Subject: Cutting-Edge Advances in Signal Processing and Machine Learning
Hi Elman,
This collection of preprints explores diverse applications of signal processing and machine learning across wireless communications, sensing, and biomedical domains. Several papers focus on enhancing wireless system performance through novel architectures and algorithms. AlZubi and Alouini (2025) AlZubi & Alouini (2025) investigate the potential of TV White Spaces for narrowband IoT applications, demonstrating throughput of up to 97 Kbps in outdoor experiments using an SDR testbed. Tang et al. (2025) Tang et al. (2025) propose a two-stage channel estimation scheme for XL-MIMO systems with decentralized baseband processing, combining local sparse reconstruction with global refinement. For wideband terahertz MIMO-OFDM systems, Li and Vu (2025) Li & Vu (2025) introduce a GNN-based hybrid beamforming design to mitigate beam squint, achieving performance comparable to all-digital beamforming. Addressing interference management, Jiang et al. (2025) Jiang et al. (2025) propose federated learning strategies for coordinated beamforming in multi-cell ISAC, while Shi et al. (2025) Shi et al. (2025) leverage RIS technology for robust over-the-air federated learning.
Another prominent theme is the application of deep learning. Verma et al. (2025) Verma et al. (2025) explore the reconstruction of Pulsed-wave Doppler signals from non-invasive fetal ECG, while Chintapenta et al. (2025) Chintapenta et al. (2025) propose MFConvTr, a multi-frequency convolutional transformer, for fetal arrhythmia detection. For indoor radio map prediction, Li et al. (2025) Li et al. (2025) introduce TransPathNet, a two-stage framework combining transformer-based feature extraction and multiscale convolutional attention decoding. In autonomous vehicles, Wang et al. (2025) Wang et al. (2025) present a modular framework for uncertainty prediction in motion forecasting, and Shanto et al. (2025) Shanto et al. (2025) employ supervised contrastive learning for robust deepfake detection. Mishra et al. (2025) Mishra et al. (2025) utilize graph convolutional variational autoencoders for subject representation learning from EEG data.
Several contributions focus on specific signal processing techniques. Xu et al. (2025) Xu et al. (2025) propose a two-stage sparse DOA estimation framework for incoherently-distributed sources with gain-phase uncertainty. Kurisaki (2025) Kurisaki (2025) presents a new proof for linear filtering and smoothing equations and explores the asymptotic expansion of nonlinear filtering. Castelli Lacunza and Sing Long (2025) Castelli Lacunza & Sing Long (2025) investigate adaptive multipliers for extrapolation in frequency, establishing connections with multiresolution analysis. For RF signal classification, Plouet et al. (2025) Plouet et al. (2025) demonstrate the use of simple spintronic devices, metallic spin-diodes, to perform convolutions. Nomeir et al. (2025) Nomeir et al. (2025) determine the asymptotic capacity of Byzantine symmetric private information retrieval.
Focusing on practical implementations, Li and Guo (2025) Li & Guo (2025) utilize deep learning with Wi-Fi probe requests for vehicle occupancy estimation in automated guideway transit. Meier and Holz (2025) Meier & Holz (2025) introduce BMAR, a barometric and motion-based alignment and refinement method for offline signal synchronization across devices. Ardizzon et al. (2025) Ardizzon et al. (2025) assess the performance of crystal oscillators in OSNMA-enabled receivers for automotive applications. Wu et al. (2025) Wu et al. (2025) propose efficient inventory solutions for 3GPP ambient IoT considering device unavailability due to energy harvesting. Finally, Wang et al. (2025) Wang et al. (2025) fuse millimeter-wave radar and pulse oximeter data for low-burden diagnosis of obstructive sleep apnea.
These preprints highlight the integration of advanced signal processing, machine learning, and hardware design. The research spans theoretical advancements to practical system implementations. The emphasis on deep learning, GNNs, federated learning, and RIS technology underscores their growing importance.
Meta-Federated Learning: A Novel Approach for Real-Time Traffic Flow Management by Bob Johnson, Michael Geller https://arxiv.org/abs/2501.16758
Caption: This diagram illustrates the architecture of a Meta-Federated Learning system for traffic management. It shows the process of client selection, local training, global aggregation, and personalized model distribution across a network of IoT devices. The system utilizes double deep Q-learning for client selection and incorporates knowledge distillation and model interpolation for enhanced performance.
Urban traffic management requires systems that can adapt to dynamic conditions while preserving privacy. Traditional centralized systems struggle with scalability and privacy, while standard Federated Learning (FL) lacks adaptability. This paper introduces Meta-Federated Learning, combining decentralized FL with the adaptability of Meta-Learning (ML).
The system uses a network of IoT devices collecting local traffic data (X<sub>i,t</sub> = {X<sub>1,t</sub>, X<sub>2,t</sub>,..., X<sub>n,t</sub>}). These devices perform initial processing and training. The system minimizes a global loss function: min f(θ) = Σ<sup>K</sup><sub>k=1</sub> p<sub>k</sub>F<sub>k</sub>(θ), where θ are global parameters, K is the number of nodes, p<sub>k</sub> is the node weight, and F<sub>k</sub>(θ) is the local loss. Model-Agnostic Meta-Learning (MAML) enables adaptation to new traffic conditions with minimal training.
Evaluation in a simulated urban traffic network using SUMO shows Meta-Federated Learning outperforms centralized ML and standard FL in accuracy, response time, throughput, and latency. Accuracy improves by 5% to 7% in various traffic scenarios, with response time improvements of 0.8s to 1.2s in high traffic. Throughput increases to 1300 vehicles/hour, and latency reduces to 0.45s. This superior performance stems from decentralized processing for privacy and MAML's adaptability. Future work includes applying this to other smart city applications and exploring advanced meta-learning algorithms.
Brain-Inspired Decentralized Satellite Learning in Space Computing Power Networks by Peng Yang, Ting Wang, Haibin Cai, Yuanming Shi, Chunxiao Jiang, Linling Kuang https://arxiv.org/abs/2501.15995
Caption: This figure illustrates the proposed decentralized learning framework for Space Computing Power Networks (Space-CPNs). Satellites within and across orbital planes (intra/inter-plane) exchange model updates using the RelaySum algorithm for efficient aggregation. The framework utilizes spiking neural networks (SNNs) for local updates, leveraging their energy efficiency for on-board processing.
Satellite networks are vital for real-time Earth observation, but centralized processing creates bottlenecks. Space Computing Power Networks (Space-CPNs) enable on-board processing, but energy constraints pose a challenge. This paper proposes a framework using spiking neural networks (SNNs) and decentralized learning.
SNNs, mimicking the brain, offer high energy efficiency due to sparse computation. A hybrid activation function, combining a surrogate differentiable function and the Heaviside function, allows for effective backpropagation during training. The RelaySum algorithm enables efficient model update distribution across orbit planes.
Theoretical analysis shows sublinear convergence, with speed linked to the inter-plane topology diameter (ĩ). A minimum diameter spanning tree problem is formulated, considering line-of-sight distance and Doppler shift, with edge weights based on average SNR (ξi,j): w(ei,j) = 1/ξi,j + 1/ξ.
Experiments on EuroSAT show RelaySum outperforms other aggregation methods, achieving higher accuracy. SNNs significantly reduce energy consumption compared to ANNs while maintaining performance. Optimizing the routing tree for minimum diameter further accelerates convergence. This highlights the potential of this framework for efficient on-board processing in Space-CPNs.
Combating Interference for Over-the-Air Federated Learning: A Statistical Approach via RIS by Wei Shi, Jiacheng Yao, Wei Xu, Jindan Xu, Xiaohu You, Yonina C. Eldar, Chunming Zhao https://arxiv.org/abs/2501.16081
Caption: This figure illustrates a RIS-aided over-the-air federated learning (AirFL) system. Target devices and unknown interference sources transmit their local gradients (g<sub>t,k</sub>) through an RIS to the parameter server (PS). The RIS manipulates the wireless channels (h<sub>r,k</sub> and h<sub>p</sub>) to enhance the desired signals and suppress interference, enabling robust global model aggregation at the PS.
Over-the-air computation (AirComp) enhances federated learning (FL) efficiency, but its analog nature makes it susceptible to interference. This paper uses reconfigurable intelligent surfaces (RIS) to improve AirFL robustness.
The key is phase-manipulated favorable propagation and channel hardening. Adjusting RIS phase shifts (θ<sub>n</sub> = -∠h<sub>p,n</sub> + Σ<sub>k∈K</sub> w<sub>k</sub>h<sub>r,k,n</sub>) statistically eliminates interference and reduces gradient estimation variance.
Two robust aggregation schemes are proposed. Scheme I uses power control (P<sub>k</sub> = β<sub>k</sub><sup>-1</sup>ζ). Scheme II uses maximum power transmission and adjusts weight factors (w<sub>k</sub>). Both ensure unbiased gradient estimation.
Analysis shows MSE decreases by O(1/N), where N is the number of RIS elements. Convergence rate is O(1/T<sup>ω<sub>u</sub></sup>). As N increases, the system approaches the ideal convergence rate without interference. Numerical results validate these findings, showing Scheme I excels in computation-dominant systems, while Scheme II performs better in interference-dominant environments.
This newsletter highlights the convergence of advanced signal processing, machine learning, and practical system design. The selected papers demonstrate innovative approaches to address critical challenges in diverse domains. From managing urban traffic flow with meta-federated learning to enabling efficient on-board processing in space computing networks using SNNs and decentralized learning, these works push the boundaries of what's possible. Furthermore, the innovative use of RIS technology for robust over-the-air federated learning showcases the potential of intelligent surfaces to transform wireless communication in interference-prone environments. These advancements collectively pave the way for more intelligent, efficient, and robust systems across various applications.