Marco Ciccone
Welcome!
I am Postdoctoral Researcher in the VANDAL group at Politecnico di Torino. I am also part of the ELLIS Postdoc Program working with Prof. Barbara Caputo and Prof. Carlo Ciliberto.
I received my Ph.D. in Computer Science at Politecnico di Milano advised by Prof. Matteo Matteucci and Jonathan Masci, where I focused on developing deep learning models for computer vision problems that perform iterative and conditional representation learning.
During my Ph.D I was also a research intern at NVIDIA and
NNAISENSE.
Previously, I proudly worked at the development of Horus.
News
- April 2023: Invited Lecture @ Bocconi
- March 2023: Serving as NeurIPS 2023 Competition Track co-chair
- October 2022: Invited talk @ DeepMind Meta Learning Reading group
- October 2022: Our new work SCROLL on Online Continual Learning is out!
- September 2022: I am again a lab tutor at M2L. This time finally in-person! See you in Milan!
- August 2022: 3 works on Federated Learning accepted at IROS, ECCV and WACV!
- March 2022: I am serving as Competition co-chair for NeurIPS 2022 with Gustavo Stolovitzky. Stay tuned for the Call for Competitions!
- March 2022: Moved to London at UCL to work with Carlo Ciliberto!
- February 2022: One paper on incremental weakly semantic segmentation accepted to CVPR2022!
- June 2021: I joined the VANDAL group as postdoc researcher!
- April 2021: Call me Doc! I successfully defended by PhD thesis (cum Laude)!
- March 2021: ICLR 2021 Outstanding Reviewer Award!
- March 2021: I am helping Douwe Kiela and Barbara Caputo organizing the Demos and Competitions tracks at NeurIPS 2021! See the Call for Competitions here!
Interests
My research focuses on representation learning and Meta/Continual Learning.
I am currently interested in understanding Meta/Continual Learning algorithms and scaling them to complex tasks. My long term research mission is to understand and build intelligent modular systems that can solve new problems with very weak supervision and few data by re-using and improving previously acquired skills.
I recently gained interest in Federated Learning, especially in understanding and developing algorithms to mitigate the issue of clients statistical heterogeneity.
Although I mainly work on computer vision problems, I have always been fascinated by Game Theory and Multi-Agent Reinforcement Learning (MARL). In particular, I am interested in the study of the learning dynamics of multiple agents and the emergence of cooperative behaviors.
Featured Work
|
A Differentiable Recurrent Surface for Asynchronous Event-Based Data
ECCV, 2020 project page / code / 1min video / 10min video /
We reconstruct a dense feature representation from sparse events using a grid of LSTM cells.
The dense representation can be used as input to standard frame-based CNNs architectures for tasks such as image recognition and optical flow estimation.
|
|
Coordination in Adversarial Sequential Team Games via Multi-Agent Deep Reinforcement Learning
preprint, 2019
We condition team members over signals implementing a coordination-device via hyper-networks
to induce cooperative behaviors and learn to play correlated equilibria.
Joint work with
Andrea Celli,
Raffaele Bongo and
Nicola Gatti
|
|
Asynchronous Convolutional Networks for Object Detection in Neuromorphic Cameras
Event-based Vision and Smart Cameras CVPR Workshop, 2019 (Best Paper Award) Video
We propose a novel approach to passing event-data through a CNN that respects the asynchronous nature of DVS sensors.
|
|
We propose two attentive models for event-based vision: an algorithm that tracks events activity within
the field of view to locate regions of interest and a fully differentiable attention procedure based on
DRAW neural model.
|
|
ReConvNet: Video Object Segmentation with Spatio-Temporal Features Modulation
The DAVIS Challenge on Video Object Segmentation - CVPR Workshop, 2018 Leaderboard
We consider Video Object Segmentation from a Meta-Learning perspective where each task consists of segmenting objects in a video given a single annotation.
We learn how to adapt the activations of a neural network to segment a given object via two modulation networks conditioned on the available information at test time.
|
|
NAIS-Net: Stable Deep Networks from Non-Autonomous Differential Equations
NeurIPS, 2018 Poster / Jack Clark's Import AI newsletter
We propose a non-autonomous architecture where each block is derived from a time-invariant non-autonomous dynamical system.
Each block is asymptotically stable and can be unrolled indefinitely upon convergence to an input-dependent attractor.
|
|
ReSeg: A Recurrent Neural Network-Based Model for Semantic Segmentation
DeepVision CVPR Workshop, 2016 - (Best Paper Award) Code
We propose to use recurrent layers (ReNet)
to capture global and local context in images and improve semantic segmentation performance.
Joint work with
Francesco Visin,
Adriana Romero,
Kyle Kastner,
Kyunghyun Cho,
Yoshua Bengio,
Matteo Matteucci and
Aaron Courville
|
|
My neural baptism: recurrent and convolutional networks for semantic segmentation.
Advised by
Francesco Visin and
Matteo Matteucci
|
Personal
In my spare time, I practice yoga, read novels, and play guitar. When I need to relax, I cook. When I'm not lazy, I go hiking and trekking. When I am sad or happy, I sing. When I can, I travel the world to learn about new cultures.
If you want to discuss anything or just connect, please drop me an email or reach me on Twitter or Linkedin!