Endowing artificial objects with intelligence is a longstanding computer science and engineering vision that recently converged under the umbrella of Artificial Intelligence of Things (AIoT). Nevertheless, AIoT’s mission cannot be fulfilled if objects rely on the cloud for their “brain,” at least concerning inference. Thanks to heterogeneous hardware, it is possible to bring Machine Learning (ML) inference on resource-constrained embedded devices, but this requires careful co-optimization between model training and its hardware acceleration. This work proposes YoseUe, a memory-centric hardware co-processor for Random Forests (RFs) inference, which significantly reduces the waste of memory resources by exploiting a novel train-acceleration co-optimization. YoseUe proposes a novel ML model, the Multi-Depth Random Forest Classifier (MDRFC), in which a set of RFs are trained at decreasing depths and then weighted, exploiting a Neural Network (NN) tailored to counteract potential accuracy losses w.r.t. classical RFs. With this modeling technique, first proposed in this paper, it becomes possible to accelerate the inference of RFs that count up to 2 orders of magnitude more Decision Trees (DTs) than those the current state-of-the-art architectures can fit on embedded devices. Furthermore, this is achieved without losing accuracy with respect to classical, full-depth RF in their most relevant configurations.

Modern FPGA HLS tools are proficient at accelerating datapath applications, but they generate considerable overhead when dealing with irregular, control-driven workloads. Conversely, RTL-based approaches significantly increase development time and system integration effort. To address this tooling gap, we present SATL, a Chisel-based rapid prototyping framework for building FPGA-based spatial architectures targeting irregular workloads. We use it to re-implement YoseUe, a state-of-the-art accelerator for inferring Decision Tree Ensemble Machine Learning models. Compared to the original HLS-based work, SATL yields an average 3.4 x throughput improvement and reduces the architecture's resource consumption, allowing inference on Ensemble Models with up to 3.6 x more trees, enabling the deployment of larger models on resource-constrained devices.

The Artificial Intelligence of Things (AIoT) calls for on-site Machine Learning inference to overcome the instability in latency and availability of networks. Thus, hardware acceleration is paramount for reaching the Cloud's modeling performance within an embedded device's resources. In this paper, we propose Entree, the first automatic design flow for deploying the inference of Decision Tree (DT) ensembles over Field-Programmable Gate Arrays (FPGAs) at the network's edge. It exploits dynamic partial reconfiguration on modern FPGA-enabled Systems-on-a-Chip (SoCs) to accelerate arbitrarily large DT ensembles at a latency a hundred times stabler than software alternatives. Plus, given Entree's suitability for both hardware designers and non-hardware-savvy developers, we believe it has the potential of helping data scientists to develop a non-Cloud-centric AIoT.

The Cox Proportional Hazards regression is among the most widely used models in clinical and epidemiological research for investigating the association between time-to-event outcomes and multiple predictors, that, in the modern perspective of personalized medicine, tend to belong to ever wider spheres relating to the patient and his medical condition. When the goal is to include a large number of variables in a prediction model, feature selection techniques are often required to ensure a certain level of interpretability of the results and federated learning is necessary to recruit in the study the sufficient number of patients for reliable model outcomes, overcoming the main problems of data privacy and ownership. In this regard, we here propose an adaptation for federated learning of the optimization algorithm of the Cox Proportional Hazards regression model with LASSO regularization as feature selector and we demonstrate the efficacy of our algorithm on real and simulated data sets in a simulated distributed environment with no patient-level data sharing by comparing its model parameter estimation performances with its centralised version.

The widespread application of specialized embedded devices and their ever-growing sensing capabilities lead to an overload of raw data reaching the Cloud via the Internet of Things. Despite this massive amount of available data, data-driven applications still produce a limited impact on the real world. We argue that this limitation arises from the complexity the developers still face when engineering and deploying solutions on a plurality of embedded devices. Instead of focusing on constructing valuable information from raw data and designing algorithms to exploit it, the developers still have to figure out on their own: how to partition the computation between Edge, Fog, and Cloud; how to move partial results around in highly distributed solutions; how to create abstract interfaces to expose information-rich endpoints. This paper introduces the Virtual Sensor Domain-Specific Language (DSL), which builds the keystone for designing and developing powerful Computer-Aided Design tools and Integrated Development Environments to support the developers in focusing on information manipulation. The Virtual Sensor DSL is built on top of C++ to ease its adoption by the embedded development community. It introduces a reactive approach to data collection, technology mapping, and workload distribution. It allows the creation of adaptable Virtual Sensors, streaming processed information coming from the integration of multiple raw data streams, and it comes with integrated simulation capabilities, supporting prototyping. In this paper, we describe the entities the language offers, how the developers can exploit them, and analyze a real-world use case that benefitted from the Virtual Sensor DSL.

After the COVID-induced lock-downs, augmented/virtual reality turned from leisure to desired reality. Real-time 3D audio is a crucial enabler for these technologies. Nevertheless, systems offering object spatialization in 3D audio fall in two limited cases. They either require long-running pre-renders or involve powerful computing platforms. Furthermore, they mainly focus on active audio sources, while humans rely on the sound's interactions with passive obstructions to sense their environment. We propose a hardware co-processor for real-time 3D audio spatialization supporting passive obstructions. Our solution attains similar latency w.r.t. workstations while draining a tenth of the power, making it suitable for embedded applications.

“Cloud-native” is the umbrella adjective describing the standard approach for developing applications that exploit cloud infrastructures’ scalability and elasticity at their best. As the application complexity and user-bases grow, designing for performance becomes a first-class engineering concern. As an answer to these needs, heterogeneous computing platforms gained widespread attention as powerful tools to continue meeting SLAs for compute-intensive cloud-native workloads. We propose BlastFunction, an FPGA-as-a-Service full-stack framework to ease FPGAs’ adoption for cloud-native workloads, integrating with the vast spectrum of fundamental cloud models. At the IaaS level, BlastFunction time-shares FPGA-based accelerators to provide multi-tenant access to accelerated resources without any code rewriting. At the PaaS level, BlastFunction accelerates functionalities leveraging the serverless model and scales functions proactively, depending on the workload’s performance. Further lowering the FPGAs’ adoption barrier, an accelerators’ registry hosts accelerated functions ready to be used within cloud-native applications, bringing the simplicity of a SaaS-like approach to the developers. After an extensive experimental campaign against state-of-the-art cloud scenarios, we show how BlastFunction leads to higher performance metrics (utilization and throughput) against native execution, with minimal latency and overhead differences. Moreover, the scaling scheme we propose outperforms the main serverless autoscaling algorithms in workload performance and scaling operation amount.

Scientific and clinical research have advanced the ability of healthcare professionals to more precisely define diseases and classify patients into different groups based on their likelihood of responding to a given treatment, and on their future risks. However, a significant gap remains between the delivery of stratified healthcare and personalization. The latter implies solutions that seek to treat each citizen as a truly unique individual, as opposed to a member of a group with whom they share common risks or health-related characteristics. Personalisation also implies an approach that takes into account personal characteristics and conditions of individuals. This paper investigates how these desirable attributes can be developed and introduces a holistic environment, the iHELP, that incorporates big data management and Artificial Intelligence (AI) approaches to enable the realization of data-driven pathways where awareness, care and decision support is provided based on person-centric early risk prediction, prevention and intervention measures.

Process Mining is of growing importance in the healthcare domain, where the quality of delivered services depends on the suitable and efficient execution of processes encoding the vast amount of clinical knowledge gained via the evidence-based medicine paradigm. In particular, to assess and measure the quality of delivered treatments, there is a strong interest in tools able to perform conformance checking. In process mining for the healthcare domain, a number of major challenges are posed by: (i) the complexity of involved data, that refers to patients' aspects such as disease, behaviour, clinical history, psychology, etc; (ii) the availability of data, that come from the heterogeneous, fragmented and scant connected healthcare system; and (iii) the wide range of available standards for communication (DICOM, IHE, etc.) or data representation (ICD9, SNOMED, etc.) purposes. To effectively perform process mining in the healthcare domain, it is crucial to build event logs capturing all the steps of running processes, which have to be derived by the knowledge stored in the Electronic Health Records. It is therefore crucial to cope with aforementioned data-related challenges. In this paper, we aim at supporting the exploitation of process mining in the healthcare domain, particularly with regards to conformance checking. We therefore introduce a set of specifically-designed techniques, provided as a suite of software packages written in R. In particular, the suite provides a flexible and agile way to automatically and reliably build Event Log from clinical data sources, and to effectively perform conformance checking.

Autonomicity is a golden feature when dealing with a high level of complexity. This complexity can be tackled partitioning huge systems in small autonomous modules, i.e., agents. Each agent then needs to be capable of extracting knowledge from its environment and to learn from it, in order to fulfill its goals: this could not be achieved without proper modeling techniques that allow each agent to gaze beyond its sensors. Unfortunately, the simplicity of agents and the complexity of modeling do not fit together, thus demanding for a third party to bridge the gap. Given the opportunities in the field, the main contributions of this work are twofold: (1) we propose a general methodology to model resource consumption trends and (2) we implemented it into MARC, a Cloud-service platform that produces Models-as-a-Service, thus relieving self-aware agents from the burden of building their custom modeling framework. In order to validate the proposed methodology, we set up a custom simulator to generate a wide spectrum of controlled traces: this allowed us to verify the correctness of our framework from a general and comprehensive point of view.

Process mining focuses on extracting knowledge, under the form of models, from data generated and stored in information systems. The analysis of generated models can provide useful insights to domain experts. In addition, models of processes can be used to test if a considered process complies with some given specifications. For these reasons, process mining is gaining significant importance in the healthcare domain, where the complexity and flexibility of processes makes extremely hard to evaluate and assess how patients have been treated.

Multi-tenant virtualized infrastructures allow cloud providers to minimize costs through workload consolidation. One of the largest costs is power consumption, which is challenging to understand in heterogeneous environments. We propose a power modeling methodology that tackles this complexity using a divide-and-conquer approach. Our results outperform previous research work, achieving a relative error of 2% on average and under 4% in almost all cases. Models are portable across similar architectures, enabling predictions of power consumption before migrating a tenant to a different hardware platform. Moreover, we show the models allow us to evaluate colocations of tenants to reduce overall consumption.

In the last decade we have witnessed the birth and dramatic growth of mobile devices, from cellular-to smart-phones. Despite the huge amount of information achievable from an always-connected reality, researchers that work in the mobile devices field fight against the impossibility to explore, inspect and test their work on such a vast set of possible environments, use case scenarios, hardware and software platforms the smart mobile world is composed of. This pushed the need of a wide open dataset of real world data coming from devices in their real usage context, properly anonymized and conveniently organized to be searchable and accessible. In this paper, we present a platform that brings such a dataset to researchers of the next generation of mobile devices.

Mobile devices take an important part in everyday life. They are now cheaper and widespread, but still a lot of time is spent by the users to configure them: users adapt to their own device, not vice versa. Can our smart phones do something smarter? In this work, we propose a framework to support the development of context aware applications for Android devices: the goal of such applications is to reduce as much as possible the interaction with the user, making use of automatic and intelligent components. Moreover, these components should consume as less power and computational resources as possible, being them part of a mobile ecosystem whose battery and hardware are highly constrained. The work implies the study of a methodology that fits the Android framework and the design of a highly extensible software architecture. An open source framework based on the proposed methodology is then described. Some use cases are finally presented, analyzing the performances and the limitations of the proposed methodology.