NECST Lab Visit: Heterogeneous Systems, Hardware Acceleration, and More; Politecnico di Milano

Speaker: Ian Di Dio Lavore

Abstract: The System and Software Design research area focuses on developing innovative solutions to address the challenges of modern computing systems, where increasing application demands necessitate advanced performance, efficiency, and adaptability. Our research exploits hardware-software co-design approaches, enabling the creation of domain-specific architectures, reconfigurable systems, and frameworks for high-performance computing. We explore novel methods to enhance computational efficiency, such as leveraging specialized architectures and compilers to optimize resource-intensive workloads like pattern matching, sparse applications, and image registration leveraging FPGA-based technologies for hardware acceleration and prototyping. Additionally, we tackle the complexities of heterogeneous and distributed systems, devising abstractions and runtime strategies to simplify programming, optimize resource utilization, and improve performance portability across diverse architectures. By addressing issues such as locality, affinity, and scalability, we aim to deliver comprehensive solutions that empower developers to achieve both productivity and performance in increasingly intricate hardware environments. This integrated approach ensures the adaptability to emerging technologies and positions it at the forefront of advancing system and software design for current and future computational challenges.

Bio: Ian Di Dio Lavore is a Ph.D. student in Information Technology – Computer Science and Engineering at Politecnico di Milano. He holds an M.Sc. (2022) and a B.Sc. (2020) in Computer Science and Engineering from Politecnico di Milano. Ian worked in the HPC Team of the Scalable Computing and Data group at the Pacific Northwest National Laboratory (PNNL). His research mainly focuses on parallel and distributed computing and programming models, with a particular interest in high-level abstractions for heterogeneous HPC systems. 

Speaker: Alessandro Verosimile

Abstract: Numerical system solvers, machine/deep learning, dense/sparse linear algebra, genomics, and physics simulation, just to name a few, are examples of highly performance-demanding applications hungry for computational power. However, modern general-purpose architectures fail to deliver the ever-increasing required performance in such a context, causing long energy-inefficient runtimes. Consequently, both High-Performance Computing (HPC) and embedded systems started embracing heterogeneity and leveraging hardware accelerators. In this way, the adoption of devices like Graphics Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs) alongside general-purpose processors offers a unique solution to offload workloads, boost performance, and increase energy efficiency. This research area showcases our efforts on the acceleration of computationally demanding applications on GPUs and FPGAs.

Bio: Alessandro Verosimile is a second-year PhD student in Information Technology at Politecnico di Milano. He worked for 6 months as a research intern in the RAD team of Advanced Micro Devices (AMD). His research focuses on HW-SW co-design techniques that aim to co-optimize the training of large Machine Learning models and the design of the hardware architecture for their inference on embedded devices, with a focus on both Deep learning models and Decision Tree based ensemble models.

Speaker: Mirko Coggi

Abstract: The exponential growth in high-quality biomedical data has created an urgent need for advanced computational methods to analyze and interpret this wealth of information effectively. The “Computational Methods for Biomedical Data Analysis” area at NECSTLab addresses critical challenges in data analysis following three different directions across multiple scales. (i) In Genome Analysis, this area focuses on optimized and hardware-accelerated (GPU) algorithms for processing massive sequencing datasets, also leveraging novel graph-based pangenomic architectures to overcome the limitations of traditional linear approaches in variant identification and characterization. In addition, (ii) Graph Machine Learning techniques are used to decipher the intricate relationships between DNA and biological pathways, particularly to predict the structural and functional effects of genomic variations on generated proteins. Finally, (iii) Deep Learning architectures for medical image analysis are investigated, combining information from multiple image modalities and scales to make more precise diagnostic, prognostic, and therapeutic determinations. By integrating these computational approaches, this research area advances a holistic understanding of biological systems, driving innovations in precision medicine on multiple levels.

Bio: Mirko Coggi is a third-year PhD Candidate in Information Technology, specifically in Computer Science, but he comes from BSc and MSc in Biomedical Engineering at Politecnico di Milano. Mirko’s work centers on software tools for analyzing pangenome graphs. He has extensive experience with tools like VG Toolkit, GraphAligner, and GedMap, and he developed several tools himself for manipulating genome graphs, for optimizing the alignment task, for comparing aligners, and for parsing and querying on VCFs. Additionally, he co-founded GenoGra, a startup dedicated to creating platforms for genome graph analysis.

Speaker: Laura Ginestretti

Abstract: This research area focuses on developing advanced solutions for distributed ambient monitoring and physiological data analysis, with the goal of optimizing mental and physical health, as well as cognitive and athletic performance. The proposed approach integrates distributed smart sensors in IoT environments, along with data from wearable devices, to collect and analyze physiological and environmental parameters in real time. This enables both the exploration of how lifestyle-related factors, such as nutrition, sleep, and other variables, and environmental conditions influence cognitive and athletic performance, as well as the monitoring of mental and physical health across diverse contexts. These processes operate within a distributed wireless communication framework, where data from IoT sensors and wearable devices are seamlessly transmitted to enable continuous and personalized assessment. Advanced data processing, supported by deep learning and machine learning techniques, drives the development of predictive models applied to these contexts. An innovative aspect of this research line is the introduction of recommendation systems, leveraging real-time data to suggest personalized interventions aimed at enhancing mental and physical well-being while simultaneously optimizing cognitive and athletic performance.

Bio: Laura Ginestretti is an MSc student in Biomedical Engineering at Politecnico di Milano. Her research focuses on cognitive performance, particularly exploring how physiological data from wearables can be utilized with deep learning and machine learning to predict and enhance mental and cognitive outcomes.

Speaker: Filippo Carloni

Abstract: Regular Expressions (REs) are a pervasive yet computationally intensive task critical for identifying data patterns in fields like personalized medicine and computer security. These applications demand massive data processing, where the high data dependency of REs often results in long execution times and elevated energy consumption. Currently, RE engines rely on either flexibility in run-time RE adaptability and broad operators to support impairing performance or fixed high-performing accelerators implementing few simple RE operators.

This poster presents how combining tailored hardware and software solutions to address these challenges by integrating RE-specific compiler infrastructures with advanced domain-specific architectures. These advancements deliver substantial improvements in execution speed, energy efficiency, and processing throughput, outperforming conventional CPU-based and ASIC-based approaches in low-latency and near-data processing scenarios.

Bio: Filippo Carloni is a PhD candidate in Information Technology (Computer Science and Engineering) at Politecnico di Milano. His PhD research focuses on domain-specific architectures and compilers, with a particular emphasis on the regular expressions domain. He works extensively with hardware description languages and FPGAs to address architectural challenges and implement advanced solutions. During the final year of his PhD, he was a visiting student at the COMMIT lab at MIT, where he began exploring SmartNIC hardware acceleration. His broader research interests include RISC-V architecture and ISA design.

Speaker: Beatrice Branchini, PhD student

Abstract: Variant Calling is one of the core procedures for highlighting genetic mutations and HaplotypeCaller from the Genome Analysis ToolKit is the de-facto standard for this procedure. However, this tool requires long runtimes for real-life datasets, where most is spent computing the Pair Hidden Markov Model (PairHMM)  algorithm. Given the compute-intensiveness of this algorithm, general-purpose CPUs struggle to deliver the required performance to analyze large-size genomic data. Thus, offloading this task to hardware accelerators like GPUs represents a suitable approach to speed up the execution. However, literature lacks in performance and flexibility. This talk illustrates a GPU-based implementation of the PairHMM, seamlessly integrated into HaplotypeCaller, that overcomes the performance and flexibility issues of the currently-available solutions.

Bio: Beatrice is a third-year PhD student in Information Technology at Politecnico di Milano, Milan, Italy. She worked in the HPC Team of the Scalable Computing and Data group at the Pacific Northwest National Laboratory (PNNL). Her research interests include High-Performance Computing (HPC) and heterogeneous architectures. In particular, her research focuses on developing novel methodologies to exploit HPC systems and hardware accelerators, such as GPUs and FPGAs, to overcome genomics analyses’ computational intensity, paving the way for a personalized and safer approach to medicine.

Speaker: Alessandro Verosimile

Abstract: 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”. Thanks to heterogeneous hardware, it is possible to bring Machine Learning 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 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. Lastly, the proposed co-processor has been optimized by re-implementing it through SATL, a SoA spatial template to implement accelerators for irregular workloads. With the proposed approach it becomes possible to accelerate the inference of RFs that count up to 3 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.

Bio: Alessandro Verosimile is a second-year PhD student in Information Technology at Politecnico di Milano. He worked for 6 months as a research intern in the RAD team of Advanced Micro Devices (AMD). His research focuses on HW-SW co-design techniques that aim to co-optimize the training of large Machine Learning models and the design of the hardware architecture for their inference on embedded devices, with a focus on both Deep learning models and Decision Tree based ensemble models.

Speaker: Susanna Bardini

Abstract: In a world where healthcare systems are increasingly strained by rising demands and limited resources, there is a pressing need for innovative IoT based solutions that can improve the quality and efficiency of care. Underlying this work is a desire to improve care for people with mobility impairments and the elderly, leveraging cutting-edge technology to create intelligent and adaptive systems that redefine efficiency and personalization. Central to this work is the development of non-invasive, distributed systems that integrate environmental and physiological monitoring to support real-time decision-making. These systems address critical issues such as caregiver shortages, inter-facility collaboration, and the need for privacy-preserving data management in personalized healthcare. A key challenge of this research is the integration of federated learning to enable robust and privacy-conscious data sharing between medical infrastructures while enhancing scalability and adaptability across diverse healthcare settings. In addition, an intelligent alert prioritization mechanism designed to ensure that critical events are addressed in a timely manner would optimize operator response time in resource-constrained environments. Furthermore, adaptive sensor management improves energy efficiency, fostering sustainable and long-term functionality in distributed IoT networks. This talk will explore the proposed architecture of this system that I would like to develop during my PhD, focusing on heterogeneous IoT networks, federated learning frameworks, and scalable alert management solutions.

Bio: Susanna Bardini is a first-year PhD student in Information Technology at Politecnico di Milano. Her research focuses on smart sensing technologies, IoT heterogeneous networks, and microcontroller-based systems for personalized health and smart homes. She is particularly interested in edge computing frameworks for real-time data processing and adaptive systems designed to enhance well-being through low-power solutions.

Speaker: Leonardo De Grandis

Abstract: Self-Supervised Learning has emerged as a promising paradigm to learn powerful representations. It allows to leverage large amounts of unlabeled data, often easy to obtain, and build a knowledge foundation that can be specialized for downstream tasks. However, while Computer Vision (CV) led the development with models like BYOL, DINO and Barlow Twins, Graph Machine Learning (GML) methods still often rely on contrastive approaches, which present known limitations such as batch size sensitivity and the need for negative samples. Therefore, in this work we explore the application of the DINO architecture, initially developed for the CV domain to overcome contrastive issues, to GML models. We evaluate the methodology for molecular property prediction tasks, pre-training on a subset of the ZINC20 dataset, and on different use cases provided in the TUDataset benchmarks, ranging from social to biological 8networks. Furthermore, we carefully analyze how representations are evolving through the model and how they behave in the latent space to ensure efficient information processing.

Bio: Leonardo De Grandis is a first-year PhD student in Information Technology at Politecnico di Milano. After gaining experience working with Graph Machine Learning methods applied to molecules and biological networks during the MSc, his research now focuses on applying such techniques to genomics. He aims to develop self-supervised approaches, which help leveraging large amounts of unlabeled data, combining multi-omic information to discover hidden patterns and advance our understanding of complex biological processes.

Speaker: Laura Ginestretti

Abstract: As the integration of wearable devices and IoT sensors continues to evolve, the ability to extract meaningful, data-driven insights on cognitive performance and mental well-being has increased exponentially. This talk presents a pilot study that leverages multimodal data—including nutrition, sleep, activity patterns, and real-time physiological monitoring—to track and predict cognitive performance.

Cognitive performance includes memory, reasoning, holding attention, thinking, reading, and learning capabilities. The study develops models able to predict cognitive performance using machine learning and deep learning, then applying explainable artificial intelligence (XAI) techniques, including SHAP and Occlusion methods, to provide actionable insights into feature importance and temporal dynamics.

This pilot study also leverages a recommender system that translates model predictions into personalized interventions, aimed at improving cognitive performance across diverse contexts. By detailing the data collection protocols, preprocessing methods, and experimental outcomes, this talk presents the first study in the state of the art correlating physiological data from wearable devices, physical activity and nutrition to cognitive performance.

Bio: Laura Ginestretti is an MSc student in Biomedical Engineering at Politecnico di Milano. Her research focuses on cognitive performance, particularly exploring how physiological data from wearables can be utilized with deep learning and machine learning to predict and enhance mental and cognitive outcomes.

Speaker: Riccardo Fiorentini

Abstract: The Variant Call Format (VCF) is a de facto standard file type in genomic analysis to report variations found during the alignment task. However, its complex and highly explicit structure poses significant computational challenges for parsing and querying tasks in this file format. Current high-performance tools, predominantly written in C++, deliver fast analyses but often lack flexibility and accessibility for Python users, creating barriers to their adoption. This work introduces a hybrid solution that combines the computational efficiency of C++ with Python’s usability while leveraging GPU acceleration through CUDA to achieve maximum performance on heterogeneous systems. The tool offers faster VCF parsing and efficient data organization, enabling seamless integration into data science and machine learning workflows. By optimizing memory consumption, multithreaded processing, and harnessing GPU power, this tool aims to redefine the standard for genomic data analysis, bridging the gap between performance and accessibility.

Bio: Riccardo Fiorentini is a second-year MSc student in Computer Science and Engineering. His research specializes in hardware acceleration for computational genomics, focusing on GPU-based optimization of computationally intensive algorithms such as PairHMM within HaplotypeCaller for the GATK toolkit. He was awarded second prize for this work in the EUROCON 2023 “IEEE Best Student Research Video & Poster Competition.” Currently, he is developing a GPU-accelerated VCF data parsing library, designed to enhance the efficiency of genomic workflows. Riccardo’s work bridges high-performance computing with modern genomics, focusing on heterogeneous architectures for scalable data analysis.

Speaker: Jacopo Lazzari

Abstract: Peer evaluation is a vital component of academic and educational settings, providing valuable opportunities for students to receive feedback, refine their work, and develop critical thinking skills. When combined with machine learning and statistical methods, peer evaluation can evolve into a more efficient and unbiased assessment system. However, traditional peer review processes are often compromised by inherent biases, such as those stemming from personal relationships, expectations, or subjective preferences of reviewers. These biases can lead to inaccurate feedback, especially in large-scale environments like Massive Open Online Courses (MOOC), where the volume of evaluations makes it difficult to ensure fairness and consistency. The proposed variational inference (VI) approach iteratively refines these parameters by adjusting for the biases and variances of each reviewer, leading to a more objective and accurate estimate of a student’s performance. This model is evaluated against several state-of-the-art techniques, demonstrating its ability to handle biases across diverse reviewer behaviors. Experiments, conducted using both simulated and real-world datasets, show that the VI-based method achieves better results compared to traditional ones, providing a scalable and reliable solution for peer evaluation systems.

Bio: Jacopo Lazzari is a second-year MSc student in Mathematical Engineering at Politecnico di Milano. His research focuses on applying Bayesian statistics and machine learning techniques to improve peer evaluation systems by mitigating biases. He aims to develop models that enhance the accuracy of assessments, with a particular focus on their application in large-scale educational environments and other domains. In addition to this, Jacopo has worked on applying Functional Data Analysis to neuroscience. His research explored innovative methods for extracting and interpreting signal features, improving classification accuracy in cognitive studies.