Mixed-precision quantization (MPQ) is crucial for deploying deep neural networks on resource-constrained devices, but finding the optimal bit-width for each layer represents a complex combinatorial optimization problem. Current state-of-the-art methods rely on computationally expensive search algorithms or local sensitivity heuristic proxies like the Hessian, which fail to capture the cascading global effects of quantization error. In this work, we argue that the quantization sensitivity of a layer should not be measured by its local properties, but by its impact on the information flow throughout the entire network. We introduce InfoQ, a novel framework for mixed-precision quantization that is training-free in the bit-width search phase. InfoQ assesses layer importance by performing a single forward pass to measure the change in mutual information in the remaining part of the network, thus creating a global sensitivity score. This approach directly quantifies how quantizing one layer degrades the information characteristics of subsequent layers. The resulting scores are used to formulate bit-width allocation as an integer linear programming problem, which is solved efficiently to minimize total sensitivity under a given budget (e.g., model size or BitOps). Our retraining-free search phase provides a superior search-time/accuracy trade-off (using two orders of magnitude less data compared to state-of-the-art methods such as LIMPQ), while yielding up to a 1% accuracy improvement for MobileNetV2 and ResNet18 on ImageNet at high compression rates (14.00x and 10.66x).

The deployment of deep neural networks on resource-constrained devices relies on quantization. While static, uniform quantization applies a fixed bit-width to all inputs, it fails to adapt to their varying complexity. Dynamic, instance-based mixed-precision quantization promises a superior accuracy-efficiency trade-off by allocating higher precision only when needed. However, a critical bottleneck remains: existing methods require a costly dequantize-to-float and requantize-to-integer cycle to change precision, breaking the integer-only hardware paradigm and compromising performance gains. This paper introduces Dynamic Quantization Training (DQT), a novel framework that removes this bottleneck. At the core of DQT is a nested integer representation where lower-precision values are bit-wise embedded within higher-precision ones. This design, coupled with custom integer-only arithmetic, allows for on-the-fly bit-width switching through a near-zero-cost bit-shift operation. This makes DQT the first quantization framework to enable both dequantization-free static mixed-precision of the backbone network, and truly efficient dynamic, instance-based quantization through a lightweight controller that decides at runtime how to quantize each layer. We demonstrate DQT state-of-the-art performance on ResNet18 on CIFAR-10 and ResNet50 on ImageNet. On ImageNet, our 4-bit dynamic ResNet50 achieves 77.00% top-1 accuracy, an improvement over leading static (LSQ, 76.70%) and dynamic (DQNET, 76.94%) methods at a comparable BitOPs budget. Crucially, DQT achieves this with a bit-width transition cost of only 28.3M simple bit-shift operations, a drastic improvement over the 56.6M costly Multiply-Accumulate (MAC) floating-point operations required by previous dynamic approaches - unlocking a new frontier in efficient, adaptive AI.

Transformer architectures based on the attention mechanism have revolutionized natural language processing (NLP), driving major breakthroughs across virtually every NLP task. However, their substantial memory and computational requirements still hinder deployment on ultra-constrained devices such as wearables and Internet-of-Things (IoT) units, where available memory is limited to just a few megabytes. To address this challenge, we introduce EmbBERT, a tiny language model (TLM) architecturally designed for extreme efficiency. The model integrates a compact embedding layer, streamlined feed-forward blocks, and an efficient attention mechanism that together enable optimal performance under strict memory budgets. Through this redesign for the extreme edge, we demonstrate that highly simplified transformer architectures remain remarkably effective under tight resource constraints. EmbBERT requires only 2 MB of total memory, and achieves accuracy performance comparable to the ones of state-of-the-art (SotA) models that require a 10 × memory budget. Extensive experiments on the curated TinyNLP benchmark and the GLUE suite confirm that EmbBERT achieves competitive accuracy, comparable to that of larger SotA models, and consistently outperforms downsized versions of BERT and MAMBA of similar size. Furthermore, we demonstrate the model's resilience to 8-bit quantization, which further reduces memory usage to just 781 kB, and the scalability of the EmbBERT architecture across the sub-megabyte to tens-of-megabytes range. Finally, we perform an ablation study demonstrating the positive contributions of all components and the pre-training procedure. All code, scripts, and checkpoints are publicly released to ensure reproducibility: https://github.com/RiccardoBravin/tiny-LLM.

Training neural networks under Torus Fully Homomorphic Encryption (TFHE) is severely constrained by the native restriction of the scheme to boolean and integer arithmetic, forcing prior work to rely on quantized integer pipelines limited to shallow MLPs. We present a framework that enables approximate floating-point training within TFHE by reinterpreting IEEE 754 representations as encrypted integers and operating on them with redesigned arithmetic. Our core contribution adapts L-mul, an approximate integer-based multiplication, to the encrypted setting, introducing numerical safeguards that prevent catastrophic wrap-around errors near zero and in subnormal ranges. We extend this approach to approximate division and implement exact addition and square root, yielding a sufficient arithmetic for backpropagation with SGD. For CPU-based FP32 encrypted operations, our approach achieves up to 2.1x faster multiplication and 29.3x faster division compared to state-of-the-art exact TFHE floating-point arithmetic, alongside up to 57x lower peak memory. Using these primitives, we perform, to our knowledge, the first fully encrypted training of a small-scale CNN under TFHE. To bypass the prohibitive overhead of training deep networks in this encrypted environment, we utilize plain-text emulation. We ensure strict output alignment by verifying that the network parameters generated during emulation are identical to those produced by the encrypted execution. Leveraging this equivalence, we use emulation to evaluate convergence on larger architectures (LeNet-5, VGG-style CNNs, and ResNet-20) across six benchmarks from MNIST variants to CIFAR-10 and medical imaging, matching exact-arithmetic baselines within 1% accuracy. Fully encrypted training of these deeper models remains beyond current hardware; we provide wall-clock projections quantifying this gap.

Quantization is a pivotal technique for managing the growing computational and memory demands of Deep Neural Networks (DNNs). By reducing the number of bits used to represent weights and activations (typically from 32-bit Floating-Point (FP) to 16-bit or 8-bit integers), quantization reduces memory footprint, energy consumption, and execution time of DNNs. However, most existing methods typically target DNN inference, while training still relies on FP operations, limiting applicability in environments where FP arithmetic is unavailable. To date, only one prior work has addressed integer-only training, and only for Multi-Layer Perceptron (MLP) architectures. This paper introduces NITRO-D, a novel framework for training deep integer-only Convolutional Neural Networks (CNNs) that operate entirely in the integer domain for both training and inference. NITRO-D enables training of integer CNNs without requiring a separate quantization scheme. Specifically, it introduces a novel architecture that integrates multiple local-loss blocks, which include the proposed NITRO-Scaling layer and NITRO-ReLU activation function. The proposed framework also features a novel learning algorithm that employs local error signals and leverages IntegerSGD, an optimizer specifically designed for integer computations. NITRO-D is implemented as an open-source Python library. Extensive evaluations on state-of-the-art image recognition datasets demonstrate its effectiveness. For integer-only MLPs, NITRO-D improves test accuracy by up to +5.96% over the state-of-the-art. It also successfully trains integer-only CNNs, reducing memory requirements and energy consumption by up to 76.14% and 32.42%, respectively, compared to the traditional FP backpropagation algorithm.

Early exits (EEs) offer a promising approach to reducing computational costs and latency by dynamically terminating inference once a satisfactory prediction confidence on a data sample is achieved. Although many works integrate EEs into encoder-only Transformers, their application to decoder-only architectures and, more importantly, Mamba models, a novel family of state-space architectures in the LLM realm, remains insufficiently explored. This work introduces DYNAMAX, the first framework to exploit the unique properties of Mamba architectures for early exit mechanisms. We not only integrate EEs into Mamba but also repurpose Mamba as an efficient EE classifier for both Mamba-based and transformer-based LLMs, showcasing its versatility. Our experiments employ the Mistral 7B transformer compared to the Codestral 7B Mamba model, using data sets such as TruthfulQA, CoQA, and TriviaQA to evaluate computational savings, accuracy, and consistency. The results highlight the adaptability of Mamba as a powerful EE classifier and its efficiency in balancing computational cost and performance quality across NLP tasks. By leveraging Mamba’s inherent design for dynamic processing, we open pathways for scalable and efficient inference in embedded applications and resource-constrained environments. This study underscores the transformative potential of Mamba in redefining dynamic computing paradigms for LLMs.

Tiny Machine Learning (TinyML) algorithms have seen extensive use in recent years, enabling wearable devices to be not only connected but also genuinely intelligent by running machine learning (ML) computations directly on-device. Among such devices, smart glasses have particularly benefited from TinyML advancements. TinyML facilitates the on-device execution of the inference phase of ML algorithms on embedded and wearable devices, and more recently, it has expanded into On-device Learning (ODL), which allows both inference and learning phases to occur directly on the device. The application of ODL techniques to wearable devices is particularly compelling, as it enables the development of more personalized models that adapt based on the user’s data. However, one of the major challenges of ODL algorithms is the scarcity of labeled data collected on-device. In smart wearable contexts, requiring users to manually label large amounts of data is often impractical and could lead to user disengagement with the technology.To address this issue, this paper explores the application of Active Learning (AL) techniques, i.e., techniques that aim at minimizing the labeling effort, by actively selecting from a large quantity of unlabeled data only a small subset to be labeled and added to the training set of the algorithm. In particular, we propose TActiLE, a novel AL algorithm that selects from the stream of on-device sensor data the ones that would help the ML algorithm improve the most once coupled with labels provided by the user. TActiLE is the first Active Learning technique specifically designed for the TinyML context. We evaluate its effectiveness and efficiency through experiments on multiple image classification datasets. The results demonstrate that, within the stringent resource constraints of TinyML and wearables environments, TActiLE outperforms both a simple random selection baseline and more complex Stream Active Learning algorithms, demonstrating its suitability for tiny and wearable devices.

Privacy-Preserving Machine and Deep Learning (PP-MDL) leverages Homomorphic Encryption (HE) to enable the inference and training of Machine and Deep Learning (ML and DL) models on encrypted data, addressing the stringent privacy requirements of domains such as healthcare and finance. While PP-MDL has achieved maturity in inference tasks, training on encrypted data remains a significant challenge due to the computational overhead and operational constraints of HE. This paper introduces Homomorphically Encrypted Distributed Ensemble Learning (HEDEL), a distributed training architecture that enables the training of encrypted ensemble models, starting from encrypted base models and encrypted datasets, by combining Transfer Learning (TL) and Multi-Key Homomorphic Encryption (MKHE). HEDEL presents two unique features. First, TL enables HEDEL to dramatically reduce the computational cost of encrypted training by requiring only a few training epochs to achieve high-accuracy models, making it particularly effective in data-scarce scenarios. Second, MKHE provides a fine-grained access control mechanism for datasets and base models. Through collaborative decryption, model providers can not only regulate who accesses their models but also limit how often these models are used for training and inference. This ensures secure, controlled usage while preventing unauthorized or excessive access, even when a third party offers HEDEL. By combining the efficiency of TL with the security and control of MKHE, HEDEL provides a robust and scalable solution for PP-MDL in collaborative environments involving model providers, who share encrypted models to enable TL, and users who seek accurate, privacy-preserving predictions on their private datasets. Experimental results highlight its ability to overcome HE constraints while delivering high accuracy and strong privacy guarantees.

Early-exit neural networks (EENNs) endow a standard deep neural network (DNN) with early-exit classifiers (EECs) to provide predictions at intermediate points of the processing when enough confidence in classification is achieved. This leads to many benefits in terms of effectiveness and efficiency. Currently, the design of EENNs is carried out manually by experts, a complex and time-consuming task that requires accounting for many aspects, including the correct placement, the thresholding, and the computational overhead of the EECs. For this reason, the research is exploring the use of neural architecture search (NAS) to automate the design of EENNs. Currently, few comprehensive NAS solutions for EENNs have been proposed in the literature, and a fully automated, joint design strategy taking into consideration both the backbone and the EECs remains an open problem. To this end, this work presents neural architecture search for hardware-constrained early exit neural networks (NACHOS), the first NAS framework for the design of optimal EENNs satisfying constraints on the accuracy and the number of multiply and accumulate (MAC) operations performed by the EENNs at inference time. In particular, this provides the joint design of backbone and EECs to select a set of admissible (i.e., respecting the constraints) Pareto optimal solutions in terms of the best trade-off between the accuracy and the number of MACs. The results show that the models designed by NACHOS are competitive with the state-of-the-art EENNs. Additionally, this work investigates the effectiveness of two novel regularization terms designed for the optimization of the auxiliary classifiers of the EENN.

Edge AI seeks for the deployment of deep neural network (DNN) based services across distributed edge devices, embedding intelligence close to data sources. Due to capacity constraints at the edge, a difficult challenge lies in planning a dependable deployment that minimizes the data loss rate so as to meet application Quality-of-Service (QoS) goals. In this paper, we present ChainNet, a customized graph neural network (GNN) model serving as a surrogate to assess the reliability of alternative deployments and guide the loss-aware search for an optimal edge AI deployment plan. Extensive results show that ChainNet delivers a substantial improvement in loss prediction accuracy by over 50% compared to established GNN models, such as graph attention networks (GATs). Moreover, we show that ChainNet provides significantly more dependable deployment decisions under a fixed time budget compared to simulation-based search across a spectrum of systems from small to large-scale.

Accurate classification of medical images is essential for modern diagnostics.Deep learning advancements led clinicians to increasingly use sophisticated models to make faster and more accurate decisions, sometimes replacing human judgment.However, model development is costly and repetitive.Neural Architecture Search (NAS) provides solutions by automating the design of deep learning architectures.This paper presents ZO-DARTS+, a differentiable NAS algorithm that improves search efficiency through a novel method of generating sparse probabilities by bilevel optimization.Experiments on five public medical datasets show that ZO-DARTS+ matches the accuracy of state-of-the-art solutions while reducing search times by up to three times.

Neural Architecture Search (NAS) paves the way for the automatic definition of Neural Network (NN) architectures, attracting increasing research attention and offering solutions in various scenarios. This study introduces a novel NAS solution, called Flat Neural Architecture Search (FlatNAS), which explores the interplay between a novel figure of merit based on robustness to weight perturbations and single NN optimization with Sharpness-Aware Minimization (SAM). FlatNAS is the first work in the literature to systematically explore flat regions in the loss landscape of NNs in a NAS procedure, while jointly optimizing their performance on in-distribution data, their Out-of-Distribution (OOD) robustness, and constraining the number of parameters in their architecture. Differently from current studies primarily concentrating on OOD algorithms, FlatNAS successfully evaluates the impact of NN architectures on OOD robustness, a crucial aspect in real-world applications of machine and deep learning. FlatNAS achieves a good trade-off between performance, OOD generalization, and the number of parameters, by using only in-distribution data in the NAS exploration. The OOD robustness of the NAS-designed models is evaluated by focusing on robustness to input data corruptions, using popular benchmark datasets in the literature.

Tiny Machine Learning (TinyML) is a branch of Machine Learning (ML) that constitutes a bridge between the ML world and the embedded system ecosystem (i.e., Internet-of-Things devices, embedded devices, and edge computing units), enabling the execution of ML algorithms on devices constrained in terms of memory, computational capabilities, and power consumption. Video Streaming Analysis (VSA), one of the most interesting tasks of TinyML, consists in scanning a sequence of frames in a streaming manner, with the goal of identifying interesting patterns. Given the strict constraints of these tiny devices, all the current solutions rely on performing a frame-by-frame analysis, hence not exploiting the temporal component in the stream of data. In this paper, we present StreamTinyNet, the first TinyML architecture to perform multiple-frame VSA, enabling a variety of use cases that requires spatial-temporal analysis that were previously impossible to be carried out at a TinyML level. Experimental results on public-available datasets show the effectiveness and efficiency of the proposed solution. Finally, StreamTinyNet has been ported and tested on the Arduino Nicla Vision, showing the feasibility of what proposed.

Early exit neural networks (EENs) reduce the processing times of deep convolutional neural networks by means of internal classifiers (ICs) that allow jobs, being the input of the EEN, to exit early from the processing pipeline. However, the current designs used in pervasive systems ignore variability in data arrival rates, exposing EEN-based services to potential loss of the incoming jobs, due to finite input buffer capacity. Motivated by this issue, we introduce and study theearly exitscheduling problem, which aims at dynamically configuring IC thresholds at runtime to achieve effective trade-offs between job classification accuracy, processing time, and job loss ratio. We argue that deciding the EEN exit layer for a job at the start of its processing makes the problem mathematically tractable, allowing us to develop policies to control buffer backlog, classification accuracy, and processing time across the EEN layers. The main contribution of the paper is the introduction of single-exit IC threshold configurations as a mechanism to allow the scheduling policy to reliably predict the best EEN exit layer of each input job. Three scheduling policies that leverage this idea are proposed to dynamically schedule job arrivals to an EEN-based service. The proposed solution, here tailored to EENs based on convolutional neural networks (CNNs), is fairly general and can be applied to different use cases. The two application scenarios considered in this paper focus on image classification and intrusion detection. Experiments on some popular CNNs for the two aforementioned application scenarios indicate that the proposed policies can achieve significant savings in processing times and improve job loss ratio compared to both ordinary EENs and CNNs while still providing high mean classification accuracy.

Incremental on-device learning is one of the most relevant and interesting challenges in the field of Tiny Machine Learning (TinyML). Indeed, differently from traditional TinyML solutions where the training is typically carried out on the Cloud and inference only occurs on the tiny devices (e.g., embedded systems or Internet-of-Things units), incremental on-device TinyML allows both the inference and the training of TinyML models directly on tiny devices. This ability paves the way for TinyML-enabled intelligent devices that can learn directly on the field and adapt to evolving environments, different working conditions, or specific users. The literature in this field is quite limited with very few solutions focusing only on the incremental fine-tuning of machine learning models, whereas a general solution encompassing algorithms and code generation for incremental on-device TinyML is still perceived as missing. The aim of this article is to introduce, to the best of our knowledge for the first time in the literature, a toolbox called TyBox for the automatic design and code generation of incremental on-device TinyML classification models. In more detail, starting from a “static” TinyML model, TyBox is able to (i) automatically design the “incremental” on-device version of the TinyML model that has been suitably designed to take into account the technological constraint on the RAM memory of the target tiny device, and (ii) autonomously provide the C++ codes and libraries to support the inference and learning of the incremental on-device TinyML model directly on the tiny devices. TyBox has been extensively compared with a state-of-the-art incremental learning solution for TinyML and tested on an off-the-shelf tiny device (i.e., the Arduino Nano 33 BLE) in three relevant TinyML application tasks and scenarios: binary image classification, multi-class image classification, and ultra-wide-band human activity recognition. In addition, TyBox is released to the scientific community as a public repository.

Tiny machine learning (TML) is a new research area whose goal is to design machine and deep learning (DL) techniques able to operate in embedded systems and the Internet-of-Things (IoT) units, hence satisfying the severe technological constraints on memory, computation, and energy characterizing these pervasive devices. Interestingly, the related literature mainly focused on reducing the computational and memory demand of the inference phase of machine and deep learning models. At the same time, the training is typically assumed to be carried out in cloud or edge computing systems (due to the larger memory and computational requirements). This assumption results in TML solutions that might become obsolete when the process generating the data is affected by concept drift (e.g., due to periodicity or seasonality effect, faults or malfunctioning affecting sensors or actuators, or changes in the users' behavior), a common situation in real-world application scenarios. For the first time in the literature, this article introduces a TML for concept drift (TML-CD) solution based on deep learning feature extractors and a k -nearest neighbors ( k -NNs) classifier integrating a hybrid adaptation module able to deal with concept drift affecting the data-generating process. This adaptation module continuously updates (in a passive way) the knowledge base of TML-CD and, at the same time, employs a change detection test (CDT) to inspect for changes (in an active way) to quickly adapt to concept drift by removing obsolete knowledge. Experimental results on both image and audio benchmarks show the effectiveness of the proposed solution, whilst the porting of TML-CD on three off-the-shelf micro-controller units (MCUs) shows the feasibility of what is proposed in real-world pervasive systems.

Coupled simulation, also known as co-simulation, has been proposed to provide more information to a task scheduler by simulating at runtime the Quality of Service (QoS) arising from a scheduling action. To do so, co-simulation algorithms run the simulation assuming a static set of arrival time series, restricting the diversity of the traffic scenarios. To ensure the co-simulator can provide valuable and representative results, we present an online adaptive arrival forecasting framework that contains a change-point detection module and a probabilistic transformer model to couple co-simulators with arrival series forecasting. The framework can also update the prediction model to adapt to dynamic environments. Our experiments show that our online adaptive forecasting framework has lower forecasting errors than established prediction models, such as autoregressive processes, and lower on real-world traces the co-simulator prediction error by up to 27 % on average response time and 39% on average service-level agreement (SLA) violation.

Personalization is among the most promising outcomes of using Machine Learning models that can be trained on data representing a specific individual. Personalization is particularly promising in areas such as health and medicine, as several crucial aspects and determinants of health are individual. Yet additional ethical issues arise with increasingly personalized models, including privacy, acceptability, reliability and trade-offs. In this paper we discuss and propose ML models for health that can be personalized on individual users, while guaranteeing both their privacy and quality from an ethical and epistemic (knowledge-related) point of view. To achieve these goals, we argue that we need to control the learning and evolution of personalized models. We propose soft personalization as an ethically-informed framework to limit personalization and respect epistemic and ethical values that are specific for the health context, including representativity, quality, non-maleficence, beneficence, privacy. Based on an interdisciplinary approach combining the philosophical and computer science scholarship of our group, soft personalization is a way of developing different models that can be selected depending on their quality and safety. We characterize the approach theoretically and technically and make it concrete with a case study of glucose monitoring and anomaly detection through privacy-preserving ML. Our framework shows that, even when individual issues such as privacy can be mitigated, tradeoffs with other values remain and choices are necessary as to which values should be prioritized.

Early Exit Neural Networks (EENNs) endow neural network architectures with auxiliary classifiers to progressively process the input and make decisions at intermediate points of the network. This leads to significant benefits in terms of effectiveness and efficiency such as the reduction of the average inference time as well as the mitigation of overfitting and vanishing gradient phenomena. Currently, the design of EENNs, which is a very complex and time-consuming task, is carried out manually by experts. This is where Neural Architecture Search (NAS) comes into play by automatically designing neural network architectures focusing also on the optimization of the computational demand of these networks. These requirements are crucial in the design of machine and deep learning solutions meant to operate in devices constrained by the technology (computation, memory, and energy) such as Internet-Of-Things and embedded systems. Interestingly, few NAS solutions have taken into account the design of early exiting mechanisms. This work introduces, for the first time in the literature, a framework called Early exit aDAptive Neural Architecture Search (EDANAS) for the automatic design of both the EENN architecture and the parameters that manage its early exit mechanism in order to optimize both the accuracy in the classification tasks and the computational demand. EDANAS has proven to compete with expert-designed early exit solutions paving the way for a new era in the prominent field of NAS.

Designing and developing machine and deep learning solutions able to guarantee the privacy of users' data is a novel and promising research area. Homomorphic Encryption (HE) is playing a primary role in this area thanks to its ability to support the processing of machine and deep learning solutions directly on encrypted data. Currently, the research in this field focuses on HE-based machine and deep learning solutions for the processing of images and text, while the privacy-preserving processing of time series has been mostly left unattended due to the strong constraints imposed by HE on the machine and deep learning forecasting models. This paper introduces, for the first time in the literature, a general privacy-preserving solution for time series prediction based on HE and Temporal Convolutional Neural Networks. The novel content brought by the paper is twofold. From the algorithmic point of view, this paper introduces a family of Temporal Convolutional Neural Networks, called PINPOINT, which is integrated with a HE scheme to support the privacy-preserving time series prediction. From the technical point of view, this paper introduces and details a Cloud-based privacy-preserving system for the forecasting-as-a-service based on the proposed PINPOINT models. Experimental results on publicly available benchmarks show the effectiveness of the proposed solution for privacy-preserving time series prediction.

Tiny Machine Learning (TinyML) is a novel research area aiming at designing machine and deep learning models and algorithms able to be executed on tiny devices such as Internet-of-Things units, edge devices or embedded systems. In this paper we introduce, for the first time in the literature, a TinyML solution for presence-detection based on UltrawideBand (UWB) radar, which is a particularly promising radar technology for pervasive systems. To achieve this goal we introduce a novel family of tiny convolutional neural networks for the processing of UWB-radar data characterized by a reduced memory footprint and computational demand so as to satisfy the severe technological constraints of tiny devices. From this technological perspective, UWB-radars are particularly relevant in the presence-detection scenario since they do not acquire sensitive information of users (e.g., images, videos or audio), hence preserving their privacy. The proposed solution has been successfully tested on a public-available benchmark for the indoor presence detection and on a real-world application of in-car presence detection.

Neural Architecture Search (NAS) paves the way for the automatic definition of neural networks architectures. The research interest in this field is steadily growing with several solutions available in the literature. This study introduces, for the first time in the literature, a NAS solution, called Constrained NAS (CNAS), able to take into account constraints on the search of the designed neural architecture. Specifically, CNAS is able to consider both functional constraints (i.e., the type of operations that can be carried out in the neural network) and technological constraints (i.e., constraints on the computational and memory demand of the designed neural network). CNAS has been successfully applied to Tiny Machine Learning and Privacy-Preserving Deep Learning with Homomorphic Encryption being two relevant and challenging application scenarios where functional and technological constraints are relevant in the neural network search.

Abstract Pipeline infrastructures, carrying either gas or oil, are often affected by internal corrosion, which is a dangerous phenomenon that may cause threats to both the environment (due to potential leakages) and the human beings (due to accidents that may cause explosions in presence of gas leakages). For this reason, predictive mechanisms are needed to detect and address the corrosion phenomenon. Recently, we have seen a first attempt at leveraging Machine Learning (ML) techniques in this field thanks to their high ability in modeling highly complex phenomena. In order to rely on these techniques, we need a set of data, representing factors influencing the corrosion in a given pipeline, together with their related supervised information, measuring the corrosion level along the considered infrastructure profile. Unfortunately, it is not always possible to access supervised information for a given pipeline since measuring the corrosion is a costly and time-consuming operation. In this paper, we will address the problem of devising a ML-based predictive model for internal corrosion under the assumption that supervised information is unavailable for the pipeline of interest, while it is available for some other pipelines that can be leveraged through Transfer Learning (TL) to build the predictive model itself. We will cover all the methodological steps from data set creation to the usage of TL. The whole methodology will be experimentally validated on a set of real-world pipelines.

Federated Learning (FL) is a promising research area in the machine learning field. Techniques and solutions belonging to this area operate in distributed scenarios, comprising a server and pervasively distributed clients, aiming at learning a single central model without sending (possibly sensitive) data from the clients to the server. Such an approach allows mitigating the privacy concerns that are nowadays perceived as relevant in distributed machine learning solutions leveraging data belonging to different users or companies. The literature in the field of FL is wide and many state-of-the-art solutions are available. Unfortunately, all these solutions assume (implicitly or explicitly) that the process generating the data is stationary (hence not changing its statistical behavior over time); an assumption that rarely holds in real-world conditions where concept drift occurs due to, e.g., seasonality or periodicity effects, faults in sensors or actuators or changes in the users' behaviour. In this paper, we introduce, for the first time in the literature, a novel FL algorithm called Adaptive-FedAVG, able to operate with nonstationary data generating processes affected by concept drifts. Following a passive approach, Adaptive-FedAVG is able to increase the accuracy in stationary conditions and promptly react to concept drift by adapting the learning rate to increase the plasticity of the learning phase. A wide experimental campaign shows the effectiveness of the proposed Adaptive-FedAVG algorithm by comparing it with a state-of-the-art FL algorithm present in the literature both in stationary and non-stationary conditions.

Understanding the distribution of bird species and populations and learning how birds behave and communicate are of great importance in wildlife biology, animal ecology, conservation of ecosystems, and assessing the effects of climate change and urbanization. The temporal and spatial limitations of human observation have motivated significant efforts to develop technology for bird song and vocalization detection and classification. While solutions based on signal processing and machine learning are extant, they are limited in various combinations of speed, computational complexity, and memory use, as well as in detection/classification capability in real-world conditions. This paper introduces ToucaNet, a deep neural network for birdsong detection based on transfer-learning, a deep learning mechanism allowing us to exploit knowledge acquired on various tasks: this enables us to speed up training and shows improved detection accuracy. ToucaNet provides birdsong detection accuracy in line with the best solutions in the literature but with much less computational complexity and memory demand. We also introduce BarbNet, an approximated version of ToucaNet tailored for Internet-of-Things (IoT) units. We show the proposed solution’s effectiveness and efficiency in terms of detection accuracy and the implementation feasibility in real-world IoT devices, with specific results for the STM32 Nucleo H7 board, which is based on an ARM Cortex-M7 processor. To our best knowledge, this is the first birdsong detection algorithm designed to take into account constraints on memory, computational speed, and power usage of embedded devices. Thus, this work points the way to cost-effective IoT technology for at-scale intelligent birdsong data collection and analysis in the field.

In most of the transfer learning approaches to reinforcement learning (RL) the distribution over the tasks is assumed to be stationary. Therefore, the target and source tasks are i.i.d. samples of the same distribution. Unfortunately, this assumption rarely holds in real-world conditions, e.g., due to seasonality or periodicity, evolution in the environment or faults in the sensors/actuators. In the context of this work, we consider the problem of transferring value functions through a variational method when the distribution that generates the tasks is time-variant, proposing a solution that leverages this temporal structure inherent in the task generating process. Furthermore, by means of a finite-sample analysis, the previously mentioned solution is theoretically compared to its time-invariant version. Finally, the experimental evaluation of the proposed technique is carried out on the lake Como water system representing a real-world scenario and on three different RL environments with three distinct temporal dynamics.

Severe constraints on memory and computation characterizing the Internet-of-Things (IoT) units may prevent the execution of Deep Learning (DL)-based solutions, which typically demand large memory and high processing load. In order to support a real-time execution of the considered DL model at the IoT unit level, DL solutions must be designed having in mind constraints on memory and processing capability exposed by the chosen IoT technology. In this article, we introduce a design methodology aiming at allocating the execution of Convolutional Neural Networks (CNNs) on a distributed IoT application. Such a methodology is formalized as an optimization problem where the latency between the data-gathering phase and the subsequent decision-making one is minimized, within the given constraints on memory and processing load at the units level. The methodology supports multiple sources of data as well as multiple CNNs in execution on the same IoT system allowing the design of CNN-based applications demanding autonomy, low decision-latency, and high Quality-of-Service.

In most Reinforcement Learning (RL) studies, the considered task is assumed to be stationary, i.e., it does not change its behavior or its characteristics over time, as this allows to generate all the convergence properties of RL techniques. Unfortunately, this assumption does not hold in real-world scenarios where systems and environments typically evolve over time. For instance, in robotic applications, sensor or actuator faults would induce a sudden change in the RL settings, while in financial applications the evolution of the market can cause a more gradual variation over time. In this paper, we present an adaptive RL algorithm able to detect changes in the environment or in the reward function and react to these changes by adapting to the new conditions of the task. At first, we develop a figure of merit onto which a hypothesis test can be applied to detect changes between two different learning iterations. Then, we extended this test to sequentially operate over time by means of the CUmulative SUM (CUSUM) approach. Finally, the proposed change-detection mechanism is combined (following an adaptive-active approach) with a well known RL algorithm to make it able to deal with non-stationary tasks. We tested the proposed algorithm on two well-known continuous-control tasks to check its effectiveness in terms of non-stationarity detection and adaptation over a vanilla RL algorithm.

Deep-learning-as-a-service is a novel and promising computing paradigm aiming at providing machine/deep learning solutions and mechanisms through Cloud-based computing infrastructures. Thanks to its ability to remotely execute and train deep learning models (that typically require high computational loads and memory occupation), such an approach guarantees high performance, scalability, and availability. Unfortunately, such an approach requires to send information to be processed (e.g., signals, images, positions, sounds, videos) to the Cloud, hence having potentially catastrophic-impacts on the privacy of users. This paper introduces a novel distributed architecture for deep-learning-as-a-service that is able to preserve the user sensitive data while providing Cloud-based machine and deep learning services. The proposed architecture, which relies on Homomorphic Encryption that is able to perform operations on encrypted data, has been tailored for Convolutional Neural Networks (CNNs) in the domain of image analysis and implemented through a client-server REST-based approach. Experimental results show the effectiveness of the proposed architecture.

Designing adaptive machine learning systems able to operate in nonstationary conditions, also called concept drift, is a novel and promising research area. Convolutional Neural Networks (CNNs) have not been considered a viable solution for such adaptive systems due to the high computational load and the high number of images they require for the training. This paper introduces an adaptive mechanism for learning CNNs able to operate in presence of concept drift. Such an adaptive mechanism follows an "active approach", where the adaptation is triggered by the detection of a concept drift, and relies on the "transfer learning" paradigm to transfer (part of the) knowledge from the CNN operating before the concept drift to the one operating after. The effectiveness of the proposed solution has been evaluated on two types of CNNs and two real-world image benchmarks.

Learning under concept drift is a novel and promising research area aiming at designing learning algorithms able to deal with nonstationary data-generating processes. In this research field, most of the literature focuses on learning nonstationary probabilistic frameworks, while some extensions about learning graphs and signals under concept drift exist. For the first time in the literature, this paper addresses the problem of learning discrete-time Markov chains (DTMCs) under concept drift. More specifically, following a hybrid active/passive approach, this paper introduces both a family of change-detection mechanisms (CDMs), differing in the required assumptions and performance, for detecting changes in DTMCs and an adaptive learning algorithm able to deal with DTMCs under concept drift. The effectiveness of both the proposed CDMs and the adaptive learning algorithm has been extensively tested on synthetically generated experiments and real data sets.

Reducing the computational load of Convolutional Neural Networks (CNNs) is of utmost importance to allow their execution in computing systems characterized by constraints on computation and energy (e.g., embedded and cyber-physical systems and Internet-of-Things). To address this problem, which has been rarely addressed in the related literature, this paper introduces the Gate-Classification CNNs. The core of this novel family of CNNs is the presence of Gate-Classification layers that allow to incrementally process the input image through the CNN layers and take a decision as soon as “enough confidence” about the classification is gained, hence not requiring the processing of the whole CNN when not needed. The Gate-Classification CNNs rely on the ability of CNNs to process features characterized by increasing complexity and meaning and, in particular, the Gate-Classification layers allow to select the path within the CNN according to the information content provided by the input image and the processed features. A wide experimental campaign on public-available datasets supports the effectiveness of the proposed solution.

Execution of deep learning solutions is mostly restricted to high performing computing platforms, e.g., those endowed with GPUs or FPGAs, due to the high demand on computation and memory such solutions require. Despite the fact that dedicated hardware is nowadays subject of research and effective solutions exist, we envision a future where deep learning solutions -here Convolutional Neural Networks (CNNs)- are mostly executed by low-cost off-the shelf embedded platforms already available in the market. This paper moves in this direction and aims at filling the gap between CNNs and embedded systems by introducing a methodology for the design and porting of CNNs to limited in resources embedded systems. In order to achieve this goal we employ approximate computing techniques to reduce the computational load and memory occupation of the deep learning architecture by compromising accuracy with memory and computation. The proposed methodology has been validated on two well-know CNNs, i.e., AlexNet and VGG-16, applied to an image-recognition application and ported to two relevant off-the-shelf embedded platforms.

Intelligent buildings are typically endowed with sensing devices that are able to measure the concentration of specific contaminants in relevant zones. The collected measurements are subsequently processed by intelligent algorithms in order to enable the prompt detection and isolation of contaminant sources inside the building. Unfortunately, in real-world conditions, these sensing devices may suffer from faults affecting the sensors or the embedded electronics. Such faults, generally result in perturbed or missed data in the acquired data-stream, that can induce false alarms (or possibly missed alarms) and compromise the contaminant detection and isolation ability. This paper proposes a three-layer cognitive monitoring system for the detection and isolation of both contaminants and sensor faults in intelligent buildings. The first two layers are designed for the prompt detection of small variations in the concentration of a specific contaminant, while reducing the possible occurrence of false alarms. At the third layer, a cognitive mechanism employing a propagation model for the contaminant, which is based on the airflows between the building zones, allows to isolate the source zone and discriminate between sensor faults and the presence of a contaminant source. The proposed method is validated using a realistic 14-zone building scenario.

Self-adaptive Cyber-Physical Systems (CPSs) enrich CPSs functionalities by introducing self-configuration, self-management, and self-healing skills. Such skills, which are crucial to support adaptation mechanisms, take advantage of the ability to detect changes in the acquired datastreams, e.g., induced by faults affecting sensors/actuators or time-variant environments. In turn, change detection permits CPSs to enable adaptive mechanisms such as reconfiguration of some functionalities to track or mitigate the effect of the change. This paper introduces a novel methodology together with a technological implementation specifically designed for detecting changes affecting the sensor acquisitions in units of CPSs. The methodology requires: 1) learning the signal model; 2) design a model-free change detection test; 3) design a change-point method to validate the detected change. A technological implementation of the proposed methodology encompassing linear predictive models, the ICI-based change detection test and the Mann-Whitney change-point method is introduced and tested on the ST STM32 Nucleo platform. The high detection accuracy altogether with the low computational load and memory occupation make the proposed methodology (and its technological implementation) well suited for self-adaptive CPSs.

In this paper we introduce a novel engineering system for tracking animal movements. Study of animal movement has implications in several relevant and challenging research areas, i.e., from ornithology to global ecology and wildlife management. Inspired by the Internet-of-things vision, the proposed tracking system relies on GSM-based tracking devices that are directly connected to the GSM network to remotely transmit/exchange data with Application Servers/Cloud. The novel characteristic of the proposed GSM-based tracking system is the use of the GSM network for both localization and transmission. This allows to simplify the design of both hardware and software, hence reducing, size, weight and cost of the GSM-based tracking device as well as the fieldwork effort necessary to obtain tracking data. Moreover, this joint localization/transmission phase allows to reduce the energy consumption of the GSM-based tracking device, hence prolonging the life-time of the system and increasing the Quality-of-Service of the envisaged tracking application. We performed a detailed energy assessment of the GSM-based tracking device, while the whole GSM-based tracking system has been tested in a real deployment on four greater flamingos (Phoenicopterus roseus) for approximately 11 months in Northern Italy.

Cognitive fault detection and diagnosis systems are systems able to provide timely information about possibly occurring faults without requiring any a priori knowledge about the process generating the data or the possible faults. This ability is crucial in sensor network scenarios where a priori information about the data generating process, the noise level or the dictionary of the possibly occurring faults is generally hard to obtain. We here present a novel cognitive fault detection and isolation system for sensor networks. The proposed solution relies on the modeling of spatial and temporal relationships present in the acquired datastreams and an ensemble of Hidden Markov Model change-detection tests working in the space of estimated parameters for fault detection and isolation purposes. The effectiveness of the proposed solution has been evaluated on both synthetically generated and real datasets.

We present hierarchical change-detection tests (HCDTs), as effective online algorithms for detecting changes in datastreams. HCDTs are characterized by a hierarchical architecture composed of a detection layer and a validation layer. The detection layer steadily analyzes the input datastream by means of an online, sequential CDT, which operates as a low-complexity trigger that promptly detects possible changes in the process generating the data. The validation layer is activated when the detection one reveals a change, and performs an offline, more sophisticated analysis on recently acquired data to reduce false alarms. Our experiments show that, when the process generating the datastream is unknown, as it is mostly the case in the real world, HCDTs achieve a far more advantageous tradeoff between false-positive rate and detection delay than their single-layered, more traditional counterpart. Moreover, the successful interplay between the two layers permits HCDTs to automatically reconfigure after having detected and validated a change. Thus, HCDTs are able to reveal further departures from the postchange state of the data-generating process.

This paper presents a model-free method for the online identification of sensor faults and learning of their fault dictionary. The method, designed having in mind Cyber-Physical Systems (CPSs), takes advantage of functional relationships among the datastreams acquired by CPS sensing units. Existing model-free change detection mechanisms are proposed to detect faults and identify the fault type thanks to a fault dictionary which is built over time. The main features of the proposed algorithm are its ability to operate without requiring any a priori information about the system under inspection or the nature of the possibly occurring faults. As such, the method follows the model-free approach, characterized by the fact the fault dictionary is constructed online once faults are detected. Whenever available, humans can be considered in the loop to label a fault or a fault class in the dictionary as well as introduce fault instances generated thanks to a priori information. Experimental results on both synthetic and real datasets corroborate the effectiveness of the proposed fault diagnosis system.

Recent studies have shown that the IEEE 802.15.4 MAC protocol suffers from severe limitations, in terms of reliability and energy efficiency, when the CSMA/CA parameter setting is not appropriate. However, selecting the optimal setting that guarantees the application reliability requirements, with minimum energy consumption, is not a trivial task in wireless sensor networks, especially when the operating conditions change over time. In this paper we propose a Just-in-Time LEarning-based Adaptive Parameter tuning (JIT-LEAP) algorithm that adapts the CSMA/CA parameter setting to the time-varying operating conditions by also exploiting the past history to find the most appropriate setting for the current conditions. Following the approach of active adaptive algorithms, the adaptation mechanism of JIT-LEAP is triggered by a change detection test only when needed (i.e., in response to a change in the operating conditions). Simulation results show that the proposed algorithm outperforms other similar algorithms, both in stationary and dynamic scenarios.

In recent years, Radio frequency (RF) sensor networks have been used to localize people indoor without requiring them to wear invasive electronic devices. These wireless mesh networks, formed by low-power radio transceivers, continuously measure the received signal strength (RSS) of the links. Radio Tomographic Imaging (RTI) is a technique that generates, starting from these RSS measurements, 2D images of the change in the electromagnetic field inside the area covered by the radio transceivers to spot the presence and movements of animates (e.g., people, large animals) or large metallic objects (e.g., cars). Here, we present a RTI system for localizing and tracking people outdoors. Differently than in indoor environments where the RSS does not change significantly with time unless people are found in the monitored area, the outdoor RSS signal is time-variant, e.g., due to rainfall or wind-driven foliage. We present a novel outdoor RTI method that, despite the nonstationary noise introduced in the RSS data by the environment, achieves high localization accuracy and dramatically reduces the energy consumption of the sensing units. Experimental results demonstrate that the system accurately detects and tracks a person in real-time in a large forested area under varying environmental conditions, significantly reducing false positives, localization error and energy consumption compared to state-of-the-art RTI methods.

Intelligent buildings are equipped with sensing systems able to measure the contaminant concentration in the different building zones for safety purposes. The aim of these systems is to promptly detect the presence of a contaminant so that appropriate actions can be taken to ensure the safety of the people. At the same time, these sensing systems, which operate in real-world conditions, suffer from noise and sensor degradation faults. Both noise and faults can induce false alarms (resulting in unnecessary disruptive actions such as building evacuation) or missed alarms (when the presence of a contaminant is not detected). This paper proposes a novel cognitive monitoring system for performing contaminant detection in intelligent buildings with real-time point-trigger sensors. The proposed system reduces the occurrence of false alarms by means of a three-layered architecture, which employs cognitive mechanisms to validate possible detections and discriminate between the presence of a real contaminant source and a degradation fault affecting the sensors of the sensing system. In addition, the proposed system is able to isolate the building zone containing the contaminant source (or the faulty sensor) and estimate the onset time of the release (or the fault).

Time-series data are often characterized by a large degree of self-similarity, which arises in application domains featuring periodicity or seasonality. While self-similarity has shown to be an effective prior for modeling real data in the signal and image-processing literature, it has received much less attention in time-series literature, where only few works leveraging the self-similarity for anomaly detection have been presented. Here we introduce a novel change-detection test to detect structural changes in time series by analyzing their self-similarity. The core of the proposed solution is the definition of a change indicator to quantitatively assesses the self-similarity of the time-series data over time. In particular, the change indicator is obtained by comparing each patch to be analyzed with its most similar counterpart in a change-free training set. Experimental results on the flow measurements in the water distribution network of the Barcelona city show the effectiveness of the proposed solution.

Cognitive fault diagnosis systems differentiate from more traditional solutions by providing online strategies to create and update the fault-free and the faulty classes directly from incoming data. This aspect is of paramount relevance within the big data framework, since measurements are there immediately processed to detect and identify the upsurge of potential faults. This paper introduces a novel cognitive fault diagnosis framework for processes described by nonlinear dynamic systems that inspects changes in the existing relationships among sensors. The proposed framework is based on an evolving clustering algorithm that operates in the parameter space of time invariant linear models approximating such relationships. During the operational life, parameter vectors associated with models thought not to belong to the nominal state are either labeled as outlier or fault. New classes of faults, here considered to propagate to the model parameters according to an abrupt profile, are created online as they appear. At the same time, existing classes can merge, depending on the information content carried by incoming data.

The literature has shown that model ensemble techniques are particularly effective to solve regression/classification applications by providing, given a suitable aggregation mechanism, a better generalization ability than the generic model of the ensemble. However, only few recent results consider the use of ensembles for a time-dependent framework, with focus on time-series forecasting. Here, we propose the use of ensemble of models to an on-line reconstruction of missing data coming from a sensor network. Reconstructing missing data is of paramount importance for any further data processing and must be carried out on-line not to introduce unnecessary latency when data lead to a decision or control action. The ensemble is designed by both exploiting temporal and spatial dependencies existing among the sensors composing the network. An effective aggregation mechanism is proposed for the considered models to improve the generalization ability of the ensemble. Results demonstrate the effectiveness of the proposed approach in reconstructing missing data.

Just-in-time (JIT) classifiers operate in evolving environments by classifying instances and reacting to concept drift. In stationary conditions, a JIT classifier improves its accuracy over time by exploiting additional supervised information coming from the field. In nonstationary conditions, however, the classifier reacts as soon as concept drift is detected; the current classification setup is discarded and a suitable one activated to keep the accuracy high. We present a novel generation of JIT classifiers able to deal with recurrent concept drift by means of a practical formalization of the concept representation and the definition of a set of operators working on such representations. The concept-drift detection activity, which is crucial in promptly reacting to changes exactly when needed, is advanced by considering change-detection tests monitoring both inputs and classes distributions.

This paper introduces a novel cognitive fault diagnosis system (FDS) for distributed sensor networks that takes advantage of spatial and temporal relationships among sensors. The proposed FDS relies on a suitable functional graph representation of the network and a two-layer hierarchical architecture designed to promptly detect and isolate faults. The lower processing layer exploits a novel change detection test (CDT) based on hidden Markov models (HMMs) configured to detect variations in the relationships between couples of sensors. HMMs work in the parameter space of linear time-invariant dynamic systems, approximating, over time, the relationship between two sensors; changes in the approximating model are detected by inspecting the HMM likelihood. Information provided by the CDT layer is then passed to the cognitive one, which, by exploiting the graph representation of the network, aggregates information to discriminate among faults, changes in the environment, and false positives induced by the model bias of the HMMs.

High-frequency sampling is not only a prerogative of high-energy physics or machinery diagnostic monitoring: critical environmental and structural health monitoring applications also have such a challenging constraint. Moreover, such unique design constraints are often coupled with the requirement of high synchronism among the distributed acquisition units, minimal energy consumption, and large communication bandwidth. Such severe constraints have led scholars to suggest wired centralized monitoring solutions, which have only recently been complemented with wireless technologies. This article suggests a hybrid wireless-wired monitoring system combining the advantages of wireless and wired technologies within a distributed high-frequency-sampling framework. The suggested architecture satisfies the mentioned constraints, thanks to an ad-hoc design of the hardware, the availability of efficient energy management policies, and up-to-date harvesting mechanisms. At the same time, the architecture supports adaptation capabilities by relying on the remote reprogrammability of key application parameters. The proposed architecture has been successfully deployed in the Swiss-Italian Alps to monitor the collapse of rock faces in three geographical areas.

Handling dynamic environments and building up algorithms operating at low supervised-sample rates are two main challenges for classification systems designed to operate in real-life scenarios. Here, changes in the probability density function of classes characterizing the data-generating process (also called concept drift) should be detected as soon as possible to prevent the classifier from becoming obsolete. Moreover, when the rate of supervised samples during the operational life is low (as in those situations where the sample inspection is costly or destructive) both detecting the change and re-training the classifier become even more critical aspects. We present an adaptive classifier that exploits both supervised and unsupervised data to monitor the process stationarity. The classifier follows the just-in-time (JIT) approach and relies on two different change-detection tests (CDTs) to reveal changes in the environment and reconfigure the classifier accordingly. The proposed solution assesses the stationary in both the joint probability density function (CDT at the classification error) and the distribution of the inputs (CDT on unlabeled data). In addition, we integrate in the JIT adaptive classifier a procedure able to handle recurrent concepts within an ensemble of classifiers framework. Experiments show that monitoring unsupervised samples and handling recurrent concepts is essential for classifying in non-stationary environments when few supervised samples are available.

Data streams from remote monitoring systems such as wireless sensor networks show immediately that the “you sample you get” statement is not always true. Not rarely, the data stream is interrupted by intermittent communication or sensors faults, resulting in missing data in the received sequence. This has a negative impact in many algorithms assuming continuous data stream; as such, the missing data must be suitably reconstructed, in order to guarantee continuous data availability. We suggest a general methodology for reconstructing missing data that exploits both temporal and spatial redundancy characterizing the phenomenon being monitored and the distributed system, a situation proper of many monitoring systems constituted by sensor and actuator networks. Temporal and spatial dependencies are learned through linear and non-linear non-parametric models, also encompassing neural -possibly recurrent- networks, which become the spatial transfer functions connecting the different views of the phenomenon under investigation. Missing data are finally reconstructed by exploiting the forecasting ability provided by such transfer functions. The experimental section shows the effectiveness of the proposed methodology.

In this work we address the problem of automatically detecting changes either induced by faults or concept drifts in data streams coming from multi-sensor units. The proposed methodology is based on the fact that the relationships among different sensor measurements follow a probabilistic pattern sequence when normal data, i.e. data which do not present a change, are observed. Differently, when a change in the process generating the data occurs the probabilistic pattern sequence is modified. The relationship between two generic data streams is modelled through a sequence of linear dynamic time-invariant models whose trained coefficients are used as features feeding a Hidden Markov Model (HMM) which, in turn, extracts the pattern structure. Change detection is achieved by thresholding the log-likelihood value associated with incoming new patterns, hence comparing the affinity between the structure of new acquisitions with that learned through the HMM. Experiments on both artificial and real data demonstrate the appreciable performance of the method both in terms of detection delay, false positive and false negative rates.

The recent increase in number and complexity of wireless sensor networks (WSN)-based deployments made evident the limits of traditional hardware platforms designed to work in a controlled environment (laboratory) for a limited amount of time. The need for high-performance, low-power, flexible and scalable hardware platforms able to work in real-world (possibly harsh) environments led us to design and develop the NetBrick platform. The novelty and the advantages of the proposed platform w.r.t. other existing hardware platforms reside in: 1) flexibility at the board level (each module composing the board can be enabled/disabled by the software) 2) flexibility at the network level (the NetBrick natively allows for creating wireless, wired and hybrid networks); 3) the high performance guaranteed by the 32bit microprocessor at a very reasonable power consumption thanks to the ultra-low power Cortex M3 architecture; 4) the fine-grain energy management of the board modules (each module provides information about its power consumption), hence allowing the designer for defining advanced energy management policies. The NetBrick platform has been tested with success in a distributed monitoring system for landslide forecasting designed and developed by our group and deployed in the Alps (north Italy).

Classification systems designed to work in nonstationary conditions rely on the ability to track the monitored process by detecting possible changes and adapting their knowledge-base accordingly. Adaptive classifiers present in the literature are effective in handling abrupt concept drifts (i.e., sudden variations), but, unfortunately, they are not able to adapt to gradual concept drifts (i.e., smooth variations) as these are, in the best case, detected as a sequence of abrupt concept drifts. To address this issue we introduce a novel adaptive classifier that is able to track and adapt its knowledge base to gradual concept drifts (modeled as polynomial trends in the expectations of the conditional probability density functions of input samples), while maintaining its effectiveness in dealing with abrupt ones. Experimental results show that the proposed classifier provides high classification accuracy both on synthetically generated datasets and measurements from real sensors.

Design of applications working in nonstationary environments requires the ability to detect and anticipate possible behavioral changes affecting the system under investigation. In this direction, the literature provides several tests aiming at assessing the stationarity of a data generating process; of particular interest are nonparametric sequential change-point detection tests that do not require any a-priori information regarding both process and change. Moreover, such tests can be made automatic through an on-line inspection of sequences of data, hence making them particularly interesting to address real applications. Following this approach, we suggest a novel two-level hierarchical change-detection test designed to detect possible occurrences of changes by observing incoming measurements. This hierarchical solution significantly reduces the number of false positives at the expenses of a negligible increase of false negatives and detection delays. Experiments show the effectiveness of the proposed approach both on synthetic dataset and measurements from real applications.

Designing tests able to effectively detect changes in the stationarity of a process generating data is a challenging problem, in particular when the process is unknown, and the only information available has to be extracted from a set of observations. This work proposes a novel approach for detecting changes in a process generating data whose distribution is unknown. Peculiarity of the approach is the use of the Intersection of Confidence Intervals (ICI) rule to monitor the process evolution. A change detection test derived from this approach is also presented. Experimental results show that the proposed test outperforms state-of-the art solutions, both in terms of efficiency and effectiveness, in particular when a reduced test configuration set is available.

LOO and k-fold cross validation are widely used validation methods assessing the accuracy of a model at the expenses of a high computational load (several models need to be trained and performance averaged). To mitigate such phenomenon a virtual LOO method has been suggested in which, by relying on the concept of leverages, provides the LOO estimate of the generalization error in a closed form without the need to re-training different models. In this paper, we extend and generalize such an approach by introducing the virtual k-fold cross validation method which provides a k-fold cross validation estimate without requiring training multiple models. Results, correct for linear models, are approximations for nonlinear ones. Simulation results show the effectiveness of the proposed virtual method which can be suitably extended to cover different figures of merit and performance assessment techniques.

Rock face collapses are one of the most dangerous and sudden natural risks in mountain environments. Traditional investigation techniques (e.g., strain gauge, inclinometer) cannot guarantee neither a non-invasive monitoring within the rocks nor a prompt alarm in case of possible collapse. Real-time monitoring systems, which might provide an effective evaluation of the dynamics of the phenomenon for a subsequent forecasting phase, generally rely on wired solutions that are unfeasible in environmental monitoring applications. In this paper, we present a real-time monitoring system for the rock collapse forecasting that exploits MEMS accelerometers and geophones (in addition to traditional sensors) for a non-invasive detection of micro-acoustic bursts associated with the formation and the evolution of the cracks within the rocks. The proposed monitoring system relies on an hybrid wireless-wired architecture that allows for detecting and localizing micro-acoustic emissions in real-time yet maintaining an high energy-efficiency by means of effective energy management policies and sophisticated energy harvesting mechanisms. The deployment area is the St. Martin mountain that dangerously insists on the town of Lecco (Italy).

A classifier expected to work in a non-stationary environment has to: (i) detect changes in the process generating the data; (ii) suitably react to the change by adapting to the new working condition. Just-in-time adaptive classifiers, a classification structure addressing stationary and nonstationary conditions, have been presented to the computational intelligence community. Such classifiers require a temporal detection of a (possible) process deviation followed by an adaptive management of the knowledge base characterizing the classifier to cope with the process change. This paper improves just-in-time adaptive classifiers by integrating temporal information about the state of the process under monitoring. An index for the process deviation is defined which, coupled with an adaptive weighted k-NN classifier, shows to be particularly effective in dealing with smooth process drifts and ageing phenomena.

The paper proposes a theory-based method for estimating the optimal value of k in k-NN classifiers based on a n-sized training set. As expected, experiments show that the suggested k is such that k/n rarr 0 when both k and n tend to infinity, as required by the asymptotical consistency condition. Interestingly, it appears that the generalization error is robust w.r.t. to k when n becomes large (probably as a consequence of the k/n rarr 0 relationship); the immediate consequence is that there is no need to provide an accurate estimate for the optimal k and an approximated coarser value, e.g., provided with cross validation, 1-fold cross validation or leave one out is more than adequate.

The stationarity requirement for the process generating the data is a common assumption in classifiers' design. When such hypothesis does not hold, e.g., in applications affected by aging effects, drifts, deviations, and faults, classifiers must react just in time, i.e., exactly when needed, to track the process evolution. The first step in designing effective just-in-time classifiers requires detection of the temporal instant associated with the process change, and the second one needs an update of the knowledge base used by the classification system to track the process evolution. This paper addresses the change detection aspect leaving the design of just-in-time adaptive classification systems to a companion paper. Two completely automatic tests for detecting nonstationarity phenomena are suggested, which neither require a priori information nor assumptions about the process generating the data. In particular, an effective computational intelligence-inspired test is provided to deal with multidimensional situations, a scenario where traditional change detection methods are generally not applicable or scarcely effective.

Aging effects, environmental changes, thermal drifts, and soft and hard faults affect physical systems by changing their nature and behavior over time. To cope with a process evolution adaptive solutions must be envisaged to track its dynamics; in this direction, adaptive classifiers are generally designed by assuming the stationary hypothesis for the process generating the data with very few results addressing nonstationary environments. This paper proposes a methodology based on k-nearest neighbor (NN) classifiers for designing adaptive classification systems able to react to changing conditions just-in-time (JIT), i.e., exactly when it is needed. k-NN classifiers have been selected for their computational-free training phase, the possibility to easily estimate the model complexity k and keep under control the computational complexity of the classifier through suitable data reduction mechanisms. A JIT classifier requires a temporal detection of a (possible) process deviation (aspect tackled in a companion paper) followed by an adaptive management of the knowledge base (KB) of the classifier to cope with the process change. The novelty of the proposed approach resides in the general framework supporting the real-time update of the KB of the classification system in response to novel information coming from the process both in stationary conditions (accuracy improvement) and in nonstationary ones (process tracking) and in providing a suitable estimate of k. It is shown that the classification system grants consistency once the change targets the process generating the data in a new stationary state, as it is the case in many real applications.

Integrating new information in classification systems during their operational life requires adaptive mechanisms able to identify first the presence of valuable information and update then the knowledge base onto which the classifier is configured. In this paper we provide a design solution for adaptive classifiers operating in stationary environments; information provided (whenever available by a supervisor over time) is used to improve the performance of the classification system hence mimicking the asymptotical behavior suggested by the theory. The adaptive classifier relies on k -NNs, here chosen for their learning-free modality (hence easily supporting a real time adaptation mechanism); a novel method is proposed for matching the optimal k (measuring the complexity of the classifier) with the incremental knowledge acquired over time. A large experimental campaign shows the effectiveness of the proposed approach.

In real world applications ageing effects, process drifts, soft and hard faults may affect the data generation mechanism and, as a consequence, data coming from it. Intelligent measurement systems developed for such processes (e.g., industrial quality assessment and control, environmental monitoring) require adaptive techniques which, by tracking the system evolution, allow the intelligent system for keeping acceptable performance. Here we focus on adaptive classifiers embedded in intelligent measurement systems designed to cope with non-stationary environments, yet well performing in stationary conditions. The novelty of the approach resides in the possibility to update in a just-in-time fashion, i.e., only when it is really needed, the knowledge base of the classifier. A large experimental campaign shows the effectiveness of the proposed design.

Energy conservation techniques for sensor networks typically rely on the assumption that data sensing and processing consume considerable less energy than communication. This assumption does not hold in some practical application scenarios, where ad hoc developed sensor units require power consumption comparable with, or even larger than, that of the radio. In this paper we focus on an embedded sensor for monitoring snow composition in mountain slopes for avalanche forecasting. To lower the sensor energy consumption we propose an adaptive sampling algorithm able to dynamically estimate the optimal sampling frequency of the signal to be monitored. In turn, this minimizes the activity of both the sensor and the radio (hence saving energy) while maintaining an acceptable accuracy on the acquired data. Simulation experiments show that the suggested solution can save up to 97% of the energy consumed for sensing when the sensor is always on, while maintaining the error at acceptable levels.

The stationarity hypothesis is largely and implicitly assumed when designing classifiers (especially those for industrial applications) but it does not generally hold in practice. The paper goal is to provide an automatic, general purpose, easy to use and effective index for estimating deviations, drifts or ageing effects in the process generating the data (e.g., classifier inputs); in turns this will allow the designer for identifying when to intervene to update the knowledge space of adaptive classifiers. More specifically, we suggest a robust extension of the adaptive CUSUM test procedure which addresses a set of features (in contrast to the literature which considers a single feature) for detecting drifts. The application of the change detection test to real applications shows that its real additional value resides in the ability to detect continuous and small drifts, a critical situation for traditional tests.

Many applications, e.g., fault detection, quality of industrial process, monitoring and prediction of climatic phenomena assume the stationary hypothesis or require identification of the process change. Change detection tests satisfy the change detection necessity by identifying a drift, a different expected behavior, a deviation; their effectiveness is generally based on statistical confidence tests whose parameters are configured at design-time (generally through a trial-and-error approach). Here, we suggest an extension of the widely-used CUSUM change detection test which improves effectiveness and timeliness in detecting changes by adaptively configuring its test parameters

The large amount of documents to be daily managed in modern offices requires development of automatic document classification tools aiming at (semi)automatically classifying the office documents into semantically similar classes. This paper presents an automatic invoice-documents classification system based on the analysis of the graphical information present in the document and able to perform both closed (the number of classes is fixed) and open world (the number of classes increases during operational life) classification. Invoice-documents of real companies prove that the classification system achieves a 99% correct classification in closed world and 79% in the open world case.

The automatic visual inspection (AVI) systems are nowadays used with good results in a wide broad of applications. In particular we considered the problem of the defect detection of particle boards by means of a visual inspection of the printed surface. We propose an innovative defect detection approach capable to automatically extract the repetitive patterns that are generally present in the printed-matters. The extracted patterns are then used in the proper defect detection phase that can thus achieve high defect detection performances. Real wood patterns prove that our defect detection system achieves very high classification capability.

The development of ambient intelligence (AmI) applications that effectively adapt to the needs of the users and environments requires, among other things, the presence of planning mechanisms for goal-oriented behavior. Planning is intended as the ability of an AmI system to build a course of actions that, when carried out by the devices in the environment, achieve a given goal. The problem of planning in AmI has not yet been adequately explored in literature. We propose a planning system for AmI applications, based on the hierarchical task network (HTN) approach and called distributed hierarchical task network (D- HTN), able to find courses of actions to address given goals. The plans produced by D-HTN are flexibly tailored to exploit the capabilities of the devices currently available in the environment in the best way. We discuss both the architecture and the implementation of D-HTN. Moreover, we present some of the experimental results that validated the proposed planner in a realistic application scenario in which an AmI system monitors and answers the needs of a diabetic patient.