Academic projects repositories

Fed2RC

In this work, we have developed a technique for federated time series classification based on Rocket features. Rocket is a method that exploits random mappings to accomplish time series classification tasks, and it is based on random kernel selection. We developed a FL system that shares only the generator of the random kernel and finds a closed-form solution for the general Ridge classifier, improving communication cost, computation cost and performances. GitHub

Hyperbolic Benchmark

In this work, we have benchmarked different geometries, answering one key question: when do hyperbolic geometries give you something for free? Not often. We found that, differently from what said by many papers, hyperbolic spaces are not magical, and they don’t bring improvements if not specifically tailored. We benchmarked several metric: accuracy, robustness and OOD detection in image classification. GitHub

Personalized Aggregation for Federated Prototypical Learning

In this work, we developed a Federated Learning strategy to optimize Federated systems that exchange only prototypes under adversarial conditions, i.e. with model poisoning attacks. We have introduced a novel technique to prevent these type of attacks and considered usually neglected metrics, i.e. fairness. GitHub

Hyperbolic Prototypical Entailment Cones

In this work, we have introduced a novel hyperbolic method based on Hyperbolic Entailment Cones to improve the performances of Hyperbolic Representation Learning techniques using prototype learning. In particular, we have introduced a novel metric based on entailment cones to learn and a novel clipping technique called backclip. GitHub

fluke

Federated Learning Utility frameworK for Experimentationg is our framework used for experimenting and fast prototyping of novel FL techniques. My main part in this project has been testing and improvements. GitHub

Federated Survival Analysis

In this work, we have developed an adaptation of Adaboost.F to survival analysis in a FL context, which required redesigning some part of the original algorithm, and which outperformed the state-of-the-art methods in FL survival analysis. GitHub

Parallel Interior Point Solver

We implemented an interior point solver (Karmarkar’s Interior Point method) and compared its implementation under different parallel settings: MPI, PSTL, CUDA. While achieving the best performance with tailored CUDA code, using PSTL made the code easily portable from CPU-only regimes to GPUs, while CUDA requested a complete code revision. GitHub

Personal projects repositories

Taiwanese Credit Risk

In this project we implemented a credit risk model trying to predict wether a customer could return a loan or not. We benchmarked many ML algorithms and conducted carefull EDA. GitHub

Digital Wallet Transactions

In this project I implemented a full end-to-end ML pipeline, including EDA, model development for fraud detection and an interface using fastAPI and Streamlit. GitHub

More

For a complete and up-to-date list of repositories, visit my
GitHub profile