Alessio Martino
Alessio Martino graduated summa cum laude in Communications Engineering at University of Rome "La Sapienza", Italy, October 2016. His Bachelor and Master's Degree Theses regarded EU-FP7 and EU-FP8 projects, respectively. From November 2016 to October 2019, he served as PhD Research Fellow in Information and Communications Technologies at the same University (Department of Information Engineering, Electronics and Telecommunications), with a final dissertation on pattern recognition techniques in non-metric domains. During his PhD, he also served as scientific collaborator with Consortium for Research in Automation and Telecommunication, Rome, Italy.
After obtaining the PhD, he has been granted a 1-year PostDoctoral Research Fellowship at University of Rome "La Sapienza" and a 1-year PostDoctoral Research Fellowship at the Italian National Research Council. Currently, he is Assistant Professor of Computer Science at LUISS University.
His research interests include machine learning, computational intelligence and knowledge discovery. Currently he's focusing on large-scale machine learning, advanced pattern recognition systems, big data analysis, parallel & distributed computing, granular computing and complex systems modelling, in applications including bioinformatics and computational biology, natural language processing and energy distribution networks.
After obtaining the PhD, he has been granted a 1-year PostDoctoral Research Fellowship at University of Rome "La Sapienza" and a 1-year PostDoctoral Research Fellowship at the Italian National Research Council. Currently, he is Assistant Professor of Computer Science at LUISS University.
His research interests include machine learning, computational intelligence and knowledge discovery. Currently he's focusing on large-scale machine learning, advanced pattern recognition systems, big data analysis, parallel & distributed computing, granular computing and complex systems modelling, in applications including bioinformatics and computational biology, natural language processing and energy distribution networks.
less
Interests
Uploads
Papers by Alessio Martino
Dense cross-layer connectivity can shorten gradient paths and promote feature reuse, potentially improving optimization under fixed training budgets.
Objective
We test whether concatenation-based dense historical connectivity improves decoder-only autoregressive language modeling under controlled comparison protocols.
Methods
We compare a standard Transformer decoder and a dense decoder on Penn Treebank and WikiText-2 under two fairness regimes: (i) a same training recipe setting with a fixed baseline and a bounded dense architectural search, and (ii) a same parameter budget setting where the dense model is resized to not exceed the baseline parameter count.
Results
Dense connectivity does not consistently reduce test perplexity; on WikiText-2, the baseline remains better in both regimes, while gains on Penn Treebank are small and regime-dependent. Ablations within the dense family show that depth and feed-forward capacity are the most reliable drivers of perplexity improvements.
Conclusions
Probes and attention diagnostics do not reveal a clear advantage for dense connectivity in our limited probe set, while Zipf–RQA analysis of long-form generations reveals systematic structural differences between baseline and dense outputs. Specifically, Zipf–RQA is used here as a descriptive structural probe rather than a performance metric.
To apply a machine learning analysis to clinical and presynaptic dopaminergic imaging data of patients with rapid eye movement (REM) sleep behavior disorder (RBD) to predict the development of Parkinson disease (PD) and dementia with Lewy bodies (DLB).
Methods:
In this multicenter study of the International RBD study group, 173 patients (mean age 70.5 ± 6.3 years, 70.5% males) with polysomnography-confirmed RBD who eventually phenoconverted to overt alpha-synucleinopathy (RBD due to synucleinopathy) were enrolled, and underwent baseline presynaptic dopaminergic imaging and clinical assessment, including motor, cognitive, olfaction, and constipation evaluation. For comparison, 232 RBD non-phenoconvertor patients (67.6 ± 7.1 years, 78.4% males) and 160 controls (68.2 ± 7.2 years, 53.1% males) were enrolled. Imaging and clinical features were analyzed by machine learning to determine predictors of phenoconversion.
Results:
Machine learning analysis showed that clinical data alone poorly predicted phenoconversion. Presynaptic dopaminergic imaging significantly improved the prediction, especially in combination with clinical data, with 77% sensitivity and 85% specificity in differentiating RBD due to synucleinopathy from non phenoconverted RBD patients, and 85% sensitivity and 86% specificity in discriminating PD-converters from DLB-converters. Quantification of presynaptic dopaminergic imaging showed that an empirical z-score cutoff of −1.0 at the most affected hemisphere putamen characterized RBD due to synucleinopathy patients, while a cutoff of −1.0 at the most affected hemisphere putamen/caudate ratio characterized PD-converters.
Interpretation:
Clinical data alone poorly predicted phenoconversion in RBD due to synucleinopathy patients. Conversely, presynaptic dopaminergic imaging allows a good prediction of forthcoming phenoconversion diagnosis. This finding may be used in designing future disease-modifying trials.
MRI studies reported that ALS patients with bulbar and spinal onset showed focal cortical changes in corresponding regions of the motor homunculus. We evaluated the capability of brain 2-[18F]FDG-PET to disclose the metabolic features characterizing patients with pure bulbar or spinal motor impairment.
Methods
We classified as pure bulbar (PB) patients with bulbar onset and a normal score in the spinal items of the ALSFRS-R, and as pure spinal (PS) patients with spinal onset and a normal score in the bulbar items at the time of PET. Forty healthy controls (HC) were enrolled. We compared PB and PS, and each patient group with HC. Metabolic clusters showing a statistically significant difference between PB and PS were tested to evaluate their accuracy in discriminating the two groups. We performed a leave-one-out cross-validation (LOOCV) over the entire dataset. Four classifiers were considered: support vector machines (SVM), K-nearest neighbours, linear classifier, and decision tree. Then, we used a separate test set, including 10% of patients, with the remaining 90% composing the training set.
Results
We included 63 PB, 271 PS, and 40 HC. PB showed a relative hypometabolism compared to PS in bilateral precentral gyrus in the regions of the motor cortex involved in the control of bulbar function. SVM showed the best performance, resulting in the lowest error rate in both LOOCV (4.19%) and test set (9.09 ± 2.02%).
Conclusions
Our data support the concept of the focality of ALS onset and the use of 2-[18F]FDG-PET as a biomarker for precision medicine-oriented clinical trials.