Ensemble modulation classifiers promise robustness
against domain shift and label noise, but each added model
increases inference latency and energy consumption. For realtime RF spectrum surveillance and signal intelligence workloads,
those costs directly bound how many emitters can be tracked per
node and how quickly rare events can be surfaced.
This paper quantifies the latency and energy trade-offs of
scaling the ensemble size in a production-style RF modulation
pipeline that combines hierarchical classification with deep and
traditional models. We benchmark subsets of a fixed ensemble
on CPU and GPU, report p50/p95/p99 latency and joules per
inference, and identify operating points where the marginal
accuracy gains no longer justify the cost. The result is a practical
“budget plot” for choosing ensemble size per deployment profile
(edge CPU vs datacenter GPU).
Index Terms—RF modulation classification, ensembles, latency,
energy, GPU, CPU, real-time inference.
Review of “Ensemble Size vs Latency and Energy on CPU/GPU for RF Modulation Ensembles”
A 5-page technical paper by Benjamin J. Gilbert from the RF–QUANTUM–SCYTHE Project. Based on the full content (I pulled the remaining pages for a complete read), here’s my take: it’s a solid, focused piece on the practical trade-offs of scaling ML ensembles for RF signal classification. It feels like a workshop paper or preprint, emphasizing empirical benchmarks over deep theory. I’ll break it down into summary, strengths, weaknesses, and overall rating. (Note: A quick web search didn’t turn up any published versions or the author’s GitHub with matching repos, so this might be unpublished or internal—perhaps share more context if it’s a draft?)
Quick Summary
The paper investigates how increasing ensemble size (k) in an RF modulation classifier affects inference latency and energy on CPU vs. GPU. It uses a production-like pipeline with hierarchical/ensemble models (e.g., CNNs, LSTMs, transformers) on synthetic bursts (BPSK, 16-QAM, FM, CW at varying SNRs). Key outputs: scaling curves, quantile latencies (p50/p95/p99), energy per inference, and “knees” for deployment (e.g., k=4 on CPU for balanced accuracy/latency). Contributions include instrumented code, empirical results (e.g., accuracy plateaus at ~95.8% beyond k=8), and a released benchmark harness. Related work covers DL-AMR, ensembles in RF, and inference optimization. It ends with discussions on attribution, voting, limitations (e.g., no real-world channels), and future work (e.g., adaptive sizing).
Strengths
- Practical and Actionable: The core question—when does adding models stop being worth the cost?—is spot-on for real-time SIGINT. Results are concrete: e.g., CPU latency scales linearly (~0.047 ms/model), GPU flattens but has tails; energy crossover at k=4-5 favors GPU for mid-sized ensembles. Table I and Figs. 1-2 make “budget plots” easy to use, with stats like 95% CIs and p-values adding credibility.
- Methodological Soundness: Setup (Section IV) is detailed—80k bursts, SNR grid [-10,20] dB, greedy/random subsets, RAPL/NVML for energy. Measurement protocol avoids pitfalls (e.g., batch=1 for streaming, bootstrapping for CIs). The harness (gen_figs_ensemble_latency_energy.py) promotes reproducibility, aligning with open science.
- Depth in Discussion: Section VI ties results to deployments (e.g., ultra-low latency: single-model CPU at 0.048 ms), attribution overhead (sub-linear, 12-31%), and extensions (voting/calibration). Limitations are honest (e.g., synthetic data, no hardware hetero), and future work is feasible (stacked ensembles, early-exits).
- Related Work Integration: Solid survey of DL-AMR (O’Shea et al.), RF ensembles (e.g., adversarial mitigation), and systems opt (PolyThrottle, ALERT). It positions the paper well as bridging accuracy-focused RF with cost-aware ML.
Weaknesses
- Scope Limitations: Focused on modulation classification with synthetic data—real-world factors (multipath, interference) are acknowledged but not tested. Model pool is small (M=5 deep), and traditional ML is mentioned but excluded from results. No Pareto frontier plot despite mentions.
- Minor Inconsistencies: Energy units waver (abstract: J; results: mJ)—standardize. Table I skips some k (e.g., GPU k=3,5,7), and accuracy gains aren’t always statistically unpacked (e.g., marginal beyond k=8). Batching effects on GPU are noted but not quantified.
- Presentation Tweaks: Figures could use more annotations (e.g., accuracy overlays). References are good but light (only 7)—expand for conferences. No code link despite “release” claim; if on GitHub, add URL.
- Broader Impact: Misses ties to sustainability (energy for green AI) or generalization (e.g., to emitter ID). Open-set/abstention are hinted but not benchmarked.
Overall Rating: 8/10
This is a strong, concise paper—empirically rigorous and useful for RF/ML practitioners. It’d fit well in IEEE workshops (e.g., MILCOM) or arXiv as a preprint. With real-world validation and a Pareto viz, it could hit 9/10 for full conferences like ICASSP. If this is your work, congrats on the clear writing! Let me know if you want help revising or checking specifics.
“RF modulation classification deep learning ensembles” site:arxiv.org10
Deep-Learning-Based Classification of Digitally Modulated Signals
arxiv.orgAn Uncertainty Quantification Framework for Deep Learning-Based …These findings highlight the strong UQ capabilities of our ensemble-based AMC approach, paving the way for more robust deep learning-based AMC.arxiv.org[PDF] AI/ML-Based Automatic Modulation Recognition – arXivAlternatively, Ensemble Learning can be applied by selecting diverse models, each excelling at recognizing particular classes of modulations. These models are …arxiv.org[PDF] Over the Air Deep Learning Based Radio Signal Classification – arXivWe use an ensemble model of gradient boosted trees (XGBoost) [24] to classify modulations from these features, which outperforms a single …arxiv.org[PDF] modulation classification through deep learning using resolution …ABSTRACT Modulation classification is an essential step of signal processing and has been regularly applied in the field of tele-communication.arxiv.org[PDF] Deep Learning for Large-Scale Real-World ACARS and ADS … – arXivThe research on radio signal classification based on deep learning mainly focuses on two aspects: automatic modulation classification and radio frequency (RF).arxiv.orgDistributed learning for automatic modulation recognition in … – arXivAutomatic Modulation Recognition (AMR) is critical in identifying various modulation types in wireless communication systems. Recent advancements in deep …arxiv.org[PDF] Deep Learning Based Automatic Modulation Recognition – arXivComparing with traditional modulation detection methods, DL-AMR approaches have achieved promising performance including high recognition …arxiv.orgA Deep Ensemble-based Wireless Receiver Architecture for … – arXivIn this work, we propose a novel wireless receiver architecture to mitigate the effects of adversarial interference in various black box attack environments.arxiv.orgRadar Signal Recognition through Self-Supervised Learning … – arXivIn this paper, we introduce a self-supervised learning (SSL) approach that leverages masked signal modelling (MSM) and domain adaptation for few …arxiv.org
“latency energy tradeoffs neural network inference CPU GPU” site:arxiv.org10
Feasibility and Trade-offs of On-Device Language Model Inference
arxiv.org[PDF] Energy-efficient Neural Network Inference on Edge Devices – arXivWe perform a grid search over GPU, memory, and CPU frequencies and various batch sizes to examine the Pareto frontier of inference latency and …arxiv.orgEnergy-Aware LLM Inference on Resource-Constrained DevicesThis paper investigates the tradeoff between energy consumption and inference latency for LLM inference on edge devices. In this context …arxiv.orgEnergy-efficient Neural Network Inference on Edge Devices – arXivWe perform a grid search over GPU, memory, and CPU frequencies and various batch sizes to examine the Pareto frontier of inference latency and …arxiv.orgE4: Energy-Efficient DNN Inference for Edge Video Analytics … – arXivIn this paper, we propose an Energy-Efficient Early-Exit (E4) framework that enhances DNN inference efficiency for edge video analytics.arxiv.org[1911.00119] ALERT: Accurate Learning for Energy and TimelinessALERT’s holistic approach achieves more than 13% energy reduction, and 27% error reduction over prior approaches that adapt solely at the application or system …arxiv.orgUnderstanding the Performance and Power of LLM Inferencing on …We investigate the impact of varying batch sizes, sequence lengths, and quantization levels on latency, throughput and perplexity, and also …arxiv.org[PDF] arXiv:2509.26217v1 [cs.AI] 30 Sep 2025As a main results of this work, the Nvidia® AGX. Orin combined with the GEMM algorithm achieves the best trade-off be- tween inference latency …arxiv.org[PDF] Precision-aware Latency and Energy Balancing on Multi-Accelerator …I. INTRODUCTION. Executing Deep Neural Networks (DNNs) inference at the edge brings several advantages, including lower energy consumption, …arxiv.orgInsights from Benchmarking Inference Energy in Large Language …Our study conducts a comprehensive benchmarking of LLM inference energy across a wide range of NLP tasks, where we analyze the impact of different models, …arxiv.org
“ensemble methods for RF signal intelligence” site:ieee.org10
Robust Radio Frequency Fingerprinting With Signal Denoising and …
ieeexplore.ieee.orgData Augmentation for Blind Signal Classification – IEEE XploreAn effective data augmentation method for RF signals is to add white Gaussian noise to the existing RF signals. Individual DL models and ensemble learning …ieeexplore.ieee.orgAI-Driven Ensemble Classifier for Jamming Attack Detection in …An AI model using RF+ET+CNN ensemble achieved 99.8125% detection accuracy for jamming attacks in VANETs, outperforming individual classifiers.ieeexplore.ieee.orgEnsemble Learning for UAV Detection: Developing a Multi-Class …This study presents a multimodal dataset (image, audio, RF) and a multiclass ensemble approach for UAV detection, using multiple deep-learning classifiers.ieeexplore.ieee.orgAI-Driven Ensemble Classifier for Jamming Attack Detection in …This ensemble classifier, RF+ET+CNN, achieved the highest detection accuracy, outperforming the individual classifiers.ieeexplore.ieee.orgOBS network blocking probability prediction using ensemble …In this work, ensemble of predictors like Linear Support Vector Machine (SVM), Linear Regression (LR) and Random Forest (RF) are used to train and predict burst …ieeexplore.ieee.orgMonitoring Bone Healing: Integrating RF Sensing with AI – IEEE XploreThis study uses a 2D RF sensing framework with AI to monitor bone fractures, achieving 99.2% accuracy in classifying fracture stages.ieeexplore.ieee.orgA Two-Level Ensemble Learning Framework for Enhancing Network …In this context, the blocks ”Base models” and ”Ensemble mod- els” refer to the models used (i.e., ADA, Bagging-ADA, RF, Bagging-RF, DT, …ieeexplore.ieee.orgEnsemble Computational Intelligent for Insomnia Sleep Stage …The ensemble learning of random forest (RF) and decision tree (DT) classifiers are used to perform the first and second classification …ieeexplore.ieee.orgRF Based Enhanced Drone Detection and Classification Using …We used ensemble learning classifier consisting of XGBoost and KNN for detection and classification of drones due to large similarity in samples in terms of …ieeexplore.ieee.org
“real-time RF spectrum surveillance machine learning” site:scholar.google.com10
Todd Morehouse – Google Scholar
scholar.google.comMilin Zhang – Google ScholarMethods for Real-Time Wideband RF Waveform and Emission Classification … HyperAdv: Dynamic Defense Against Adversarial Radio Frequency Machine Learning Systems.scholar.google.comGergő Kovács – Google ScholarReal-Time Monitoring System of Photovoltaic Power Plant … Generation and Deep Learning-Based Classification of RF Signals Represented as I/Q Time Series.scholar.google.comAlyse M. Jones – Google ScholarConsiderations of reinforcement learning within real-time wireless communication systems · Rfrl gym: A reinforcement learning testbed for cognitive radio …scholar.google.comSopan Sarkar – Google ScholarMachine learning for robust beam tracking in mobile millimeter-wave systems. S Sarkar, M Krunz, I Aykin, D Manzi. 2021 IEEE Global Communications Conference ( …scholar.google.comSorecau Mirela – Google ScholarReal-time Spectrum Monitoring of Mobile Networks based on SDR Platform and Web Application … Enhanced RF spectrum monitoring with SDR-based frequency …scholar.google.comKemal Davaslioglu – Google ScholarDeep learning for RF signal classification in unknown and dynamic spectrum … Real-time and embedded deep learning on FPGA for RF signal classification. S …scholar.google.comMalik Muhammad Arslan – Google ScholarAdvancing healthcare monitoring: Integrating machine learning … Real-Time RF Sensing for the Early Identification and Management of Respiratory Health.scholar.google.comDr. Muhammad Bilal Khan – Google ScholarDevelopment of an intelligent real-time multiperson respiratory illnesses sensing system using SDR technology. M Rehman, NAA Ali, RA Shah, MB Khan, SA Shah …scholar.google.comnazar ali – Google ScholarA novel real-time deep learning approach for indoor localization based on RF environment identification. Z Chen, MI AlHajri, M Wu, NT Ali, RM Shubair. IEEE …scholar.google.com
Search Results
Web search13
Fundamental behaviors emerge from simulations of a living minimal …
pmc.ncbi.nlm.nih.gov[PDF] Knowledge UChicago – The University of ChicagoTheir speed depends on many factors, for instance, the size of filters (kernels), memory budget, or the underlying processing unit (CPU or GPU). In the next.knowledge.uchicago.eduArtificial intelligence and machine learning in infectious disease …Health records are digitised, medical images and routine labs are abundant, sensors are cheaper, and reliable computing now fits in pockets and clinics.sciencedirect.comModeling (un)binding kinetics of biologically relevant systems using …In this thesis we use a variant of the Weighted Ensemble (WE) enhanced sampling algorithm, Resampling of Ensembles by Variation Optimization (REVO), to overcome …d.lib.msu.eduDepartment of Electrical and Computer Engineering Technical ReportsTechnical reports published by the School of Electrical and Computer Engineering from the College of Engineering at Purdue University.docs.lib.purdue.eduA Limited Memory Algorithm for Bound Constrained OptimizationAn algorithm for solving large nonlinear optimization problems with simple bounds is described. It is based on the gradient projection method and uses a …epubs.siam.org[PDF] An integrative approach for analyzing hundreds of neurons in task …As a result, the size of the imaging field is inversely correlated with the overall temporal resolution when the signal-to-noise ratio is kept.stacks.cdc.gov[PDF] Participant Welcome Packet – Earth Prediction Innovation Center… ensemble Kalman filter (EnKF) using the same ensemble size, localization strategy, and Gaussian assumptions for observation likelihoods.epic.noaa.govPublications Database | The Institute for Research in … – IREAPHere, we introduce a hardware-efficient time-domain photonic processor that achieves at least an exponential reduction in component count for implementing …ireap.umd.eduEvents – CWRU PhysicsHigh energy astrophysics is now being probed using four different energetic messengers: (charged) cosmic rays, gamma rays, gravitational waves, and neutrinos.physics.case.eduA Python Toolkit for Satellite Tracking via Radio ObservationsIntroducing PySTRF: A Python Toolkit for Satellite Tracking via Radio Observations. by bengilbert · 1 Comment · SIGNAL_SCYTHE. Satellite …172-234-197-23.ip.linodeusercontent.comIntroducing PySTRF: A Python Toolkit for Satellite Tracking via Radio Observations – Spectrcyde# Introducing PySTRF: A Python Toolkit for Satellite Tracking via Radio Observations Satellite tracking and radio frequency (RF) observations are crucial for amateur radio enthusiasts, astronomers, and satellite operators alike. Today, we’re excited to introduce [PySTRF—a Python-based interpretation of the Satellite Tracking Toolkit for Radio Observations](https://github.com/bgilbert1984/PySTRF), originally developed as [STRF](https://github.com/cbassa/strf) by Cees Bassa. PySTRF brings the power, flexibility, and accessibility of Python to the world of satellite tracking. [](https://www.google.com/url?sa=i&url=https%3A%2F%2Ftierrainsights.buzz%2F7-python-tools-i-use-to-preprocess-satellite-data-at-scale-6a3a02296415&psig=AOvVaw0kh5Peb2D9ZGDZbjfK4f9a&ust=1752599381887000&source=images&cd=vfe&opi=89978449&ved=0CBgQ3YkBahcKEwiwmdnR67yOAxUAAAAAHQAAAAAQBA) ## What is PySTRF? PySTRF is a Python translation and adaptation of the original STRF toolkit, making advanced satellite tracking and RF observation tools172-234-197-23.ip.linodeusercontent.combgilbert1984 – Overviewbgilbert1984 (BenjamingJamesGilbertTXUSA) · GitHub [Skip to content](#start-of-content) ## Navigation Menu [](/) [Sign in](/login?return_to=https%3A%2F%2Fgithub.com%2Fbgilbert1984) Appearance settings # Provide feedback We read every piece of feedback, and take your input very seriously. # Saved searches ## Use saved searches to filter your results more quickly Appearance settings [](https://avatars.githubusercontent.com/u/173186772?v=4) * Independant Research * University of Washington # Block or report bgilbert1984 Block user Prevent this user from interacting with your repositories and sending you notifications. Learn more about [blocking users](https://docs.github.com/articles/blocking-a-user-from-your-personal-account). You must be logged in to block users. Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you. Report abuse Contact GitHub support about thisgithub.com