/ 01
Binary & Quantized Networks
Training networks that abandon floating point almost entirely — cyclic-precision schedules (CycleBNN), binary backbones for real-time perception, and quantization that survives deployment.
Postdoctoral Researcher · Sapienza University of Rome
Federico Fontana
Efficient & Trustworthy Deep Learning · Computer Vision Laboratory, Sapienza
I make deep learning extremely efficient — compressing state-of-the-art networks into binary and low-bit forms that run where full-precision models cannot — and I ask whether what survives that compression can still be trusted: over time, under attack, and when a model must forget.
- Citations
- 228+
- h-index
- 6
- Publications
- 14
- Countries collaborated
- 9
01 — About
Small models, full accuracy
I am a postdoctoral researcher at Sapienza University of Rome, where I earned my Ph.D. in efficient deep learning in January 2026 within the Computer Vision Laboratory. My work sits exactly where deep learning meets its hardware limits: I design networks that retain state-of-the-art accuracy while shedding the compute, memory and energy that modern models take for granted.
Two questions drive the work. First, how far can a network be compressed — binary and low-bit weights, structured pruning, distillation — before it stops working. Second, whether what survives that compression stays trustworthy: robust as the world shifts under it, able to spot synthetic media, and able to forget on demand.
I am first or sole author on the core efficiency methods, published at CVPR, ECCV, ICML and in Q1 journals including IEEE Transactions on Artificial Intelligence and the International Journal of Neural Systems. The work has been recognised with a Best Paper Award at ICVSS 2024, funded by competitive grants I hold as principal investigator, and released as open source that the community actually uses.
The collaborations are self-initiated and span nine countries — from a research visit I arranged myself in Klagenfurt to joint work with KTH, Imperial College London, Warwick, UAB Barcelona and others. I am always open to new ones, across academia and industry.
International collaborations
02 — Research
Research directions
A single thread runs through all of them: getting more out of neural networks with radically less computation, memory and energy.
/ 02
Pruning & Compression
Distilled gradual pruning with pruned fine-tuning: removing redundant structure while a teacher keeps the student honest, producing compact models for resource-constrained hardware.
/ 03
Deepfake Detection
Real-time detection on binary backbones, plus a chronological continual-learning view of the problem that exposes how poorly detectors generalise to tomorrow's generators.
/ 04
Machine Unlearning
A unified f-divergence framework for removing concepts from diffusion models — the efficiency lens applied to what a generative model should be able to forget.
/ 05
UAV & Cross-View Geolocalization
Matching drone, SAR and satellite imagery on board, under strict latency and power budgets — from an industrial internship at MBDA to BiCrossNet with the University of Klagenfurt.
/ 06
Neuro-AI
Transformer architectures for EEG: subject-aware emotion recognition (SATEER) and neural transcoding from EEG to fMRI (NT-ViT), with KTH and the University of Calabria.
03 — Experience
Career
2026 — present
Postdoctoral Researcher
Sapienza University of Rome · Computer Vision Laboratory — Rome, Italy
- Research on extreme quantization, binary neural networks, model compression and deepfake detection.
- Secures competitive funding, publishes at top-tier venues and presents at international conferences.
- Lectures and co-supervises B.Sc. and M.Sc. students within the group.
2022 — 2026
Doctoral Researcher — Efficient Deep Learning
Sapienza University of Rome · Computer Vision Lab — Rome, Italy
- Doctoral research on compressing deep networks — binarization, pruning and distillation — without sacrificing accuracy.
- Published across CVPR-W, ECCV, IEEE T-AI, IEEE Access, MLST and ACM SIGGRAPH.
- Best Paper Award, ICVSS 2024; Sapienza Research Grants in 2023 and 2024.
- Supervisors: Prof. Danilo Avola, then Prof. Luigi Cinque.
2024 · 3 months
Visiting Researcher
University of Klagenfurt (AAU) · Klagenfurt, Austria
Self-initiated research visit on efficient cross-view geolocalization, resulting in BiCrossNet — a binary neural network approach published in Machine Learning: Science and Technology.
11/2022 — 04/2023
Machine Learning Scientist (Internship)
MBDA Italia · Aerospace & defence — Rome, Italy
Responsible for the cross-view geolocation algorithm matching SAR and RGB imagery, designed for real-time, low-power integration on board.
01/2019 — 06/2019
Machine Learning Scientist (Internship)
Soonapse · Agri-tech — Rome, Italy
Machine learning applied to water management in agriculture: data mining and predictive modelling for irrigation decisions.
01/2018 — 12/2018
App Software Engineer
Nextrek s.r.l. · Rome, Italy
Lead developer on mobile application development projects.
01/2016 — 12/2016
Software Engineer — Web
INFN — Istituto Nazionale di Fisica Nucleare · Frascati, Italy
Gathered requirements from research groups and built flexible WordPress plugins and PHP tooling so scientists could present their experiments online.
04 — Education
Academic background
Jan 2026
Ph.D. in Efficient Deep Learning
Sapienza University of Rome · 2022–2026
Binarization, pruning and distillation for resource-constrained vision. Advisors: Prof. Danilo Avola, then Prof. Luigi Cinque.
Jul 2022
M.Sc. in Computer Science 110/110 cum laude
Sapienza University of Rome · LM-18
Thesis: Training binary neural networks without floating-point precision. Selected for the Department of Computer Science Path of Excellence.
Schools & training
- 2025ACACES 2025 — HiPEAC summer school on high-performance computer architecture & compilation
- 2024ICVSS 2024 — International Computer Vision Summer School, Sicily (Best Paper Award)
- 2023AI-DLDA — International Summer School on AI, University of Udine (student grant)
Funding & honours
- 202680,000 GPU hours on CINECA Leonardo — competitive compute allocation
- 2024Best Paper Award — ICVSS 2024
- 2023–25Sapienza “Avvio alla Ricerca” grants (principal investigator); Sapienza Research Grants 2023, 2024
- 2015€40,000 innovation grant — Regione Lazio “Creativi Digitali / APP-ON”, for the SMART CHARGING app (won as a teenage founder)
- 2022Path of Excellence — Dept. of Computer Science, Sapienza
- 2016–22LazioDisCo merit grants
Teaching, talks & service
-
2024–25
- Guest lecturer, Deep Learning — ACSAI, Sapienza
- Guest lecturer, Computer Vision — Computer Science, Sapienza
- Co-supervised ~10 B.Sc./M.Sc. students
-
2026
Invited seminar, Digital Futures @ KTH, Stockholm — “Binary Neural Networks: From the Roots to Trustworthy Intelligence on the Edge”
-
2024
- Invited talk, University of Klagenfurt (AAU), Austria
- Speaker, CADL @ ECCV, Milan — hosted by G. Fiameni (NVIDIA)
- Co-speaker, DFAD @ CVPR, Seattle
- Speaker, Rome Technopole Spoke 6 — International Young Researcher Workshop
-
ongoing
- Reviewer — CVPR, ECCV, ICCV, ICML, ICLR, WACV (4 years)
05 — Publications
Selected work
228 citations · h-index 6 · full and up-to-date list on Google Scholar.
-
2026
A Unified Framework for Diffusion Model Unlearning with f-Divergence
-
2026
DualFocusNet: Dual-Stream Spatio-Temporal Transformer for Generalizable Deepfake Detection
-
2025
BiCrossNet: Resource-Efficient Cross-View Geolocalization with Binary Neural Networks
-
2025
SATEER: Subject-Aware Transformer for EEG-Based Emotion Recognition
-
2025
Revisiting Deepfake Detection: Chronological Continual Learning and the Limits of Generalization
-
2024
CycleBNN: Cyclic Precision Training in Binary Neural Networks
-
2024
Faster Than Lies: Real-Time Deepfake Detection using Binary Neural Networks
-
2024
Distilled Gradual Pruning with Pruned Fine-Tuning
-
2024
Semantically Guided Representation Learning for Action Anticipation
-
2024
NT-ViT: Neural Transcoding Vision Transformers for EEG-to-fMRI Synthesis
-
2024
UAV Geo-Localization for Navigation: A Survey
-
2024
BNNAction-Net: Binary Neural Network on Hand Gesture Recognition
-
2023
Hand Gesture Recognition Exploiting Handcrafted Features and LSTM
-
2023
Training Binary Neural Networks without Floating Point Precision
Citation counts from Google Scholar, August 2026.
06 — Skills
Expertise
Research domains
Programming
Frameworks & tooling
Academic practice
Languages
- Italian Native
- English B2 / C1 — professional
07 — Contact
Let's build smaller models
Open to collaboration on efficient deep learning across academia and industry — joint papers, research visits, student projects and industrial pilots.
fontana.f@di.uniroma1.it