- 📍 Rome, 🇮🇹
- 🎓 BSc and MSc in CS @ Sapienza University of Rome
- 📚 High-performance computing, machine learning, and deep learning
- 🎓 PhD student in deep learning @ 🎭 GLADIA Research Group
- 🔬 Model merging, audio generation, and music editing
- Free time fav activities: 📺, 🎶, 🧘🏻♂️, 🧠, 📖, 🏃🏻♂️
| Name | Paper | TLDR |
|---|---|---|
| 🏧 ATM | 📜 | Demonstrate and leverage task vectors and multi-task gradients relationship to improve model merging by up to 20% |
| Name | TLDR |
|---|---|
| Kindergarten-VQ-VAE | Devised custom disentanglement modules for Transformers |
| Biscuits | Achieved 75% of original validation accuracy using just 1% of the parameters, inspired by Training BatchNorm and Only BatchNorm |
| CreActive | Devised techniques to observe statistics, patterns, and behaviors of Deep Neural Networks activations |
| GPHungarian | 800x speedup via custom, CUDA-accelerated version of the Hungarian Algorithm |
| Name | TLDR |
|---|---|
| AML_Homework | Devised novel, custom Spatio-Temporal Graph Attention Transformers and CNNs for 3D human pose estimation |
| Big Data FC | Devised novel representations for football teams, implemented custom ML algos to predict end-of-the-season ranking, and coded a scraper for '07-'12 seasons |
| Braver Cat | Reduced train time by over 60% via Transfer Learning in medical image segmentation |
| Cheap Clasp | Devised custom algorithms to use JPEG image encoder as an alternative to CLIP-like Neural image encoders |
| Distillo, Ergo Sum | Improved student model performances by 10% while using only 10% of its parameters using (knowledge distillation) on medical images segmentation |
| GeNNuS | Achieved 87% validation accuracy on a music genre classification task, and solved vanishing gradient issues |