Financed scholarships:

• Artificial intelligence applications to physics research
• Large scale ozone risk assessment for vegetation, from past years to the end of the century under different climate change scenarios
• Study of Cooperative Optical Phenomena in Artificial Solids of Nanocrystalline Metal Halides
• Learning by Prediction and Integration: Human-inspired Approaches for Natural Language Understanding

Artificial intelligence applications to physics research

Student: Cassio Cristani

Background and motivation

This research project lays in the flourishing intersection between Computer Science and Physics. In last years it has been observed that machine learning is coming to hold a crucial position in fields of physical science ranging from particle physics to cosmology, quantum many-body physics, quantum computing, and chemical and material physics [1].
Examples of successful applications of machine learning techniques to physics research include, among others: discovering gravitational lenses [2], constructing effective variational many-body wavefunctions in interacting systems with long-range entanglement [3], identifying phase transitions from entanglement spectra [4], characterizing dynamical phases in closed quantum systems where statistical mechanics is not applicable [5], etc. 
On the other hand, machine learning development can benefit from physics research, in particular for the understanding of internal operations. In fact, one of the major open problems in machine learning is related to explainability, in particular with deep neural networks: these algorithms can perform very well in some tasks (e.g., classification) but it’s difficult to understand exactly why, both in theory and in practice.  Examples of how machine learning theoretical development can exploit physical research are the recent proposal of an explanation for the performance of deep learning based on renormalization group theory [6], and an approach to clustering based on statistical physics [7].
The goals of the present project are: exploring the connections between physics and computer science, developing ad hoc machine learning models for physics research, and proposing new techniques for explainable machine learning.

Supervisors

Dr Enrico Barbierato, UCSC, Italy, enrico.barbierato@unicatt.it
Prof. Yi-Ting Hsu, ND, USA, yhsu2@nd.edu

Large scale ozone risk assessment for vegetation, from past years to the end of the century under different climate change scenarios

Student: Pierluigi Guaita

Background and motivation

Ozone (O3) is a secondary pollutant that can cause visible injuries to vegetation in general, but also relative growth reductions in forest plants and crop yield losses in agricultural species1. The phytotoxic ozone dose (POD) index seems the best suited for the estimation of the impact of stomatal ozone deposition under future climate. The project will be focused on modeling activities aimed at defining an integrated procedure based on meteo-chemical data, spatialization techniques and deposition models to produce regional scale maps for the ozone risk assessment for vegetation in Europe.

The main goals of the project are:

• the calibration and validation of a model for the estimation of the phytotoxic ozone dose absorbed by vegetation (wheat and poplar will be used as reference receptors for crop and forest species, respectively) at 1km2 resolution in non-complex terrain;
• the production of O3 risk assessment maps for vegetation based on future climate change scenarios in order to outline possible mitigation strategies for the O3 impacts.

The candidate is expected to acquire the knowledge in the fields of environmental data elaboration, atmospheric pollutants deposition, plant physiology and soil water dynamics modelling, risk assessment.

Supervisors

Prof. Giacomo Gerosa, UCSC, Italy, giacomo.gerosa@unicatt.it
Prof. Paola Crippa, ND, USA, pcrippa@nd.edu ; 
Dr. Riccardo Marzuoli, UCSC, Italy, riccardo.marzuoli@unicatt.it

Study of Cooperative Optical Phenomena in Artificial Solids of Nanocrystalline Metal Halides

Student: Umberto Filippi

Background and motivation

Light-emitting nanocrystals of metal halides provide an opportunity to engineer artificial solids with a cooperative emission. Such solids, known as colloidal nanocrystal superlattices, are a promising platform for applications in quantum photonics and information science. The project will address two important aspects of colloidal nanocrystal superlattices and light emission from them.
First, there is a need to discover single- and multi-component nanocrystal solids made of recent generations of colloidal nanocrystals that include (but are not limited to) lead halide perovskites, lead-free double perovskites, and lead chalcohalides. Despite rapid progress in synthesis and optimization of such nanocrystals, little is known about the diversity of colloidal superlattices that they can form.
Second, there is a need to investigate the cooperative response of such superlattices upon interaction with optical excitation. For example, there are preliminary indications that pulsed laser excitation of cesium lead bromide nanocrystals produces intensity-dependent oscillations and acceleration of radiative decay. Such response could be indicative of cooperative effects, whether that is the case is not well understood.
In the present project we will combine state-of-the-art facilities in the partner institutions. The synthesis and characterization of the colloidal superlattices will be carried out at the Italian Institute of Technology (IIT) in Genova (supervisors: prof. Liberato Manna and Dr. Dmitry Baranov). The investigation of collective and cooperative phenomena in colloidal superlattices will be carried out in collaboration with the group of prof. Claudio Giannetti at Università Cattolica (UCSC) in Brescia and in collaboration with the group of prof. Masaru Kuno at Notre Dame University (ND). A specific multidimensional optical spectroscopy setup will be developed @UCSC to investigate the decoherence dynamics of optical excitons and hunt for evidence of changes in the ultrafast decoherence driven by cooperative phenomena.

The main goals of the project are:

• Synthesis and characterization of novel colloidal superlattices
• Investigation of collective and cooperative phenomena in colloidal superlattices excited by tunable ultrafast light pulses in the weak and strong excitation regimes to manage cooperative absorption and emission on ultrafast timescales.  

Supervisors

Prof. Liberato Manna (IIT) Italy, liberato.manna@iit.it
Dr. Dmitry Baranov (IIT) Italy, dmitry.baranov@iit.it
Prof. Masaru Kuno (ND) USA, mkuno@nd.edu
Prof. Claudio Giannetti (UCSC) Italy, claudio.giannetti@unicatt.it

Learning by Prediction and Integration: Human-inspired Approaches for Natural Language Understanding

Student: Vladimir Araujo

Background and motivation

Giving machines the skills to represent and understand natural language for applications in the real world presents a significant challenge in the area of Natural Language Processing. Pre-trained language models based on neural networks have recently achieved outstanding performance in several natural language understanding tasks. Although effective, these models lack the ability humans possess to text comprehension. For example, as we read, we can anticipate what content may come next or use prior knowledge to better understand a passage.

We hypothesize that current language models could benefit from human language processing mechanisms. In this work, we investigate and propose different approaches to improve current language models, drawing inspiration from prediction and integration theories of human language comprehension. Our contributions show that augmenting models with human mechanisms leads to improvements in natural language understanding across various tasks.

First, we extend the architecture of pre-trained language models with insights from predictive coding theory. We demonstrate that introducing bottom-up and top-down computation to predict future sentences in latent space in the neural networks improves sentence and discourse-level representations. We also validate the generalization of our models in Spanish by using benchmarks and pre-trained models developed by us.

Second, we adapt and propose memory population methods for pre-trained language models under the lifelong learning with episodic memory paradigm. We show that a process that samples the entire data distribution works well enough to integrate previous knowledge and prevent forgetting in the neural network. Furthermore, we also found that some tasks benefit more from selective-based population methods.

Third, extending our second contribution, we propose a method to deal with the stability- plasticity dilemma that occurs in lifelong learning with memory. We show that entropy can be used as a plasticity factor to decide how much a layer in a neural network must be modified according to the current input. We found that this process not only improves performance on text and image classification tasks but also promotes efficiency in the model.

Fourth, we propose a method that incorporates prediction and integration ideas to solve the question answering task in a data stream. Specifically, our approach leverages cross- attention mechanisms to integrate information into external memory. This memory is supported by anticipation and rehearsal pretext tasks. We show that our model can improve the memorization of text-based and video-based sequences.

In summary, we present several approaches that follow ideas from human language processing, demonstrating that human inspiration can be a way to improve the current state of language models based on neural networks. By including human-based mechanisms, we bolster or add some abilities that models do not have and are key to obtaining human-like language processing.

Supervisors

Prof. Álvaro Soto, PUC, Chile, asotoa@uc.cl
Prof. Marie-Francine Moens, KU Leuven, Belgium, sien.moens@cs.kuleuven.be