Paolo Frasconi

Associate Professor of Computer Science

Machine Learning and Neural Networks Group
Dipartimento di Sistemi e Informatica, Università degli Studi di Firenze

• Publications
• Books
• Probabilistic Inductive Logic Programming: Theory and Applications (Springer, 2008).
• Modeling the Internet and the Web: Probabilistic Methods and Algorithms (Wiley, 2003).
• Artificial Intelligence and Heuristic Methods in Bioinformatics (IOS Press, 2003).
• CV
• Contact datails
Address:
Via di Santa Marta, 3
I-50139, Firenze – Italy
T: +39 055 4796 362
F: +39 055 4796 363
Office:
See map
Email:
• Teaching
• Apprendimento Automatico (Machine learning) Laurea magistrale in Ingegneria Informatica
• Algoritmi e Strutture Dati (Algorithms and Data Structures) CdL in Ingegneria Informatica

Research Interests

Machine Learning

+ Kernels on logical representations and discrete data structures

The aim is to bridge statistical learning and symbolic approaches such as Inductive Logic Programming. kFOIL (Landwehr et al. 2006) can be seen as an approach to learn the kernel matrix based on FOIL search. Kernels on Prolog Proof Trees compare two instances by comparing the corresponding execution traces of a special visitor program that allows the incorporation of background knowledge (Passerini et al., 2006). The method uses kernels on Prolog ground terms (Passerini & Frasconi 2005) as a sub-component. Declarative kernels are based on mereotopology (the logical foundation of parts, wholes, and places) and allow to program, in a declarative style and possibly exploiting background knowledge, what parthood and connection relations should be used to specify the kernel (Frasconi et al., 2005).Weighted Decomposition Kernels (Menchetti et al., 2005) attempt to control the trade-off between expressivity and generalization on sequence and graph kernels. They have been succesfully applied to protein subcellular localization (Costa et al., 2007) and small-molecule classification (Ceroni et al., 2007).

+ Multiclass margin-based classifiers

Error correcting output codes for (Passerini et al. 2004).

+ Neural networks and graphical models for sequential and structured data

Local Feedback RNNs (Frasconi et al. 1992)IOHMMs (Bengio & Frasconi, 1995, 1996),Bidirectional Recurrent networks (Baldi et al, 1999),Recursive Neural Networks (Frasconi et al., 1998). Prior knowledge in RNNs (Frasconi et al. 1995,1991). The difficulty of learning long-term dependencies (Bengio et al. 1994), (Hochreiter et al. 2001).

Bioinformatics

+ Protein structure prediction

Prediction of protein coarse contact maps, a representation of spatial neighborhood relations between pairs of secondary structure elements. The prediction tool is based onrecursive neural networks (Vullo & Frasconi, 2002,2003, Pollastri et al. 2002). Currently we are working on reconstruction of 3D structures from coarse maps. Some preliminary results are in (Ceroni & Frasconi 2004). In the past we have applied bidirectional recursive neural networks to secondary structure prediction (Baldi et al, 1999). The architecture is at the core of the predictors SSPro andPorter. Recently we have developed a new architecture that can exploit (partial) knowledge on residue contacts to improve prediction of secondary structure (Ceroni et al. 2005).

+ Metal binding

Prediction of metal binding sites from protein sequences is useful to help esperimental determination methods and to scan for metalloproteins at the genomic scale. We have developed a predictor that discriminated between cysteines involved in metal binding vs those involved in disulfide brdiges (Passerini & Frasconi 2004). We are currently working on several extensions of this predictor.

+ Disulfide bridges

Machine learning methods for predicting the oxidation state of cysteines (Ceroni et al., 2003) and for learning to locate disulfide bridges (Vullo & Frasconi, 2004). DISULFIND is a public prediction service based on these methods.

Other applications of machine learning

+ Natural language

Prediction of first pass attachment in incremental language processing using recursive neural networks (Costa et al. 2005, Sturt et al. 2003, Costa et al. 2003). Incremental parsing (Costa et al. 2001). Kernels vs. recursive networks in language (Menchetti et al. 2005).

+ Text categorization

Multi-page text categorization (Frasconi et al. 2002).

+ Pattern recognition

E.g. document recognition (Diligenti et al 2003), fingerprint classification (Yao et al. 2003), logo recognition (Francesconi et al 1997), speech recognition (Frasconi et al 1995).

Servers & software

MetalDetector

A predictor of protein metal binding state.

DISULFIND

A predictor of cysteines bonding state and disulfide bridges.

MLNGTS

Markov Logic Networks with Grounding-Specific Weights, a mod of the Alchemy system (written by M. Lippi).

3D and 2D Decomposition Kernels

for classification of small molecules (written by A. Ceroni and F. Costa).

Weighted Decomposition Kernels

(C++ Implementation; written by F. Costa).

Recent projects

APrIL II

Applications of Probabilistic Logic Representations. STREP funded by the European Commission.

BIOPATTERN

Computational Intelligence for Biopattern analysis in Support of eHealthcare. Network of excellence funded by the European Commission.

BIOPTRAIN

Bioinformatics Optimization Training. A Marie Curie Early Stage Training Programme funded by the European Commission.

SSAMM

Tools supporting the Mobility Metropolotan Agency. Funded by the Foundation for Research and Innovation of the University of Florence.

Professional service

Journals

Artificial Intelligence Journal, associate editor

Machine Learning Journal, editorial board member

Source Code for Biology and Medicine, editorial board member

Journal of Machine Learning Research (JMLR)

Special Topic on Mining and Learning with Graphs and Relations

ACM Transactions on Internet Technology (TOIT)

Special Issue on Machine Learning for the Internet.First issue (Volume 4, Issue 2, May 2004), Second issue (Volume 4, Issue 4, November 2004).

Cognitive Systems Research

Special Issue on Integration of Symbolic and Connectionist Systems, Volume 3, Issue 2, June 2002.

IEEE Transactions on Knowledge and Data Engineering

Special Issue on Connectionist Models for Learning in Structured Domains, March/April 2001 (Vol. 13, No. 2).

IEEE Transactions on Knowledge and Data Engineering, associate editor (2000–2005)

IEEE Transactions on Neural Networks, associate editor (2000–2004)

Conferences

20th International Conference on Inductive Logic Programming 2010, Program co-chair

AAAI 2010 Special Track on Artificial Intelligence and Bioinformatics, co-chair

Neural Information Processing Systems (NIPS), Area Chair (2008,2009)

21st Int. Joint Conference on Artificial Intelligence (IJCAI) 2009, Senior Program Committee member

5th International Workshop on Mining and Learning with Graphs (MLG'07), Firenze, Aug 1-3, 2007. Program co-chair

3rd IEEE International Conference on Data Mining (ICDM'03), Vice Chair

Other past events

Artificial Intelligence and Heuristic Methods for Bioinformatics. An Advanced Studies Institute, S. Miniato, October 2001.

Foundations of connectionist-symbolic integration: representation, paradigms, and algorithms, ECAI 2000 Workshop (Berlin).

Learning Dynamical Data Structures: From Sequences to Graphs, NIPS Workshop, Breckenridge, Colorado, 1997.

Some talks

Learning with Kernels and Logical Representations, Invited keynote at the 17th Int. Conference on Inductive Logic Programming (ILP'07)

Neural Networks and Kernels for Learning Discrete Data Structures. Invited keynote at Freiburg, Leuven & Friends, 2005.

Learning Structured Data: Theory and Applications. Invited keynote at the Workshop on Advances in Machine Learning, Montreal June 2003.

Graduate Students

Current

Matteo Bertini

Marco Lippi

Marc Vincent

Past

Dr. Alessio Ceroni (Research Assistant, Imperial College London)

Dr. Sauro Menchetti

Dr. Fabrizio Costa (Postdoctoral fellow, K.U. Leuven, Belgium)

Dr. Andrea Passerini (Assistant Professor, DISI, Università di Trento)

Dr. Alessandro Vullo (Postdoctoral Fellow, UCD, Dublin)

Music

A few tracks I made can be downloaded from timet.org