Abstract
In this paper a brief overview of Pietro Torasso’s academic career and early scientific work is provided. Particular attention is given to his interests in Artificial Intelligence topics.
The academic and scientific career
The Computer Sciences Course (“Scienze dell’Informazione”) at the University of Torino was founded in 1970 by Prof. Corrado Böhm, helped by a few assistant professors and other teachers, belonging to other faculties or universities. Notwithstanding the fact that Informatics was, at the time, almost unknown to most people, the novelty of the course attracted a group of students, curious about the innovation; among those were Pietro (alias, Piero) Torasso and Leonardo Lesmo.
Piero graduated in October 1974. Even though the main research theme at the Computer Science Department was centred, in those days, on theoretical computer science, Piero’s interest was attracted by a subject that was gaining a big momentum, namely, the study of the spoken language. His thesis, developed at the CNR Center CENS, under the supervision of Prof. Raffaele Meo, of the Polytechnic School of Torino, and Prof. Eliana Minicozzi, just arrived from the University of Naples, had the title Automatic speech synthesis: Production rules for the nasal consonants (m, n, gn) in the Italian language.
After graduation, Piero took the decision of devoting his life to research and teaching, pursuing an academic career. At that time the Ph.D. courses were not yet activated in Italy, and then, Piero obtained a ministerial bourse at the Computer Science Department for four years, during which he completed his training to become a researcher. From this point on, Piero’s academic career followed a linear evolution: first, as the responsible of the course “Elaboration of the Non-Numerical Information” (from 1978 to 1985), then, as an Associate Professor, and, then, as Full professor (from 1990) at the University of Udine, where he has been in charge of the course “Programming Techniques” at the MFN SciencesFaculty.
In the meantime, the University of Torino had opened a campus in Alessandria, where a new Faculty of Science was founded, and Piero came back, in 1991, to his original University. In the new position, Piero acts, for three years, as chair of the council for the Computer Science course, and contributes to the design and realisation of two didactic laboratories, very innovative at the time, based on a series of powerful workstations connected in a network to a set of servers.
In the end, Piero, given the opportunity, went back to the Computer Science Department in Torino, where he would work for the rest of his life. At the Department, in addition to his demanding research activity, Piero found the energy to be the department’s director for six years, and director of the Doctoral School for six more years.
The strong commitment, the honesty, and the competence spent in his didactic and scientific activities have characterised Piero in the various administrative and/or scientific responsibilities taken as a representative of the University of Torino. Among others, Piero has been a member of the Scientific Committee of “CSI-Piemonte”, President of a Committee oriented to the development of technologies for the Information society, member of the “CINI” Scientific Council, and member of the “Steering Committee for the Aerospace” in Piedmont. In every one of these activities Piero contributed to the visibility of the University of Torino and to the promotion of the relevance of Computer Science in the research and social domains.
While Piero’s academic career stands out for his dedication and the results he obtained, his scientific career is no less. In fact, Piero always distinguished himself for his coherence, his deep knowledge of the faced research subjects, and for his strong professionalism. Having approached since the beginning research themes of Artificial Intelligence, he received, in 2000, the highest European Award in that area, the ECCAI Fellowship, in recognition of his pioneering work in the field.
In addition to some awards for specific research results, Piero obtained important recognitions inside the international research community, where he was appreciated both for the seriousness in investigating complex research problems, and for the robustness and novelty of the attained results. He has been Editor-in-Chief of the journal Artificial Intelligence in Medicine, invited speaker at several conferences, and Program Chair of the AI*IA 93 Congress, and of the important international conference on Principles of Knowledge Representation and Reasoning, 1994. Moreover, he has been a member of the Executive Committee of the ECCAI, and President of the Italian Association for Artificial Intelligence (AI*IA).
Finally, we cannot ignore two relevant aspects of Piero’s activity, that he himself considered as a “mission”: on the one hand, the technological transfer toward industries and public entities, and, on the other, education. Regarding the first aspect, Piero participated to (or directed) a great number of joint projects of applied research, both at the national and the international level, thus contributing to the diffusion of the research results in various sectors of industry and society. Regarding the second one, Piero tutored a large number of master and doctoral dissertations, and helped several young researchers in the first steps of their scientific career. In his relationships with these young people, he succeeded in stimulating their curiosity for the discovery, the dedication to the tasks, and the intellectual honesty that characterised himself.
Scientific research
The research themes approached by Piero are all situated in the field of Artificial Intelligence (AI), and they span a large and diversified spectrum of topics. At the beginning we may notice a kind of exploration phase, in which different problems are faced and solved, moving then to others, more or less correlated, until a unifying and stable research line emerges, centred on diagnostic problems.
Approximate reasoning
As already mentioned in Section 1, Piero’s first research interests were oriented toward automatic speech synthesis and speech recognition. It was the period when Raj Reddy, at the CMU (USA), was foreseeing a near future of successes for the field [1]. Then, Piero joined the team working on speech recognition, directed by Prof. Renato De Mori.
From the methodological point of view, speech recognition was dominated, at that time, by the statistical approach, based on Hidden Markov Models and championed by Frederick Jelinek [2], head of the Speech Processing Group at the IBM Thomas J. Watson Research Center. Instead, De Mori was pursuing another, AI-based approach, which required reasoning mechanisms more flexible than formal logic. To this aim, with Piero, he proposed the use of Fuzzy Sets. This idea was very new in the field of speech recognition, as it appeared for the first time in Jacques Brémont’s thesis, defended at the University of Nancy the previous year [3]. De Mori and Torasso’s paper [4], was mentioned among the pioneers in the field by Didier Dubois and Henry Prade in their fundamental book [5] on Fuzzy Sets, published in 1980.
The choice of fuzzy sets to represent uncertainty of a non statistical nature was almost a must. The field had been initiated in 1965 by Lotfi Zadeh [6], Professor at U.C. Berkeley, and, after a slow start, was undergoing a very rapid expansion. The reasons mostly resided in the many successful industrial applications of the new theory of Fuzzy Logic for control, proposed by Zadeh in 1973 [7], and in the widespread adoption of the theory by the part of the Japanese scientific community.
In paper [4] the correspondence between spectrogram segments and lexicon words was described in terms of a fuzzy relation, to be exploited for emitting and verifying lexical hypotheses in a Hypothesize-and-Test approach. Context-independent and context-dependent features were extracted from the spectrogram, allowing a correspondence to be established between speech segments and their linguistic interpretation. Syllabic hypotheses were used for generating lexical ones, exploiting lexical redundancy to compensate for partial corruption of the syllabic information.
In a subsequent work [8], recognition of liquid and nasal consonants (for Italian) was performed by a pre-categorical classification, followed by speech segmentation under the control of a grammar. Syllabic nuclei and hypotheses about vowels and consonants were generated with context-dependent procedures. Such procedures were based on syntactic rules, whose associated semantics allowed each hypothesis to be evaluated using a fuzzy composite relation. Interesting enough, the rules were learned from experiments. Fuzzy production rules were also used in a real-time recognizer of isolated, spoken, Italian words [9]. The rules contained symbols that were fuzzy linguistic variables.
At the beginning of the 80’s, the Artificial Intelligence world was experiencing the rapid expansion of the Expert Systems, decision support systems based on “production rules”, mostly acquired via interviews with domain experts. Many application fields were fruitfully exploiting them, and, in particular, medical diagnosis has been one of the first domains to experiment with the novelty. Thus, in those years Piero started a cooperation with a group of physicians, under the direction of Prof. Gianni Molino, on a joint project aimed at the development of a diagnostic system for liver diseases. The project was so successful [10] that it received the HUSPI International Award for a Research on Artificial Intelligence in Medicine.
In moving to the diagnostic domain, Piero transferred the experience gained in applying fuzzy reasoning in speech. Paper [11] sets the basis for the approximate reasoning methodology to be used and improved over the years. In particular, the proposed approach relied on a classifier for assigning a class ω
c
(1 ≤ c ≤ C) to a patient x, given the values of its set of symptoms and lab tests (v1, . . . , v
j
, . . . , v
k
), associated to it. The assignment was done on the basis of a set of fuzzy decision rules. Two new aspects were introduced with respect to Piero’s previous work, namely, the use of Possibility Theory [12], and a systematic exploitation of supervised learning. The idea was to attribute patient x to the class ω
c
such that the possibility P (x ∈ ω
c
) was the largest one over the set of available classes. In order to compute P (x ∈ ω
c
), an alphabet of linguistic variablesΣ = (u1, . . . , u
k
, . . . , u
n
) had to be defined [13], and then a language L
c
over Σ for each ω
c
. Given x, P (x ∈ ω
c
) could be computed as follows:
In (1) the symbol U represented a string in language L
c
, and π (U) the degree of compatibility between U and x. Moreover, μR(U,ω
c
) represented the membership function of the relation existing between U and class ω
c
. As an example, let us consider the following simple language describing class ω
c
:
Language L
c
consists of two strings,i.e., U1 = (u1 = high and u2 = verylow) and U2 = (u1 = medium and u3 = quitehigh). Given the measurements (v1, . . . , v
j
, . . . , v
k
) on x, and the memberships of the fuzzy sets high, medium, and so on, for the linguistic variables u1, u2 and u3, it is thus possible to compute π (U1) and π (U2). On the other hand, the relationships between U
i
(i = 1, 2) and ω
c
cannot be give a priori, but have to be learned from a training set
A similar learning technique was used in characterising coronary diseases [14] in a parallel project.
The application of fuzzy set-based methodologies to different problems brought, in the end, the desire to move to the theory itself. Particular attention was devoted to the learning process, which was very innovative at the time, as (symbolic) Machine Learning was still in its infancy. In [15] a learning algorithm for automatically acquire fuzzy production rules was described. The rules, and hence the algorithm, were augmented by a mechanism which combined the evidence degrees associated to the input data to obtain the degree of evidence to be assigned to the rule’s output.
Evidence combination was a topic that attracted a lot of attention, owing to its under-constrained nature, and its criticality in approximate reasoning. A large number of approaches aimed at providing a clear semantics to the fuzzy logics had been done, and some comparison was needed in order to assess their respective merits. In [16] Piero, together with Lesmo and Saitta, contributed to the issue. In this paper they discussed some of the problems related to the representation of uncertain knowledge, and to the combination of evidence degrees in rule-based expert systems.
Some of the methods proposed in the literature were analysed, with particular attention to the Subjective Bayesian Probability (used in PROSPECTOR) and the Confirmation Theory adopted in MYCIN. The paper presented then an integrated approach based on Possibility Theory for evaluating the degree of match between the set of conditions occurring in the antecedent of a production rule and the input data; in addition, the input degree of evidence was combined with the strength of the rule itself in order to obtain the evidence of the consequent. Furthermore, an algorithm for combining evidence degrees coming from different pieces of knowledge was proposed. Finally, a new semantics for the logical operators AND and OR was introduced, as well a definition of some numerical quantifiers like AT LEAST n, AT MOST n, EXACTLY n.
Piero’s interest for evidence-based approximate reasoning in general, and for fuzzy sets in particular, lasted for a few more years [17, 18], while more sophisticated and more deep types of reasoning (causality, abduction, model-based approaches) became more and more attractive. Piero’s gradual change of focus coincided with that of his traditional coworkers, Lesmo and Saitta, the former devoting his full attention to natural language processing and the latter to machine learning.