edge computing

Edge AI System-of-Systems Reference Architecture Engineering Foundations and Multi-Dimensional Views

Edge AI systems are emerging from the convergence of IoT, edge computing, AI, agentic AI, and embodied, physical generative edge AI delivering adaptive, autonomous behaviour under physical, cyber, and operational constraints while remaining trustworthy. This article frames edge AI as a complex system-of-systems in which hardware, software, models, and data continuously co-evolve across heterogeneous “multi-X” environments: multiple systems, modalities, and agents distributed from the edge to the cloud. The article argues that as edge AI technologies are maturing, there is a need for a standardised, application-agnostic reference architecture to provide a shared lexicon and taxonomy, reduce integration errors, and expose opportunities for reusable assets and productive interoperability and standardisation. The paper grounds this need in systems engineering and introduces a quad-optimisation paradigm for balancing competing objectives during design and operation. The article presents a design framework and a multi-dimensional architecture organised into three complementary views: quality properties for trustworthiness and dependability, a layered technology stack within each tier, and a processing continuum that partitions intelligence across edge-to-cloud tiers. Finally, the article discusses value creation, interoperability, and how a 2 common baseline supports the development of complex edge AI systems-of-systems and their verification, validation, testing and benchmarking.

An Honest Effect Size for Contingency Tables: Why Nothing Can Be Unbiased, Where to Put the Error Instead, and How to Route the Report (m0f)

An Honest Effect Size for Contingency Tables: Why Nothing Can Be Unbiased, Where to Put the Error Instead, and How to Route the Report William J. Dwyer, MD, MPH, FAAP — Department of Mathematics and Statistics, University of Massachusetts Lowell. ORCID 0009-0004-0855-7222. Concept DOI (always resolves to the latest version): 10.5281/zenodo.21783660. What this is The reproducibility deposit for the point-estimate paper of the contingency-table effect-size arc. The everyday workflow reports Cramér's V and reads it against Cohen's labels; that workflow is biased at the null (Jensen's inequality makes an unbiased estimator of the square-root scale impossible) and treats a noisy point as if it were exact. The paper puts the unavoidable error where it belongs and routes the report: a closed-form O(RC) delete-one jackknife for V² that is nearly unbiased everywhere (at most 0.008 across the 180-design grid, 0.013 as a supremum), a projected confidence interval (conservative but asymptotically valid, any R×C) with an exact-conditional companion for tight coverage, a noise-floor verdict (V_{0.95}) that refuses labels a table cannot support, and a single routing rule that selects the reference (by sparsity and margin heterogeneity), the verdict, the point estimate (by purpose), the scale (always V², negatives permitted and meaningful), and the interval (by whether a guarantee is required). What the deposit contains Manuscript — the built article with the house-standard Key-points box, 8 figures, and 12 numbered/captioned tables (every table carrying an APA bold-label/italic caption and a substantive in-text callout), plus a novelty review companion. Derivations — the compressed derivations ladder and the proof-complete long-form Derivations Companion(Dwyer_M0f_Derivations_Companion_LongForm, ~38 pp, D11–D17 written out in full): the φ² non-U-estimability remark (attributed, with its self-correction), the projected interval as the projection principle instantiated, the affine-link uniqueness, the O(RC) closed-form jackknife, the Bergsma-vs-jackknife crossover equation, and the over-dispersion / two-moment interval fix. Interactive demonstrators — honest_point_estimate.html (the m0f effect-size router: computes the naive V, the jackknifed V², the noise-floor verdict, and the routed report in the browser) and honest_interval.html (the m0g guaranteed-interval sibling), both to the house flip-interpretation standard with a per-cell flip-incidence panel and a seeded null-calibration strip. Reproducibility scripts + locked outputs — every reported number traces to a named, deterministically-seeded script (the bias frontier vs Bergsma's prior, the pooling/meta-analysis collapse, the link/affine-scale test, the projected-interval benchmark, the routing grid, and the figure generators), with the analyzable public-table corpus (effect_size_tables_analyzable.csv) and the routing/flip-incidence scans (rerun/taxonomy_flip_incidence.py, taxonomy_grid_sim.py, and the m0g interval scans) included so the embedded numbers are traceable from the deposit. Real-data evidence — across 4,129 public contingency tables the label-vs-noise-floor recalibration runs both ways (615 forward / 450 reverse); the guaranteed-interval sibling withdraws a claimed resolution on 312 of 2,391 tables (13%, near-pure one-directional deflation). Figures and the deterministic deposit build (fixed timestamps → stable md5; a SHA-256 MANIFEST.txt listing every file). All evaluation is simulation-based; the one empirical component is the public-table scan, which uses only openly distributed data. Code is released under the MIT License; text and figures under CC BY 4.0. How to cite Please cite this deposit if you use the package or the method. Citing the concept DOI references the work in general and always resolves to the latest version; cite a specific version DOI to point at an exact snapshot. Dwyer, W. J. (2026). An Honest Effect Size for Contingency Tables: Why Nothing Can Be Unbiased, Where to Put the Error Instead, and How to Route the Report — reproducibility deposit [Software]. Zenodo.https://doi.org/10.5281/zenodo.21783660 BibTeX: bibtex @software{dwyer_m0f_2026, author = {Dwyer, William J.}, title = {An Honest Effect Size for Contingency Tables: Why Nothing Can Be Unbiased, Where to Put the Error Instead, and How to Route the Report --- reproducibility deposit}, year = {2026}, publisher = {Zenodo}, doi = {10.5281/zenodo.21783660}, url = {https://doi.org/10.5281/zenodo.21783660}, orcid = {0009-0004-0855-7222} } The DOI above is the concept DOI (resolves to the latest version); to cite a specific release, use that version's DOI and add version = {v1.2.NN}. When the accompanying journal article is published, please cite it as the primary reference for the method and this deposit as the reproducibility archive. Version history (consolidated changelog) Published version DOIs are marked ✅; the concept DOI above always resolves to the latest. Staged versions were rolled into the next published one unless noted. v1.2.11 ⏳ (staged 2026-09-01) — Live availability-DOI link + deepened table callouts (documents-only; no result, number, figure, or datum changed). The manuscript's Availability of data and materials statement now renders the concept DOI as a functional, clickable hyperlink in the built docx/pdf (was literal [...](...) text) — the m0-family docx builder was taught to render markdown [text](url) as a real hyperlink, resolving the reason the Behavior Research Methods editorial office returned the submission. The Table 11 and Table 12 callouts were deepened to point the reader at specific rows/values (float_callout substantive 0.35 → 0.45). Manuscript docx/pdf rebuilt from unchanged source; MANIFEST regenerated. v1.2.10 ✅ 10.5281/zenodo.22151244 (2026-08-28) — Long-form Derivations Companion added (~38 pp, D11–D17 written out in full) alongside the compressed ladder; documents-only, structure + math-render audits pass. v1.2.9 ✅ 10.5281/zenodo.22139670 (2026-08-28) — Routing-domain validation. Both browser demonstrators resolve flip incidence across the routing taxonomy (m0f: 615 forward / 450 reverse label-vs-floor recalibrations over 4,129 public tables; m0g: 312 of 2,391 guaranteed-interval withdrawals), with new reproducible scans, deposited JSON/CSV outputs, the analyzable public-table corpus, and two taxonomy deep-dives. v1.2.8 (2026-08-28) — De-bolded for verbal emphasis; Key-points highlights; refreshed novelty review.Emphasis-bold minimized (170 → ~42 spans; all 744 numeric tokens verified byte-identical), a house-standard seven-bullet Key-points box added after the abstract, and the novelty review rewritten to the clean strand-by-strand form. Presentation only; rolled into v1.2.9. v1.2.7 (2026-08-28) — Tables numbered, captioned, and called out. All 12 tables numbered (Table 1–12), given APA bold-label/italic captions, and referenced in the body (float_callout depth 1.00). Organization only; rolled into v1.2.9. v1.2.6 ✅ 10.5281/zenodo.22135868 (2026-08-27) — Routing-flowchart polish. Consistent ①–⑤ stage numbering (the scale stage is now a numbered pass-through in the figure too), vertical centering of decision-box titles, STOP-box clearance, rounded modal corners, and a legible star default marker. Presentation only. v1.2.1–v1.2.5 (2026-08-27) — The routing flowchart brought to a single house edge-label spec (figure + pop-up generated from one source), box-sizing / no-touching-boxes overlap guard, the pop-up flowchart aligned to the manuscript figure's visual style, and the browser tool taken to the flip-interpretation standard. Presentation + tooling; no result changed; rolled into v1.2.6. v1.2.0 ✅ 10.5281/zenodo.22003886 (2026-08-18) — Over-dispersion interval. The over-dispersion mechanism for the noncentral-interval drift and the derived two-moment ρ-interval (ρ_ci tight + ρ_ci_guard conservative, any R×C, JS-ported), rewriting Section 6.3 + Figure 6 and derivation D17 and superseding the earlier endpoint patch. v1.1.1 ✅ 10.5281/zenodo.21784544 (2026-08-04) — Earlier public release of the point-estimate package. Provenance: every number traces to a named, deterministically-seeded script listed in the manuscript Declarations; the two browser demonstrators reproduce the deposited Python, and the deposit build is deterministic (fixed timestamps → stable md5, SHA-256 MANIFEST).

SpikeGate: An Event-Driven Neuromorphic SNN Gatekeeper for Duty-Cycled On-Device SLM Inference

Probabilistic Disaggregation of Behind-the-Meter PV Systems Using Conformal Prediction

Disaggregating solar photovoltaics (PV) profiles from smart electricity meter data has attracted attention, as Distribution System Operators (DSOs) need street-level PV generation profiles to improve grid operations and planning. Given the importance of reliability in operational decisions, probabilistic results are preferred to avoid overlooking potential violations. This paper proposes a probabilistic disaggregation framework based on Conformal Prediction (CP), a cutting-edge uncertainty quantification methodology. This framework trains a deterministic regressor to estimate normalized PV generation profiles and proposes an efficient capacity estimation algorithm to help compute the full PV generation profiles. To obtain probabilistic results, the framework applied CP with different variants, such as Mondrian Binning (MB) and Conformal Predictive System (CPS), to enhance the reliability of prediction intervals. To address the arbitrary bin count in CP with MB, the paper proposes a novel CP variant, namely: Adaptive Mondrian Binning (AMB). Its performance, along with other CP methods, is evaluated and benchmarked against quantile regression methods on two actual datasets from the region of Amsterdam, the Netherlands, and Sydney, Australia. Results show that using LightGBM as the deterministic regressor, AMB outperforms quantile regression and other CP variants. The generalizability of the proposed framework is analysed for both probabilistic outputs and key sub-processes, such as deterministic disaggregation and capacity estimation.

Physically implemented analog in-memory distance computing with InGaZnO transistor-based capacitive unit

Modern edge devices increasingly require real-time adaptation to their environment without relying on cloud-based updates, which can introduce latency and security risks. To meet these demands, memory-augmented neural networks (MANNs) have gained traction for enabling adaptive on-device learning. Hardware implementations of MANNs commonly use non-volatile memory-based ternary content-addressable memory (TCAM), but their discrete outputs and write-verify steps limit compatibility with gradient-based learning. This work introduces analog in-memory distance computing (AIMDC), a unified architecture based on indium gallium zinc oxide (IGZO) thin-film transistors that performs both embedding extraction and similarity search using analog capacitive units (ACUs). By producing continuous, differentiable outputs directly from analog embeddings, AIMDC enables hardware-in-the-loop on-device representation learning without additional processing. We demonstrate energy-efficient one-shot learning accuracy comparable to a graphics processing unit but with up to a 576× improvement in energy efficiency. The high retention and endurance of the IGZO-based ACUs establish AIMDC as a scalable and robust solution for high-throughput, low-energy edge learning.

Joint Optimization of Memory and Computing Frequency for Energy-Efficient DNN Inference

The impact of memory frequency and computing frequency on DNN inference time is considered, and a low-complexity heuristic algorithm is proposed to effectively solve the overall problem with polynomial time complexity.

UVirtio: Enabling Ubiquitous Resource Sharing for RISC-V Industrial Edge Devices

UVirtio introduces a device-profile-based virtual hardware abstraction layer that minimizes performance overhead, and implements a live migration mechanism using differential packing, providing a scalable and agile virtualization solution for the ubiquitous computing frontier.

Beat the Counter First: A Baseline for Temporal-Graph Anomaly Detectors

SimpleCount is proposed, a reference with no parameter fitting that selects one scalar feature per dataset from a fixed pool of counts, recencies, first-occurrence indicators, and count-derived transforms that matches or exceeds SLADE on three of six datasets and exceeds IsoForest on all six.

Design And Implementation Of An ADC/DAC-Free Walsh-Hadamard Transform Based Neural Network Accelerator Using Bit-Plane Processing

An ADC/DAC-free neural accelerator based on the Walsh-Hadamard Transform and bit-plane processing that offers a multiplier-free, converter-free, regular, and scalable solution for low-power edge intelligence.

Construction and application of a platform for on-site detection data perception, analysis, and intelligent decision-making of fluids and materials entering the well

With the transformation of unconventional oil and gas resource development towards digitalization and intelligence, massive, multi-source, and dynamic well-entry fluid and material detection data have become core assets driving engineering decision-making. Traditional digital platforms aimed at process onlineization have limitations in deep perception, intelligent analysis, and value mining of data. This paper systematically expounds the next-generation on-site detection intelligent platform built around the core of "global data perception-fusion analysis-intelligent decision-making". The platform achieves high-frequency, automated real-time perception and collection of key fluid performance parameters such as drilling fluid and fracturing fluid through Internet of Things (IoT) and edge computing technology; utilizes a big data technology stack to construct a multidimensional data analysis model that integrates detection data, engineering parameters, and geological information; and develops intelligent decision-making support modules for quality prediction, anomaly warning, and formula optimization based on machine learning algorithms. Application practice shows that the platform not only realizes digital control of detection operations but also shifts quality control from "post-judgment" to "pre-prediction" and "in-process intervention" through data intelligence, significantly enhancing the precise control ability and construction optimization level of well engineering quality. The development direction of the platform is to build a "digital twin" of oil and gas well engineering quality, ultimately forming a smart quality ecosystem featuring autonomous perception, intelligent analysis, and closed-loop optimization.

A Green AI Based Approach to Optimizing Machine Learning Models for Sustainable Computing

A Green AI–driven framework is presented for optimizing machine learning models with the goal of supporting sustainable computing and offers a practical and reproducible solution for deploying energy-efficient machine learning models across both cloud and edge environments.

IDENTIFICATION OF REGIONAL TECTONIC STRUCTURES OF CAMEROON VOLCANIC LINE USING GRAVIMAGNETIC AND 2D LAPLACIAN TERRAIN ANALYSIS TECHNIQUES

Cameroon is made up of an extensive tectonic system that includes a geologically significant structure in Central Africa called the Cameroon Volcanic Line (CVL). In this study, a multi-phase approach was applied to identify new tectonic structures across the CVL using gravity data from the XGM2019e_2159 model, magnetic data from EMAG2, and topographic data from GMRT. First-order and second-order vertical derivatives of both magnetic and gravity data were used to filter geologic structures in the study areas. To understand the trend of the structures based on the potential field, edge enhancement of potential-field data using normalized statistics (standard deviation edge detection) and phase sym­metry were applied. Phase symmetry provides phase-based formulation that can detect ridges and valleys corresponding to geological structures regardless of anomaly strength, making it robust to noise and interpolation artifacts. For de­tailed analysis to delineate new and already existing lineaments, a 2D Laplacian terrain transformation of the topographic data was computed at various heights from 20 to 105 km. The gravitational and magnetic structures discovered were superimposed on the transformed topographic map and drawn out to identify the orientation and direction of the faults. The study further illustrates the occurrence of potential geologic features associated with gravity anomalies, magnetic anomalies, and the new lineaments identified in the area. These results highlight that the lineaments identified are mostly orientated SE-SW and act as evidence of the still continuing process of deformation of the CVL. This finding is essential for advanced regional tectonic modeling across the CVL.