Subject: Cutting-Edge Advancements in Signal Processing, Communication, and Biomedical Applications

Hi Elman,

This newsletter covers recent preprints exploring advancements in signal processing, communication systems, and biomedical applications, with a focus on machine learning and novel hardware architectures.

General Overview

This collection of preprints showcases significant progress in several key areas. A notable trend is the optimization of communication in challenging spectral environments like millimeter wave and terahertz bands. Wang et al. (2024) Wang et al. (2024) introduce double-side delay alignment modulation (DAM) for multi-user scenarios, demonstrating that single-side DAM is often sufficient with large antenna arrays, offering advantages over OFDM in terms of spectral efficiency and peak-to-average power ratio (PAPR). Ning et al. (2024) Ning et al. (2024) propose a wideband beamforming codebook design method for terahertz communications, addressing the beam squint effect and optimizing for wideband beam gain. Bai & Larsson (2024) Bai & Larsson (2024) tackle user activity detection in massive access scenarios, moving beyond the block-fading paradigm with a low-dimensional channel approximation framework.

Another prominent theme is the development of novel hardware and algorithms for enhanced sensing and communication. Malmberg & Feyling (2024) Malmberg & Feyling (2024) present a simplified continuous-time RC-chain ADC design boasting robustness against component variations. Ray et al. (2024) Ray et al. (2024) introduce a low-power integrating sub-sampling PLL for communication and computing applications, achieving a wide operating frequency range with minimal area and power consumption. Määttä et al. (2024) Määttä et al. (2024) leverage transformer networks to generate 3D point clouds from WiFi-CSI data, showcasing the potential of joint communication and sensing (JC&S) for spatial awareness. Wang et al. (2024) Wang et al. (2024) present Hawk, an efficient non-autoregressive language modeling (NALM) system for appliance recognition, emphasizing efficient dataset construction and accurate event recognition.

Machine learning continues to drive innovation in complex signal analysis and biomedical applications. Lia et al. (2024) Lia et al. (2024) demonstrate real-time epilepsy detection using a spiking neural network (SNN) deployed on a low-power neuromorphic processor. Li & Ruiz (2024) Li & Ruiz (2024) propose a graph sampling algorithm for scalable graph neural networks (GNNs) on homophilic graphs, improving model transferability. Das et al. (2024) Das et al. (2024) introduce Parsimonious Dynamic Mode Decomposition (parsDMD), an automated approach for sparse mode selection in complex systems, offering advantages over traditional spDMD. Bari & Sen (2024) Bari & Sen (2024) develop a computational harmonic detection algorithm for identifying data leakage through EM emanation. Douglas et al. (2024) Douglas et al. (2024) characterize the effects of electrode shift on sEMG features, aiming to improve the reliability of stroke recovery assessments.

Further contributions explore advanced signal processing techniques. Nielsen et al. (2024) Nielsen et al. (2024) investigate blind equalization using a variational autoencoder (VAE) with a second-order Volterra channel model. Konstantinou et al. (2024) Konstantinou et al. (2024) employ differential evolution for optimizing the realized gain of active and parasitic antenna arrays. Haghshenas & Naeini (2024) Haghshenas & Naeini (2024) propose a resilient temporal GCN for smart grid state estimation under topology inaccuracies. Gadamsetty et al. (2024) Gadamsetty et al. (2024) investigate sum-rate maximization in reconfigurable intelligent surface (RIS)-aided MU-MISO systems with digital and holographic beamformers. Yuan et al. (2024) Yuan et al. (2024) analyze the performance of machine learning-based error correcting decoders.

Finally, the collection touches upon quantum communication, biomedical signal processing, and integrated sensing and communication (ISAC). Zhang et al. (2024) Zhang et al. (2024) propose a security-enhanced quantum communication scheme for space-air-ground integrated networks. Hughes-Noehrer et al. (2024) Hughes-Noehrer et al. (2024) investigate clinician attitudes towards AI in ECG interpretation. Hammami et al. (2024) Hammami et al. (2024) introduce an end-to-end model-based learning approach for efficient FSS analysis. Wang et al. (2024) Wang et al. (2024) explore high-order associative learning with memristive circuits. Several papers address ISAC design challenges, including robust precoder design (Wu et al., 2024), decentralized hybrid precoding (Zhu et al., 2024), and joint space-time adaptive processing and beamforming (Liu et al., 2024).

Paper Highlights

AI Learns Physics Without Knowing the Rules

A Phenomenological AI Foundation Model for Physical Signals by Jaime Lien, Laura I. Galindez Olascoaga, Hasan Dogan, Nicholas Gillian, Brandon Barbello, Leonardo Giusti, Ivan Poupyrev https://arxiv.org/abs/2410.14724

Caption: The Ω-model processes sensor measurements (m) from various sources in the physical world (Ω) using a transformer-based architecture. During training, measurements from different sensors are encoded into a latent representation (z^emb), which is then used by decoders (g- and g+) to reconstruct past and predict future system behavior. This allows the model to generalize across diverse physical phenomena without explicit knowledge of underlying physical laws.

Archetype AI researchers have developed a groundbreaking AI foundation model, the Ω-model, capable of generalizing across diverse physical phenomena without prior knowledge of physical laws. Trained on a massive dataset of 0.59 billion cross-modal sensor measurements, the model learns the underlying structure of physical processes directly from the data.

The Ω-model employs a phenomenological framework, relying solely on observed data without incorporating explicit physical equations. Sensor data is normalized and segmented into fixed-length patches, then projected into a unified embedding space using a standard transformer architecture. The transformer's attention mechanism captures long-range temporal dependencies, creating a compressed, universal representation of the physical process. Lightweight, task-specific decoders extract information from this representation, reconstructing past trajectories (g−: z^emb → m_i,j(t), for t ≤ T) and predicting future behavior (g+: z^emb → m_i,j(t), for t > T).

Rigorous testing on unseen physical systems, including a spring-mass system and a thermoelectric demonstrator, showcased the Ω-model's zero-shot capabilities. It accurately predicted system behavior, even capturing complex phenomena like transitions from chaotic to harmonic oscillation. Further tests on real-world datasets, including electricity consumption and meteorological data, demonstrated generalization to complex, non-analytic systems. Impressively, the zero-shot Ω-model often outperformed models specifically trained on the target data, achieving a 23% lower mean squared error (MSE) in forecasting and a 34% lower MSE in reconstruction.

New Algorithm Unlocks Full Potential of High-Density Surface EMG

Unlocking the Full Potential of High-Density Surface EMG: Novel Non-Invasive High-Yield Motor Unit Decomposition by Agnese Grison, Irene Mendez Guerra, Alexander Kenneth Clarke, Silvia Muceli, Jaime Ibanez Pereda, Dario Farina https://arxiv.org/abs/2410.14800

Caption: Figure: Comparison of Swarm-Contrastive Decomposition (SCD) and conventional blind source separation (cBSS) in HD-sEMG. (a) SCD extracts significantly more motor units (MUs) across various force levels. (b-g) Analysis of common and unique MUs reveals SCD's superior performance in identifying more MUs, particularly those with lower amplitudes, faster conduction velocities, and originating from deeper muscle tissues.

High-density surface electromyography (HD-sEMG) offers a non-invasive way to study motor unit activity. However, existing decomposition algorithms struggle to isolate individual motor units in challenging conditions. This paper introduces Swarm-Contrastive Decomposition (SCD), a novel algorithm designed to overcome these limitations.

Traditional blind source separation (BSS) methods rely on fixed contrast functions. SCD dynamically adapts its contrast function, G(s) = E{sign(s)|s|^e}, for each source using particle swarm optimization. This adaptability, coupled with a peel-off strategy for sequential source removal, allows SCD to differentiate between similar motor unit action potentials (MUAPs) and identify low-amplitude motor units often missed by other methods.

SCD's performance was evaluated using simulated and experimental HD-sEMG data across various conditions. In simulations, SCD significantly outperformed a state-of-the-art cBSS method, detecting nearly double the number of motor units across varying excitation and noise levels. In ballistic contraction simulations, SCD identified approximately three times more motor units than cBSS. Experimental validation using a two-source approach, comparing HD-sEMG and HD-iEMG recordings, confirmed SCD's superior performance in identifying more motor units and achieving higher accuracy in determining motor unit discharge timings.

Memristive SoC Enables Multi-Disease Detection with Generative AI Boost

Multi-diseases detection with memristive system on chip by Zihan Wang, Daniel W. Yang, Zerui Liu, Evan Yan, Heming Sun, Ning Ge, Miao Hu, Wei Wu https://arxiv.org/abs/2410.14882

This study presents a novel system for simultaneous detection of Acute Myocardial Infarction (AMI) and liver cancer using a memristor/CMOS system-on-chip (SoC) and generative AI. This addresses limitations of traditional blood tests and previous memristor-based disease detection systems by offering a fully integrated, multi-disease diagnostic platform.

The system integrates a nanofinger platform for enhanced Raman signal capture, a conditional diffusion model for data augmentation, and a deep neural network (DNN) classifier implemented on the memristive SoC. The nanofinger platform maximizes electromagnetic field enhancement, enabling highly selective biomarker capture. The conditional diffusion model generates high-fidelity Raman signals, improving the dataset's robustness and diversity. The model's loss function is given by: L_CDM = E_t,x0,ε,c[||ε - ε_θ(x_t, t, c)||²], where ε is the added noise, ε_θ(x_t, t, c) is the U-Net's noise prediction, and c is the conditioning information.

A deeper multilayer perceptron (MLP) classifier with integrated ReLU activation functions is trained on the augmented dataset using a loss function combining cross-entropy loss and L2 regularization: L = L_{cross-entropy} + λ||W||². The trained network is transferred to the TetraMem® MX100 SoC, which accelerates inference through in-memory computing.

Experimental results demonstrate the system's effectiveness. Data augmentation significantly improved the DNN classifier's average accuracy from 87.54% to 91.82%. The SoC implementation achieved comparable accuracy to the software baseline, maintaining high accuracy for critical disease predictions.

Conclusion

This newsletter highlights a convergence of advancements across signal processing, communication, and biomedical applications. The development of the Ω-model demonstrates the potential of AI to learn complex physical phenomena directly from data, potentially revolutionizing sensor intelligence and scientific discovery. The introduction of SCD significantly enhances the capabilities of HD-sEMG, opening new avenues for studying motor unit activity. Finally, the memristive SoC for multi-disease detection showcases the power of integrating advanced hardware and generative AI for efficient and accurate medical diagnostics. These breakthroughs collectively pave the way for transformative applications across various domains.

Best,

[Your Name/Newsletter Name]