In this page you can find an overview of some thesis projects currently available in our research laboratory. For any information or to discuss additional projects you can contact me anytime.
- Automatic prediction of black carbon data
Automatic prediction of black carbon data with neural forecasting techniques. Joint work with the Institute of Methodologies for Environmental Analysis.
- Deep learning applied to cytopathology
Investigating the use of deep learning techniques in the context of cytopathological data (diagnosis at the cellular level).
- Classification of Latin handwritten manuscripts
In the context of the data from the In Codice Ratio project (promoted and coordinated by Roma Tre University), thesis works can explore the application of deep neural networks for recognizing and classifying Latin handwritten characters / documents.
- Adversarial attacks in neural networks
Adversarial attacks are patterns designed to “fool” or attack a given network. We plan on exploring their impact on different neural network architectures, including those with trainable activation functions and graph-based networks.
- Quaternion-valued activation functions
Hyper-complex neural networks extend deep networks to domains where data is more easily represented as complex or quaternion numbers (e.g., an RGBA image pixel can be modeled in this form). The thesis will explore the design of a suitable activation function for these networks.
- Dynamic selection of activation functions
In this thesis, we would like to explore the dynamical choice of an activation function in a neural network, done by the neural network itself through a process of selective attention.
- Group sparse lottery ticket hypothesis
The lottery ticket hypothesis is a novel hypothesis on the architectural properties of the neural network, where sparsity plays a key role. In this thesis we are interested in exploring group-level and neuron-level sparsity strategies when evaluating the hypothesis.
- Self-supervised learning on network’ embeddings
Self-supervised learning tries to facilitate supervised learning by creating multiple auxiliary objectives for training. In this thesis we would like to investigate novel ways for creating these auxiliary tasks.
- Branching networks for Iot applications
Analysis of differentiable techniques for networks with multiple branches and exits, especially suitable for IoT environments and on-the-edge applications.
- Techniques for graph networks explainibility
Investigating explainability in the context of graph neural networks, for the identification of important features and causal structures in dynamic scenarios.
- Konica Minolta Laboratory Europe
Multiple proposals on human activity recognition, person tracking, online robotic adaptation, AI on the edge.