I am currently working in deep learning, and deep reinforcement
learning: see also my
blog.
Some of my recent works have been focused on the following
topics:

Variational Autoencoders I recently did quite a lot of research on Variational Autoencoders, addressing the variablecollapse phenomenon, the potential mismatch between the aggregate posterior distribution Q(z) and the prior P(z), and the balancing problem between reconstruction error and KullbackLeibler divergence. Our new balancing strategy allowed us to get the best generative scores in terms of FID ever obtained with a variational approach.
See also my blog on Variational Autoencoders.



Playing Rogue with reinforcement learning techniques. Rogue is a famous dungeoncrawling videogame of the 80ies, the ancestor of its gender.
Roguelike games are known for the necessity to explore partially observable
and randomlygenerated labyrinths, preventing any form of level replay.
As such, they serve as a very natural and challenging task for reinforcement learning, requiring the acquisition of complex, nonreactive behaviors involving memory and planning.



In 2018, we took part to the Audi Autonomous Driving Cup (AADC).
The contest consists in developing fully automatic driving functions for a 1:8 model vehicle provided by Audi.
For the mandatory program, each team is supposed to complete a list of manouvres on a simple route specified in openDriveformat.
The freestyle part of this year competition explicitly required the application of deep learning and deep reinforcement learning techniques.



Automatic pointofinterest image cropping via ensembled convolutionalization Convolutionalization of discriminative neural networks, introduced by J.Long et al. for segmentation purposes, is a simple technique allowing to generate heatmaps relative to the location of a given object in a larger image. We apply this technique to automatically crop images at their actual point of interest. The use of an ensemble of fully convolutional nets sensibly reduce the risk of overfitting, resulting in reasonably accurate croppings.



Detection of Gastrointestinal Diseases from Endoscopical Images The lack, due to privacy concerns, of large public databases of medical pathologies is a wellknown and major problem, substantially hindering the application of deep learning techniques in this field. In this research, we investigate the possibility to supply to the deficiency in the number of data by means of data augmentation techniques.


Previous Research
Sometimes, I am a bit puzzled myself by the different topics I have been working
on during my scientifc career. So, to get some sense out of it, I decided to draw
a picture.
In the end, I think there is a clear line of development, I was not entirely
aware of.
I have always been interested in machine intelligence, but my first studies
have been on the logical side: lambda calculus, type theory, category theory.
From this I got involved in linear logic, optimal reduction (the BOHM machine),
implicit computational complexity.
Then I moved to more concrete topics: mathematical knowledge representation and
mechanization of formal reasoning (see my Interactive Prover Matita).
Neural Networks always intrigued me, but with the advent of
deep learning I also decided to devote some research effort in it, and
his teaching.
This does not mean I am abjuring my old topics. Actually,
I think that the integration between machine learning and deduction remains
one of the big scientific challenges for the future: let the machine learn to
prove theorems or equivalently, by the CurryHoward analogy, to write its own
programs.
Please, also have a look at my reviews of the
most influential scientifc books in my profession career.