-
Theoretical guarantees in KL for Diffusion Flow Matching
Authors:
Marta Gentiloni Silveri,
Giovanni Conforti,
Alain Durmus
Abstract:
Flow Matching (FM) (also referred to as stochastic interpolants or rectified flows) stands out as a class of generative models that aims to bridge in finite time the target distribution $ν^\star$ with an auxiliary distribution $μ$, leveraging a fixed coupling $π$ and a bridge which can either be deterministic or stochastic. These two ingredients define a path measure which can then be approximated…
▽ More
Flow Matching (FM) (also referred to as stochastic interpolants or rectified flows) stands out as a class of generative models that aims to bridge in finite time the target distribution $ν^\star$ with an auxiliary distribution $μ$, leveraging a fixed coupling $π$ and a bridge which can either be deterministic or stochastic. These two ingredients define a path measure which can then be approximated by learning the drift of its Markovian projection. The main contribution of this paper is to provide relatively mild assumptions on $ν^\star$, $μ$ and $π$ to obtain non-asymptotics guarantees for Diffusion Flow Matching (DFM) models using as bridge the conditional distribution associated with the Brownian motion. More precisely, we establish bounds on the Kullback-Leibler divergence between the target distribution and the one generated by such DFM models under moment conditions on the score of $ν^\star$, $μ$ and $π$, and a standard $L^2$-drift-approximation error assumption.
△ Less
Submitted 12 September, 2024;
originally announced September 2024.
-
KL Convergence Guarantees for Score diffusion models under minimal data assumptions
Authors:
Giovanni Conforti,
Alain Durmus,
Marta Gentiloni Silveri
Abstract:
Diffusion models are a new class of generative models that revolve around the estimation of the score function associated with a stochastic differential equation. Subsequent to its acquisition, the approximated score function is then harnessed to simulate the corresponding time-reversal process, ultimately enabling the generation of approximate data samples. Despite their evident practical signifi…
▽ More
Diffusion models are a new class of generative models that revolve around the estimation of the score function associated with a stochastic differential equation. Subsequent to its acquisition, the approximated score function is then harnessed to simulate the corresponding time-reversal process, ultimately enabling the generation of approximate data samples. Despite their evident practical significance these models carry, a notable challenge persists in the form of a lack of comprehensive quantitative results, especially in scenarios involving non-regular scores and estimators. In almost all reported bounds in Kullback Leibler (KL) divergence, it is assumed that either the score function or its approximation is Lipschitz uniformly in time. However, this condition is very restrictive in practice or appears to be difficult to establish. To circumvent this issue, previous works mainly focused on establishing convergence bounds in KL for an early stopped version of the diffusion model and a smoothed version of the data distribution, or assuming that the data distribution is supported on a compact manifold. These explorations have led to interesting bounds in either Wasserstein or Fortet-Mourier metrics. However, the question remains about the relevance of such early-stopping procedure or compactness conditions. In particular, if there exist a natural and mild condition ensuring explicit and sharp convergence bounds in KL. In this article, we tackle the aforementioned limitations by focusing on score diffusion models with fixed step size stemming from the Ornstein-Uhlenbeck semigroup and its kinetic counterpart. Our study provides a rigorous analysis, yielding simple, improved and sharp convergence bounds in KL applicable to any data distribution with finite Fisher information with respect to the standard Gaussian distribution.
△ Less
Submitted 12 September, 2024; v1 submitted 23 August, 2023;
originally announced August 2023.
-
Influence of some cardiovascular risk factors on the rela-tionship between age and blood pressure
Authors:
Giulia Silveri,
Lorenzo Pascazio,
Milos Ajcevic,
Aleksandar Miladinovic,
Agostino Accardo1
Abstract:
Blood Pressure (BP) is a biological signal related to the cardiovascular system that inevitably is affected by ageing. Moreover, it is also influenced by the presence of cardiovascular risk factors. To evaluate how the relationship be-tween BP and age changes with the presence of risk factors in hypertensive and normotensive subjects, we analyzed 880 subjects with and without smoking, obe-sity, di…
▽ More
Blood Pressure (BP) is a biological signal related to the cardiovascular system that inevitably is affected by ageing. Moreover, it is also influenced by the presence of cardiovascular risk factors. To evaluate how the relationship be-tween BP and age changes with the presence of risk factors in hypertensive and normotensive subjects, we analyzed 880 subjects with and without smoking, obe-sity, diabetes mellitus and dyslipidemia. A regression line fitted each BP/Age relation calculated separately for normotensive and hypertensive subjects with and without risk factors. For each of the four conditions the office and the 24-hour ambulatory BP monitoring (ABPM) were considered. In subjects with and without risk factors, the slopes of the Systolic BP/Age relation were higher in hypertensive than in normotensive subjects in both office and ABPM conditions. Moreover, the presence of risk factors modified the Systolic BP/Age relation in hypertensive subjects by using either office or ABPM measurements. Finally, we confirmed that the difference between the two modalities depends on age too.
△ Less
Submitted 23 November, 2020;
originally announced November 2020.
-
Influence of the gender on the relationship between heart rate and blood pressure
Authors:
Giulia Silveri,
Lorenzo Pascazio,
Milos Ajcevic,
Aleksandar Miladinovic,
Agostino Accardo
Abstract:
Blood Pressure (BP) and Heart Rate (HR) provide information on clin-ical condition along 24h. Both signals present circadian changes due to sympa-thetic/parasympathetic control system that influence the relationship between them. Moreover, also the gender could modify this relation, acting on both con-trol systems. Some studies, using office measurements examined the BP/HR re-lation, highlighting…
▽ More
Blood Pressure (BP) and Heart Rate (HR) provide information on clin-ical condition along 24h. Both signals present circadian changes due to sympa-thetic/parasympathetic control system that influence the relationship between them. Moreover, also the gender could modify this relation, acting on both con-trol systems. Some studies, using office measurements examined the BP/HR re-lation, highlighting a direct association between the two variables, linked to sus-pected coronary heart disease. Nevertheless, till now such relation has not been studied yet using ambulatory technique that is known to lead to additional prog-nostic information about cardiovascular risks. In order to examine in a more ac-curate way this relation, in this work we evaluate the influence of gender on the BP/HR relationship by using hour-to-hour 24h ambulatory measurements. Data coming from 122 female and 50 male normotensive subjects were recorded using a Holter Blood Pressure Monitor and the parameters of the linear regression fit-ting BP/HR were calculated. Results confirmed those obtained in previous stud-ies using punctual office measures in males and underlined a significant relation between Diastolic BP and HR during each hour of the day in females; a different trend in the BP/HR relation between genders was found only during night-time. Moreover, the circadian rhythm of BP/HR is similar in both genders but with different values of HR and BP at different times of the day.
△ Less
Submitted 23 November, 2020;
originally announced November 2020.
-
Novel Classification of Ischemic Heart Disease Using Artificial Neural Network
Authors:
Giulia Silveri,
Marco Merlo,
Luca Restivo,
Gianfranco Sinagra,
Agostino Accardo
Abstract:
Ischemic heart disease (IHD), particularly in its chronic stable form, is a subtle pathology due to its silent behavior before developing in unstable angina, myocardial infarction or sudden cardiac death. Machine learning techniques applied to parameters extracted form heart rate variability (HRV) signal seem to be a valuable support in the early diagnosis of some cardiac diseases. However, so far…
▽ More
Ischemic heart disease (IHD), particularly in its chronic stable form, is a subtle pathology due to its silent behavior before developing in unstable angina, myocardial infarction or sudden cardiac death. Machine learning techniques applied to parameters extracted form heart rate variability (HRV) signal seem to be a valuable support in the early diagnosis of some cardiac diseases. However, so far, IHD patients were identified using Artificial Neural Networks (ANNs) applied to a limited number of HRV parameters and only to very few subjects. In this study, we used several linear and non-linear HRV parameters applied to ANNs, in order to confirm these results on a large cohort of 965 sample of subjects and to identify which features could discriminate IHD patients with high accuracy. By using principal component analysis and stepwise regression, we reduced the original 17 parameters to five, used as inputs, for a series of ANNs. The highest accuracy of 82% was achieved using meanRR, LFn, SD1, gender and age parameters and two hidden neurons.
△ Less
Submitted 19 November, 2020;
originally announced November 2020.
-
Smoking effect on the circadian rhythm of blood pressure in hypertensive subjects
Authors:
Giulia Silveri,
Lorenzo Pascazio,
Aleksandar Miladinovic,
Milos Ajcevic,
Agostino Accardo
Abstract:
The use of office measurement of Blood Pressure (BP) as well as of the mean on day-time, on night-time or on 24h does not accurately describe the changes of the BP circadian rhythm. Moreover, several risk factors affect this rhythm but until now possible alterations, due to the presence of such risk factors considered separately, were not been yet studied. Cigarette smoking is one of the most rele…
▽ More
The use of office measurement of Blood Pressure (BP) as well as of the mean on day-time, on night-time or on 24h does not accurately describe the changes of the BP circadian rhythm. Moreover, several risk factors affect this rhythm but until now possible alterations, due to the presence of such risk factors considered separately, were not been yet studied. Cigarette smoking is one of the most relevant risk factors increasing cardiovascular morbidity and mortality. The aim of this study is to evaluate quantitatively and with a suitable temporal detail how the smoking influences the BP circadian rhythm in normotensive and hypertensive subjects excluding those who presented other risk factors like obesity, dyslipidemia and diabetes mellitus. Holter BP monitoring coming from 618 subjects was used and the behaviour on 24h was examined separately in normotensive and hypertensive subjects either smokers or non-smokers. Four intervals with alternate different characteristics were found in the BP rhythm and regression lines approximated them in order to evaluate the changing rate of BP in each period. Results showed higher values from 10:00 to 02:00 in hypertensive smokers than non-smokers and significant differences between normotensive smokers and non-smokers between 10:00 and 19:00. The changing rate between 10:00 and 14:30 was higher in non-smokers than in smokers for both normotensive and hypertensive subjects while the opposite was found in the other three periods. The different velocity rates of BP changes during 24h, could be associated with different risk levels of cardiovascular disease.
△ Less
Submitted 10 November, 2020;
originally announced November 2020.
-
Evaluation of Motor Imagery-Based BCI methods in neurorehabilitation of Parkinson's Disease patients
Authors:
Aleksandar Miladinović,
Miloš Ajčević,
Pierpaolo Busan,
Joanna Jarmolowska,
Giulia Silveri,
Manuela Deodato,
Sussana Mezzarobba,
Piero Paolo Battaglini,
Agostino Accardo
Abstract:
The study reports the performance of Parkinson's disease (PD) patients to operate Motor-Imagery based Brain-Computer Interface (MI-BCI) and compares three selected pre-processing and classification approaches. The experiment was conducted on 7 PD patients who performed a total of 14 MI-BCI sessions targeting lower extremities. EEG was recorded during the initial calibration phase of each session,…
▽ More
The study reports the performance of Parkinson's disease (PD) patients to operate Motor-Imagery based Brain-Computer Interface (MI-BCI) and compares three selected pre-processing and classification approaches. The experiment was conducted on 7 PD patients who performed a total of 14 MI-BCI sessions targeting lower extremities. EEG was recorded during the initial calibration phase of each session, and the specific BCI models were produced by using Spectrally weighted Common Spatial Patterns (SpecCSP), Source Power Comodulation (SPoC) and Filter-Bank Common Spatial Patterns (FBCSP) methods. The results showed that FBCSP outperformed SPoC in terms of accuracy, and both SPoC and SpecCSP in terms of the false-positive ratio. The study also demonstrates that PD patients were capable of operating MI-BCI, although with lower accuracy.
△ Less
Submitted 30 October, 2020;
originally announced November 2020.
-
Transfer Learning improves MI BCI models classification accuracy in Parkinson's disease patients
Authors:
Aleksandar Miladinović,
Miloš Ajčević,
Pierpaolo Busan,
Joanna Jarmolowska,
Giulia Silveri,
Susanna Mezzarobba,
Piero Paolo Battaglini,
Agostino Accardo
Abstract:
Motor-Imagery based BCI (MI-BCI) neurorehabilitation can improve locomotor ability and reduce the deficit symptoms in Parkinson's Disease patients. Advanced Motor-Imagery BCI methods are needed to overcome the accuracy and time-related MI BCI calibration challenges in such patients. In this study, we proposed a Multi-session FBCSP (msFBCSP) based on inter-session transfer learning and we investiga…
▽ More
Motor-Imagery based BCI (MI-BCI) neurorehabilitation can improve locomotor ability and reduce the deficit symptoms in Parkinson's Disease patients. Advanced Motor-Imagery BCI methods are needed to overcome the accuracy and time-related MI BCI calibration challenges in such patients. In this study, we proposed a Multi-session FBCSP (msFBCSP) based on inter-session transfer learning and we investigated its performance compared to the single-session based FBSCP. The main result of this study is the significantly improved accuracy obtained by proposed msFBCSP compared to single-session FBCSP in PD patients (median 81.3%, range 41.2-100.0% vs median 61.1%, range 25.0-100.0%, respectively; p<0.001). In conclusion, this study proposes a transfer learning-based multi-session based FBCSP approach which allowed to significantly improve calibration accuracy in MI BCI performed on PD patients.
△ Less
Submitted 29 October, 2020;
originally announced October 2020.
-
Identification of Ischemic Heart Disease by using machine learning technique based on parameters measuring Heart Rate Variability
Authors:
Giulia Silveri,
Marco Merlo,
Luca Restivo,
Beatrice De Paola,
Aleksandar Miladinović,
Miloš Ajčević,
Gianfranco Sinagra,
Agostino Accardo
Abstract:
The diagnosis of heart diseases is a difficult task generally addressed by an appropriate examination of patients clinical data. Recently, the use of heart rate variability (HRV) analysis as well as of some machine learning algorithms, has proved to be a valuable support in the diagnosis process. However, till now, ischemic heart disease (IHD) has been diagnosed on the basis of Artificial Neural N…
▽ More
The diagnosis of heart diseases is a difficult task generally addressed by an appropriate examination of patients clinical data. Recently, the use of heart rate variability (HRV) analysis as well as of some machine learning algorithms, has proved to be a valuable support in the diagnosis process. However, till now, ischemic heart disease (IHD) has been diagnosed on the basis of Artificial Neural Networks (ANN) applied only to signs, symptoms and sequential ECG and coronary angiography, an invasive tool, while could be probably identified in a non-invasive way by using parameters extracted from HRV, a signal easily obtained from the ECG. In this study, 18 non-invasive features (age, gender, left ventricular ejection fraction and 15 obtained from HRV) of 243 subjects (156 normal subjects and 87 IHD patients) were used to train and validate a series of several ANN, different for number of input and hidden nodes. The best result was obtained using 7 input parameters and 7 hidden nodes with an accuracy of 98.9% and 82% for the training and validation dataset, respectively.
△ Less
Submitted 29 October, 2020;
originally announced October 2020.
