Nicola Zomer

Nicola Zomer

Research engineer in applied data science and AI

Posts by Collection

portfolio

projects

Binary star evolution and binary black holes

Published:

A Binary Black Hole (BBH) can originate from the evolution of a massive binary star. If the binary is tight enough, it will evolve through several complex physical processes: the two stars can exchange mass either via Stable Mass Transfer (SMT) or via a Common Envelope (CE).

Sound wave distortion via FPGA using Pmod interface

Published:

In this project we implemented on a Field Programable Gate Array (FPGA) a distortion effect in sound waves that is called “Overdrive” or “Clipping”. This happens when the amplitude of a soundwave is restricted when it exceeds a given threshold. The resulting sounds are “dirty” and “fuzzy” due to the introduction of high frequency components in the signal.

Explaining microbial scaling laws using Bayesian inference

Published:

In this project, we combined methods from Statistical Physics and Bayesian Data Analysis to elucidate the principles behind cellular growth and division. We studied various classes of individual-based growth-division models and inferred individual-level processes (model structures and likely ranges of associated parameters) from sigle-cell observations.

Koln traffic regulator with parallel computing

Published:

Our work started from a project jointly developed by IBM and by the German city of Koln thought to be a first step towards traffic regulation and an efficient exploitation of transport’s resources. In particular, we analyzed a set of mobility data emulated with SUMO, consisting of 394 million records and 20 Gb in size. To reach our goals, we set up a cluster on CloudVeneto made of 5 virtual machines (4 cores and 8 GB RAM each) and created a volume, shared across the instances using a NFS. Moreover, we used Dask to parallelize the tasks.

Effective processing pipeline and advanced neural network architectures for small-footprint keyword spotting

Published:

The keyword spotting (KWS) task consists of identifying a relatively small set of keywords in a stream of user utterances. This is preferably addressed using small footprint inference models that can be deployed even on performance-limited and/or low-power devices. In this framework indeed, model performance is not the only relevant aspect and the model footprint plays an equally crucial role. In this work, first we defined a modern CNN model that outperforms our baseline model and we used it to study the impact of different pre-processing, regularization and feature extraction techniques. We saw how, for instance, the log Mel-filterbank energy features lead to the best performance and we discovered that the introduction of background noise on the train set with an optimal noise reduction coefficient of 0.5 helps the model to learn. Then, we explored different machine learning models, such as ResNets, RNNs, attention-based RNNs and Conformers in order to achieve an optimal trade-off between accuracy and footprint. We found that these architectures offer between a 30-40% improvement in accuracy compared to the baseline, while reducing up to 10× the number of parameters. We ran our tests on the Google Speech Commands dataset, one of the most popular datasets in the KWS context.

Data structure-oriented learning strategies for 3D objects classification

Published:

Nowadays, the identification and understanding of 3D objects in real-world environments has a wide range of applications, including robotics and human-computer interaction. This is typically addressed using Deep Learning techniques that deal with 3D volumetric data, as they are generally able to outperform standard Machine Learning tools.

publications

Enhancing stop location detection for incomplete urban mobility datasets

Published in ArXiv, 2024

Stop location detection, within human mobility studies, has an impacts in multiple fields including urban planning, transport network design, epidemiological modeling, and socio-economic segregation analysis. However, it remains a challenging task because classical density clustering algorithms often struggle with noisy or incomplete GPS datasets. This study investigates the application of classification algorithms to enhance density-based methods for stop identification. Our approach incorporates multiple features, including individual routine behavior across various time scales and local characteristics of individual GPS points. The dataset comprises privacy-preserving and anonymized GPS points previously labeled as stops by a sequence-oriented, density-dependent algorithm. We simulated data gaps by removing point density from select stops to assess performance under sparse data conditions. The model classifies individual GPS points within trajectories as potential stops or non-stops. Given the highly imbalanced nature of the dataset, we prioritized recall over precision in performance evaluation. Results indicate that this method detects most stops, even in the presence of spatio-temporal gaps and that points classified as false positives often correspond to recurring locations for devices, typically near previous stops. While this research contributes to mobility analysis techniques, significant challenges persist. The lack of ground truth data limits definitive conclusions about the algorithm’s accuracy. Further research is needed to validate the method across diverse datasets and to incorporate collective behavior inputs.

talks

teaching

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.