arXiv subjects

Silvia Santini

Publications and source records attributed to Silvia Santini.

8 recordsLinked to original sources

arXiv2026

Interpretability in Deep Time Series Models Demands Semantic Alignment

Deep time series models continue to improve predictive performance, yet their deployment remains limited by their black-box nature. In response, existing interpretability approaches in the field keep focusing on explaining the internal model computations, without addressing whether they align or not with how a human would reason about the studied phenomenon. Instead, we state interpretability in deep time series models should pursue semantic alignment: predictions should be expressed in terms of variables that are meaningful to the end user, mediated by spatial and temporal mechanisms that admit user-dependent constraints. In this paper, we formalize this requirement and state that, once established, semantic alignment must be preserved under temporal evolution: a constraint with no analog in static settings. Provided with this definition, we outline a blueprint for semantically aligned deep time series models, identify properties that support trust, and discuss implications for model design.

cs.LG↗

arXiv2026

NAPS: Attention-Based Fusion of Heterogeneous Physiological Signals

Physiological signals are inherently heterogeneous: they are collected under diverse acquisition setups, differ in the number and type of modalities and channels, varying in quality, reliability, and relevance across tasks. This variability poses a major challenge for machine learning models required to generalize across subjects, sensors, and clinical environments. Existing approaches typically train on limited modalities or single channels, leading to marginal representations that, on their own, fail to capture the systemic complexity of the physiological state; naive fusion of such representations, such as via pooling or voting schemes, is typically suboptimal, as it cannot adaptively weight different sources or capture temporal, spatial, and cross-modality dependencies. We introduce NAPS (Neural Aggregator of Physiological Signals), a neural module that performs principled data fusion to derive unified physiological representations, employing an ad hoc tri-axial attention mechanism and dimension-adaptive training to robustly manage varying high-dimensional sensor configurations. We test NAPS on automatic sleep staging from polysomnography (PSG), an ideal real-world application, where recordings consist of multiple physiological signals (EEG, EOG, EMG, ...), considerably varying in configuration across datasets and institutions. Leveraging frozen pretrained unimodal encoders, NAPS dynamically integrates representations or predictions, achieving state-of-the-art generalization across multiple datasets.

cs.LG↗

arXiv2026

A foundation model for electrodermal activity data

Foundation models have recently extended beyond natural language and vision to timeseries domains, including physiological signals. However, progress in electrodermal activity (EDA) modeling is hindered by the absence of large-scale, curated, and openly accessible datasets. EDA reflects sympathetic nervous system activity and is widely used to infer cognitive load, stress, and engagement. Yet very few wearable devices provide continuous, unobtrusive sensing, and the only large-scale archive to date is proprietary. To address this gap, we compile EDAMAME, a collection of EDA traces from 24 public datasets, comprising more than 25,000 hours from 634 users. Using this resource, we train UME, the first dedicated foundation model for EDA. In eight out of ten scenarios, UME outperforms baselines and matches generalist timeseries foundation models while using 20x fewer computational resources. Our findings, however, also highlight the intrinsic challenges of EDA modeling, motivating further research to unlock its full potential. All datasets, model weights, and code are released to support further research.

cs.LG↗

arXiv2026

Causally Reliable Concept Bottleneck Models

Concept-based models are an emerging paradigm in deep learning that constrains the inference process to operate through human-interpretable variables, facilitating explainability and human interaction. However, these architectures, on par with popular opaque neural models, fail to account for the true causal mechanisms underlying the target phenomena represented in the data. This hampers their ability to support causal reasoning tasks, limits out-of-distribution generalization, and hinders the implementation of fairness constraints. To overcome these issues, we propose Causally reliable Concept Bottleneck Models (C$^2$BMs), a class of concept-based architectures that enforce reasoning through a bottleneck of concepts structured according to a model of the real-world causal mechanisms. We also introduce a pipeline to automatically learn this structure from observational data and unstructured background knowledge (e.g., scientific literature). Experimental evidence suggests that C$^2$BMs are more interpretable, causally reliable, and improve responsiveness to interventions w.r.t. standard opaque and concept-based models, while maintaining their accuracy.

cs.LG↗

arXiv2025

Multi-Frequency Federated Learning for Human Activity Recognition Using Head-Worn Sensors

Human Activity Recognition (HAR) benefits various application domains, including health and elderly care. Traditional HAR involves constructing pipelines reliant on centralized user data, which can pose privacy concerns as they necessitate the uploading of user data to a centralized server. This work proposes multi-frequency Federated Learning (FL) to enable: (1) privacy-aware ML; (2) joint ML model learning across devices with varying sampling frequency. We focus on head-worn devices (e.g., earbuds and smart glasses), a relatively unexplored domain compared to traditional smartwatch- or smartphone-based HAR. Results have shown improvements on two datasets against frequency-specific approaches, indicating a promising future in the multi-frequency FL-HAR task. The proposed network's implementation is publicly available for further research and development.

cs.LG↗

arXiv2023

BiHeartS: Bilateral Heart Rate from multiple devices and body positions for Sleep measurement Dataset

Sleep is the primary mean of recovery from accumulated fatigue and thus plays a crucial role in fostering people's mental and physical well-being. Sleep quality monitoring systems are often implemented using wearables that leverage their sensing capabilities to provide sleep behaviour insights and recommendations to users. Building models to estimate sleep quality from sensor data is a challenging task, due to the variability of both physiological data, perception of sleep quality, and the daily routine across users. This challenge gauges the need for a comprehensive dataset that includes information about the daily behaviour of users, physiological signals as well as the perceived sleep quality. In this paper, we try to narrow this gap by proposing Bilateral Heart rate from multiple devices and body positions for Sleep measurement (BiHeartS) dataset. The dataset is collected in the wild from 10 participants for 30 consecutive nights. Both research-grade and commercial wearable devices are included in the data collection campaign. Also, comprehensive self-reports are collected about the sleep quality and the daily routine.

cs.HC↗

arXiv2021

Whisper: Fast Flooding for Low-Power Wireless Networks

This paper presents Whisper, a fast and reliable protocol to flood small amounts of data into a multi-hop network. Whisper makes use of synchronous transmissions, a technique first introduced by the Glossy flooding protocol. In contrast to Glossy, Whisper does not let the radio switch from receive to transmit mode between messages. Instead, it makes nodes continuously transmit identical copies of the message and eliminates the gaps between subsequent transmissions. To this end, Whisper embeds the message to be flooded into a signaling packet that is composed of multiple packlets -- where a packlet is a portion of the message payload that mimics the structure of an actual packet. A node must intercept only one of the packlets to detect that there is an ongoing transmission and that it should start forwarding the message. This allows Whisper to speed up the propagation of the flood, and thus, to reduce the overall radio-on time of the nodes. Our evaluation on the FlockLab testbed shows that Whisper achieves comparable reliability but 2x lower radio-on time than Glossy. We further show that by embedding Whisper in an existing data collection application, we can more than double the lifetime of the network.

cs.NI↗

arXiv2020

Synchronous Transmissions in Low-Power Wireless: A Survey of Communication Protocols and Network Services

Low-power wireless communication is a central building block of Cyber-physical Systems and the Internet of Things. Conventional low-power wireless protocols make avoiding packet collisions a cornerstone design choice. The concept of synchronous transmissions challenges this view. As collisions are not necessarily destructive, under specific circumstances, commodity low-power wireless radios are often able to receive useful information even in the presence of superimposed signals from different transmitters. We survey the growing number of protocols that exploit synchronous transmissions for higher robustness and efficiency as well as unprecedented functionality and versatility compared to conventional designs. The illustration of protocols based on synchronous transmissions is cast in a conceptional framework we establish, with the goal of highlighting differences and similarities among the proposed solutions. We conclude the paper with a discussion on open research questions in this field.

cs.NI↗