Daniele Berardini

Alessandro Cacciatore, Valerio Morelli, Federica Paganica et al.

Deep learning has long been dominated by multi-layer perceptrons (MLPs), which have demonstrated superiority over other optimizable models in various domains. Recently, a new alternative to MLPs has emerged - Kolmogorov-Arnold Networks (KAN)- which are based on a fundamentally different mathematical framework. According to their authors, KANs address several major issues in MLPs, such as catastrophic forgetting in continual learning scenarios. However, this claim has only been supported by results from a regression task on a toy 1D dataset. In this paper, we extend the investigation by evaluating the performance of KANs in continual learning tasks within computer vision, specifically using the MNIST datasets. To this end, we conduct a structured analysis of the behavior of MLPs and two KAN-based models in a class-incremental learning scenario, ensuring that the architectures involved have the same number of trainable parameters. Our results demonstrate that an efficient version of KAN outperforms both traditional MLPs and the original KAN implementation. We further analyze the influence of hyperparameters in MLPs and KANs, as well as the impact of certain trainable parameters in KANs, such as bias and scale weights. Additionally, we provide a preliminary investigation of recent KAN-based convolutional networks and compare their performance with that of traditional convolutional neural networks. Our codes can be found at https://github.com/MrPio/KAN-Continual_Learning_tests.

Lucia Migliorelli, Daniele Berardini, Kevin Cela et al.

Background and objectives: Patients suffering from neurological diseases may develop dysarthria, a motor speech disorder affecting the execution of speech. Close and quantitative monitoring of dysarthria evolution is crucial for enabling clinicians to promptly implement patient management strategies and maximizing effectiveness and efficiency of communication functions in term of restoring, compensating or adjusting. In the clinical assessment of orofacial structures and functions, at rest condition or during speech and non-speech movements, a qualitative evaluation is usually performed, throughout visual observation. Methods: To overcome limitations posed by qualitative assessments, this work presents a store-and-forward self-service telemonitoring system that integrates, within its cloud architecture, a convolutional neural network (CNN) for analyzing video recordings acquired by individuals with dysarthria. This architecture, called facial landmark Mask RCNN, aims at locating facial landmarks as a prior for assessing the orofacial functions related to speech and examining dysarthria evolution in neurological diseases. Results: When tested on the Toronto NeuroFace dataset, a publicly available annotated dataset of video recordings from patients with amyotrophic lateral sclerosis (ALS) and stroke, the proposed CNN achieved a normalized mean error equal to 1.79 on localizing the facial landmarks. We also tested our system in a real-life scenario on 11 bulbar-onset ALS subjects, obtaining promising outcomes in terms of facial landmark position estimation. Discussion and conclusions: This preliminary study represents a relevant step towards the use of remote tools to support clinicians in monitoring the evolution of dysarthria.

Carolina Gonçalves, João M. Lopes, Sara Moccia et al.

Gait disabilities are among the most frequent worldwide. Their treatment relies on rehabilitation therapies, in which smart walkers are being introduced to empower the user's recovery and autonomy, while reducing the clinicians effort. For that, these should be able to decode human motion and needs, as early as possible. Current walkers decode motion intention using information of wearable or embedded sensors, namely inertial units, force and hall sensors, and lasers, whose main limitations imply an expensive solution or hinder the perception of human movement. Smart walkers commonly lack a seamless human-robot interaction, which intuitively understands human motions. A contactless approach is proposed in this work, addressing human motion decoding as an early action recognition/detection problematic, using RGB-D cameras. We studied different deep learning-based algorithms, organised in three different approaches, to process lower body RGB-D video sequences, recorded from an embedded camera of a smart walker, and classify them into 4 classes (stop, walk, turn right/left). A custom dataset involving 15 healthy participants walking with the device was acquired and prepared, resulting in 28800 balanced RGB-D frames, to train and evaluate the deep networks. The best results were attained by a convolutional neural network with a channel attention mechanism, reaching accuracy values of 99.61% and above 93%, for offline early detection/recognition and trial simulations, respectively. Following the hypothesis that human lower body features encode prominent information, fostering a more robust prediction towards real-time applications, the algorithm focus was also evaluated using Dice metric, leading to values slightly higher than 30%. Promising results were attained for early action detection as a human motion decoding strategy, with enhancements in the focus of the proposed architectures.

Xuan Qi, Daniele Berardini, Dario Serez et al.

Unsupervised domain adaptation (UDA) aims to learn a target-domain classifier from labeled source data and unlabeled target data under distribution shift. Recent diffusion-based UDA methods approach this problem by synthesizing labeled target-style images and training on the resulting synthetic data. However, their performance depends heavily on the conditioning design: class prompts provide only coarse guidance, while domain adaptation modules mainly control appearance, which may leave target-style synthesis insufficiently specified. We propose VT-DUDA, a visual-token conditioning framework for diffusion-guided UDA. Instead of relying only on text prompts, VT-DUDA uses source images to provide additional instance-level visual context for target-style synthesis. Specifically, VT-DUDA maps each source image to a compact sequence of visual tokens and forms a hybrid conditioning context by concatenating these tokens with the corresponding text embeddings along the cross-attention context dimension of a latent diffusion model. This provides instance-dependent conditioning beyond text alone, while synthesis is performed with the target-domain adapter branch. Because guidance is represented explicitly as a token sequence, the same interface also permits inference-time manipulation of the conditioning signal through token selection and token-strength adjustment. The proposed method preserves the standard diffusion objective and can be integrated into existing adapter-based diffusion frameworks without modifying the backbone. Across Office-31, Office-Home, and VisDA-2017, VT-DUDA improves average target-domain accuracy over strong discriminative and diffusion-based UDA baselines. The results suggest that, in generation-based UDA, a stronger conditioning interface can improve the downstream usefulness of synthetic target-style data.

Lorenzo Stacchio, Michele Giordano, Daniele Berardini et al.

The advent of Generative Artificial Intelligence (GenAI), and in particular Large Language Models (LLMs), is reshaping educational practice, while intensifying ethical debate about its adoption. To date, the dominant paradigm remains cloud-based and text-only chatbot: a centralized service that offers limited pedagogical control, weak transparency over knowledge sources, and non-trivial risks for privacy and regulatory compliance. This model also presumes continuous connectivity and recurring API costs, creating structural barriers for many institutions, reinforcing existing digital divides. At the same time, educational interaction with LLM can benefit from multimodal cues and embodied presence, requiring interfaces that move beyond text-only tutoring. In this work, we propose ELEVATE (Efficient LLM Education with Virtual Avatar Teaching Engine), a framework to develop efficient GenAI-driven avatar tutors governed by epistemic infrastructures. ELEVATE integrates LLM-driven dialogue with embodied 3D avatars for multimodal interaction and adopts a local-first execution model enabling deployment on consumer-grade hardware. The framework formalizes a three-stratum design that separates (i) a student-facing virtual avatar interaction layer, (ii) a local GenAI execution and multimodal synthesis core, and (iii) a teacher-facing governance layer. We implemented and evaluated a working prototype deployed in a real-world educational curriculum. The system runs on standard PCs and smartphones, and we provide system-level performance evidence to show responsive interaction under realistic hardware constraints. Finally, we discuss sociotechnical and pedagogical implications for responsible adoption, positioning ELEVATE as a scalable pathway for privacy-preserving and inclusive GenAI tutoring across heterogeneous school environments.

Daniele Berardini, Vito Paolo Pastore, Vittorio Murino

One-Shot Federated Learning (OSFL) addresses extreme communication regimes in which clients interact with the server only once, amplifying the impact of heterogeneous client data distributions. In particular, the interaction of domain shift and label shift across clients induces misaligned feature representations that cannot be corrected through iterative optimization. Existing OSFL methods rely on distillation, server-side generation or ensemble-based aggregation, but assume aligned representations or address domain and label shift separately. We introduce SLOT-Align (Single-round, Learning-free Optimal Transport Alignment), a geometry-aware feature harmonization framework for OSFL. SLOT-Align uses a shared frozen encoder to extract compact feature statistics, constructs a global reference via Bures-Wasserstein barycenters, and aligns local representations using closed-form geodesic optimal transport maps. The method is computationally efficient and can be combined with existing OSFL pipelines relying on frozen encoders without modifying their training procedures. Extensive experiments across multiple benchmarks, pretrained backbones, and OSFL methods show that SLOT-Align consistently improves accuracy and robustness under joint domain and label shift.