Available Thesis (Bachelor and MSc) UNITO 🇮🇹🏴󠁧󠁢󠁥󠁮󠁧󠁿

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: time series, classification

Abstract: Time series classification is a pivotal task in modern machine learning, with widespread applications in fields such as healthcare, finance, and cybersecurity. A lot of time series classification (TSC) methods have been proposed in the last years. This thesis aims to benchmark the most widespread TSC algorithms on toy and real-world datasets.

Task: The objective is to benchmark the most famous TSC methods on toy and real-world datasets, and highlight their properties. State-of-the-art techniques in TSC include HIVE-COTE (v.1 and v.2), BOSS, Rocket (and its Mini/MultiRocket variants), Deep Learning models (InceptionTime, ResNet1D, UTime, LSTMs), and Shapelet-based methods. These methods (at least 5 or 6) should be tested on univariate and multivariate datasets, spanning from benchmark (i.e., the UCR and the UEA archives) and on real-world datasets. The thesis will consist of a first phase of literature review to fully understand how these methods work, followed by a phase of coding and experimentation.

Bibliography:
– Dempster A., “ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels”, Data Mining and Knowledge Discovery, 2020, https://link.springer.com/article/10.1007/s10618-020-00701-z
– Middlehurst M., “HIVE-COTE 2.0: a new meta ensemble for time series classification”, Machine Learning, 2021, https://link.springer.com/article/10.1007/s10994-021-06057-9
-Schäfer P., “The BOSS is concerned with time series classification in the presence of noise”, Data Mining and Knowledge Discovery, 29(6): 2015 https://link.springer.com/article/10.1007/s10618-014-0377-7

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 30%
Code implementation: 50%
Experiments: 20%

Requirements: The candidate should have good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: computer vision, classification, continual learning

Abstract: Mammoth is a framework for Continual Learning research. With more than 70 methods and 20 datasets, it includes the most complete list of competitors and benchmarks for research purposes. This thesis aims to experiment and benchmark a set (5-10) of continual learning techniques for image classification tasks.

Task: The objective is to benchmark the most famous continual learning methods for image classification tasks, and highlight their properties. According to the student’s preference, the task could be changed to detection or segmentation rather than classification. State-of-the-art techniques in continual learning include rehearsal methods such as Experience Replay, Dark Experience Replay, Learning without Forgetting, and so on… The Mammoth framework for continual learning should be used. These methods (at least 5 or 6) should be tested on benchmark and/or real-world datasets. This will depend on the student’s passions. The thesis will consist of a first phase of literature review to fully understand how these methods work, followed by a phase of coding and experimentation.

Bibliography:
– Rolnick D., “Experience Replay for Continual learning”, Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 2019, https://papers.nips.cc/paper_files/paper/2019/hash/fa7cdfad1a5aaf8370ebeda47a1ff1c3-Abstract.html
– Buzzega P., “Dark experience for general continual learning: a strong, simple baseline”, Advances in Neural Information Processing Systems 32 (NeurIPS 2020), 2020, https://dl.acm.org/doi/10.5555/3495724.3497059
– Li Z., “Learning without Forgetting”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, https://ieeexplore.ieee.org/ielaam/34/8520726/8107520-aam.pdf?tag=1

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 30%
Code implementation: 50%
Experiments: 20%

Requirements: The candidate should have good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: federated learning, time series classification

Abstract: Time series classification (TSC) is a pivotal task in modern machine learning, with widespread applications in fields such as healthcare, finance, and cybersecurity. However, TSC techniques are subject to many challenges, such as resource constraints and data privacy. Data privacy is a critical concern, particularly in key applications such as healthcare and finance. In these contexts, the use of sensitive or personal data is often governed by strict regulations such as the General Data Protection Regulation (GDPR). To address this challenge, Federated Learning (FL) presents a natural and effective solution for collaboratively training models across decentralized data sources while preserving user privacy. Most existing FL algorithms primarily rely on Neural Network (NN) models. However, recent advancements began to explore the use of traditional machine learning (ML) methods in FL settings. This thesis aims to adapt/provide a federated version of an already existing TSC algorithm.

Task: The student is first supposed to familiarize with FL and to have a careful review of TSC literature. Then, the student should identify one or more TSC techniques of interest that could be federated. A set of experiments on univariate and/or multivariate datasets (like the UCR and the UEA archives) should be conducted. Finally, depending on the achieved results, it could be possible to work towards a conference/journal publication. Some FL and TSC papers are included in the bibliography section, as well as some recent papers proposing a federated version of already existing TSC methods.

Bibliography:
– McMahan B., et al., “Communication-Efficient Learning of Deep Networks from Decentralized Data”, https://proceedings.mlr.press/v54/mcmahan17a/mcmahan17a.pdf, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2017
– Dempster A., “ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels”, Data Mining and Knowledge Discovery, 2020, https://link.springer.com/article/10.1007/s10618-020-00701-z
– Middlehurst M., “HIVE-COTE 2.0: a new meta ensemble for time series classification”, Machine Learning, 2021, https://link.springer.com/article/10.1007/s10994-021-06057-9
-Schäfer P., “The BOSS is concerned with time series classification in the presence of noise”, Data Mining and Knowledge Discovery, 29(6): 2015 https://link.springer.com/article/10.1007/s10618-014-0377-7
-Casella B., “Federated Time Series Classification with Rocket features”, Proceedings of the 32nd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN, 2024, https://dx.doi.org/10.14428/esann/2024.es2024-214
-Casella B., “Fed2RC: Federated Rocket Kernels and Ridge Classifier for Time Series Classification”, 28th European Conference on Artificial Intelligence, ECAI 2025.

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 20%
Code implementation: 50%
Experiments: 15%
Working towards a publication: 15%

Requirements: The candidate should have good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: deep learning, image classification

Abstract: Convolutional neural networks are very useful in computer vision domains, since they transformed image analytics. They are the most widely used building blocks for solving problems involving images. Many architectures have achieved exceptional accuracies in image classification tasks. However, they come with some flaws, and one of them is the time they take to train on huge datasets. Separable convolutions are helpful to tackle this problem. Separable convolutions simply divides a kernel into two, smaller kernels. Since they reduce the number of operations that should be made, separable convolutions require less computation power and time. The idea of this thesis is to mix the separable convolutions with random kernels (ROCKET) to extract random features from images and to train a linear classifier. Ideally, the final model should reach competitive learning performance with just a fraction of time required by big neural networks.

Task: There are two main types of separable convolutions: spatial separable convolutions, and depthwise separable convolutions. The spatial separable convolution is so named because it deals primarily with the spatial dimensions of an image and kernel: the width and the height. A spatial separable convolution simply divides a kernel into two, smaller kernels. The most common case would be to divide a 3×3 kernel into a 3×1 and 1×3 kernel. Unfortunately, spatial separable convolutions have some significant limitations, meaning that it is not heavily used in deep learning. The main issue with the spatial separable convolution is that not all kernels can be “separated” into two, smaller kernels. Unlike spatial separable convolutions, depthwise separable convolutions work with kernels that cannot be “factored” into two smaller kernels. Hence, it is more commonly used. The depthwise separable convolution is so named because it deals not just with the spatial dimensions, but with the depth dimension — the number of channels — as well. Similar to the spatial separable convolution, a depthwise separable convolution splits a kernel into 2 separate kernels that do two convolutions: the depthwise convolution and the pointwise convolution. With less computations, the network is able to process more in a shorter amount of time. A famous algorithm for time series classification, ROCKET, exploits random kernels to extract features from time series data, and to feed a linear classifier. ROCKET achieves state-of-the-art accuracies with a just a fraction of time. The secret is that since convolutional kernels are random, we do not spend time on fine-tuning their parameters. Additionally, we ROCKET trains a simple linear classifier, that has a few parameters compared to huge neural networks, thus requiring less computation. The idea of this thesis is to extend the ROCKET framework to deal with images, by adopting random separable convolutions to extract features.

Bibliography:
– https://medium.com/data-science/a-basic-introduction-to-separable-convolutions-b99ec3102728
– Dempster A., “ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels”, Data Mining and Knowledge Discovery, 2020, https://link.springer.com/article/10.1007/s10618-020-00701-z

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 20%
Code implementation: 50%
Experiments: 15%
Working towards a publication: 15%

Requirements: The candidate should have good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: federated learning, decentralized learning, continual learning, experience replay, medical imaging

Abstract: Federated learning encompasses a family of strategies for distributed training over multiple nodes, each with its own private dataset, which typically communicates with a central node by sending local model updates used to train the main model. In this scenario, no data is explicitly shared between nodes, thus addressing the required privacy issues.
However, the presence of a central node that aggregates local updates simplifies the communication protocol when the number of clients is very large (thousands or millions) but introduces several downsides: it represents a single point of failure; it can become a bottleneck when the number of clients increases; in general, it may not always be available or desirable in collaborative learning scenarios.
In decentralized federated learning, the central node is replaced by peer-to-peer communication between clients: there is no longer a global shared model as in standard FL, but the communication protocol is designed so that all local models approximately converge to the same solution. Decentralized learning is particularly suitable for application in the medical domain, where the number of nodes (i.e., institutions) is relatively low; however, research is still ongoing, and no effective solutions have been established.

Task: we aim to investigate some unexplored properties of the FedER algorithm, a decentralized federated learning strategy based on continual learning principles designed for medical imaging data, which outperforms server-based federated learning approaches and yields performance similar to a standard (non-federated) training settings. In particular, the goal of this thesis is to explore the possibility of model heterogeneity in FedER. Unlike all other existing methods based on parameter averaging, the FedER approach does not strictly require that all nodes share the same model architecture. Model heterogeneity could, therefore, be employed to create a shared ensemble and combine different feature learning capabilities.

Bibliography:
– McMahan B., et al., “Communication-Efficient Learning of Deep Networks from Decentralized Data”, https://proceedings.mlr.press/v54/mcmahan17a/mcmahan17a.pdf, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2017
– Pennisi M., et al. “FedER: Federated Learning through Experience Replay and privacy-preserving data synthesis”, Computer Vision and Image Understanding (CVIU), 2024, https://www.sciencedirect.com/science/article/pii/S107731422300262X

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 20%
Code implementation: 50%
Experiments: 15%
Working towards a publication: 15%

Requirements: The candidate should have good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: federated learning, video classification, resource constraints

Abstract: Video classification is a computationally intensive task due to the large amount of data and complex processing involved. Each video consists of multiple frames, often in high resolution, and each frame must be analyzed both individually and in the context of temporal dynamics. This requires significant memory and processing power, especially when working with deep learning models like 3D convolutional networks or transformer-based architectures that process both spatial and temporal information. The need to handle long sequences of frames and the variety of actions or events within videos further increase computational demands. Developing methods that require fewer resources, such as lightweight models or efficient pretraining techniques, could make video classification more accessible and reduce the strain on computational infrastructure while still maintaining strong performance. ROCKET (RandOm Convolutional KErnel Transform) is a framework designed for time series classification that stands out for its efficiency and low computational requirements. Unlike traditional deep learning models, ROCKET uses random convolutional kernels to extract features from time series data, followed by a simple linear classifier, such as logistic regression, to perform classification. This approach is both fast and scalable, as it eliminates the need for complex training procedures while still achieving state-of-the-art performance on many tasks. Given its lightweight nature and ability to handle large datasets efficiently, ROCKET could be an ideal solution for video classification tasks, particularly when computational resources are limited.

Task: By treating each video frame or sequence of frames as a time series, ROCKET’s efficient feature extraction can be adapted to video data, potentially reducing the high resource demands typically associated with deep learning-based video classification models. This makes it a promising method for scenarios where minimizing computational cost is crucial.
Two possible solutions:
– adapting ROCKET kernels from time series to videos: this means modifying ROCKET kernels, originally developed to deal with 1D arrays, in such a way they can manage 3D structures like videos.
– converting 3D signals to 1D arrays: this can be done by extracting some info (mean, median, activations) from each frame of the video, and merging them to create a time series.
A federated learning approach can also be applied.

Bibliography:
– Dempster A., “ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels”, Data Mining and Knowledge Discovery, 2020, https://link.springer.com/article/10.1007/s10618-020-00701-z

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 20%
Code implementation: 50%
Experiments: 15%
Working towards a publication: 15%

Requirements: The candidate should have good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referents: Marco Aldinucci, Bruno Casella (bruno.casella@unito.it)

Keywords: federated learning, decentralized learning

Abstract: Federated learning encompasses a family of strategies for distributed training over multiple nodes, each with its own private dataset, which typically communicates with a central node by sending local model updates used to train the main model. In this scenario, no data is explicitly shared between nodes, thus addressing the required privacy issues.
However, the presence of a central node that aggregates local updates simplifies the communication protocol when the number of clients is very large (thousands or millions) but introduces several downsides: it represents a single point of failure; it can become a bottleneck when the number of clients increases; in general, it may not always be available or desirable in collaborative learning scenarios.
In decentralized federated learning, the central node is replaced by peer-to-peer communication between clients: there is no longer a global shared model as in standard FL, but the communication protocol is designed so that all local models approximately converge to the same solution. Decentralized learning is particularly suitable for application in the medical domain, where the number of nodes (i.e., institutions) is relatively low; however, research is still ongoing, and no effective solutions have been established.

Task: we aim to investigate some unexplored properties of the FedER algorithm, a decentralized federated learning strategy based on continual learning principles designed for medical imaging data, which outperforms server-based federated learning approaches and yields performance similar to a standard (non-federated) training settings. FedER, by leveraging continual learning and generative adversarial concepts proposes
a principled way for training local models that approximately converge to the same decisions, without the need of a shared model architecture and of central coordination. In particular, the convergence is evidenced by empirical results. The goal of this thesis is to provide a theoretical proof of this property.

Bibliography:
– McMahan B., et al., “Communication-Efficient Learning of Deep Networks from Decentralized Data”, https://proceedings.mlr.press/v54/mcmahan17a/mcmahan17a.pdf, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2017
– Pennisi M., et al., “FedER: Federated Learning through Experience Replay and privacy-preserving data synthesis”, Computer Vision and Image Understanding (CVIU), 2024, https://www.sciencedirect.com/science/article/pii/S107731422300262X
– Li X., et al., “On the Convergence of FedAvg on Non-IID Data”, International Conference on Learning Representations (ICLR), 2020, https://iclr.cc/virtual_2020/poster_HJxNAnVtDS.html

Workload: Under the guidance of a team of researchers, the student will first be required to study recent works on the subject before being practically introduced to the task by experimenting with existing code.
Literature: 20%
Code implementation: 50%
Experiments: 15%
Working towards a publication: 15%

Requirements: The candidate should have high math and good programming skills (Python and PyTorch) and be motivated. A previous knowledge of the topic would be preferred but not required.

Referenti: Marco Aldinucci, Gianluca Mittone (gianluca.mittone@unito.it)

Descrizione: Con il termine Federated Learning (FL) si intende una tecnica di Machine Learning (ML) in cui il processo di apprendimento è effettuato a partire da dati partizionati su più device.

La peculiarità di questo metodo è quella di produrre un unico modello di ML efficace su tali dati senza il bisogno di dover raggruppare quest’ultimi in un solo luogo, con tutta una serie di vantaggi sulla privacy degli utenti e sui costi della gestione dei dati stessi.

Questo approccio, fin’ora utilizzato solo su Deep Neural Networks (DNN), sta venendo espanso dalla comunità scientifica anche a modelli non-deep, quali Alberi di Decisione, modelli Naive Bayes, Regressione Logistica e così via.

In questa tesi verrà esplorato l’algoritmo per non-deep FL AdaBoost.F, sviluppato nel
nostro dipartimento, nella sua implementazione nel framework per FL di Intel OpenFL al fine di indagarne le prestazioni sia di apprendimento sia computazionali.

Referenti: Marco Aldinucci, Alberto Riccardo Martinelli (albertoriccardo.martinelli@unito.it)

Descrizione: Le GPU sono ormai utilizzate in molti campi per risolvere problemi scientifici complessi in modo efficiente.

Scrivere codice portabile ed efficiente è un task notoriamente difficile e nel caso delle gpu questo problema è ancora più marcato. Negli ultimi anni sono stati molti gli sforzi per creare dei tool per facilitare la scrittura di codice parallelo per GPU.

OpenAcc e OpenMP sono I tool maggiormente usati per questo compito.

Il candidato dovrà effettuare uno studio approfondito di questi tool (confrontandoli anche con approcci di più basso livello) per capirne I punti di forza e di debolezza.
Inoltre dovrà scrivere degli esempi con entrambi I tool per avere un confronto completo sia sull’usabilità che sulle prestazioni.
Al candidato verrà fornito Spray-web, un codice scientifico che simula la dispersione di inquinanti in un ambiente 3D. Esistono diverse versioni di questo codice: OpenMP, OpenACC, CUDA e FPGA. Il candidato potrà testare le varie versioni per avere esperienza con un codice reale che gli permetta di avere una comprensione ancora più completa su tali tool.