/h. The reported limitations are explicit: extreme mobility and hardware non-idealities, including non-reciprocity, remain future work (Huttunen et al., 2022). The paper also notes that larger antenna arrays such as $S = 14$6 and massive MIMO will require deeper or wider networks, although SF-wise throughput remains linear in $S = 14$7.

The method is best understood as a hybrid design. It preserves conventional system structure—pilot-based uplink estimation, explicit CSI tensor formation, and ZF beamforming—while replacing the reciprocity-and-aging heuristic with a supervised predictor optimized through UE BER. This makes it distinct from pure model-based channel prediction and from fully end-to-end learned transceivers. The summary’s final characterization is precise: DeepTx robustly harvests the aspects of uplink CSI that matter for downlink beamforming under realistic 5G channel dynamics.

A common misconception would be to treat the method as proving that more accurate channel prediction in the MSE sense is sufficient for better beamforming. The paper directly reports the opposite outcome for one comparison: a CNN trained to minimize channel MSE yields worse BER than outdated CSI. The stronger claim supported by the reported evidence is narrower and more technical: in this setting, beamforming-aware end-to-end supervision is better aligned with downlink performance than channel-MSE supervision.

Within that scope, DeepTx constitutes a transmitter-side neural augmentation of TDD reciprocity-based precoding. Its significance lies in showing that channel aging, beamformer computation, and approximation error compensation can be learned jointly without abandoning the classical OFDM and ZF structure in which modern wireless systems are commonly analyzed.