Bernard Victorri

Transformation invariance in pattern recognition: Tangent distance and propagation

In pattern recognition, statistical modeling, or regression, the amount of data is a critical factor affecting the performance. If the amount of data and computational resources are unlimited, even trivial algorithms will converge to the optimal solution. However, in the practical case, given limited data and other resources, satisfactory performance requires sophisticated methods to regularize the problem by introducing a priori knowledge. Invariance of the output with respect to certain transformations of the input is a typical example of such a priori knowledge. We introduce the concept of tangent vectors, which compactly represent the essence of these transformation invariances, and two classes of algorithms, tangent distance and tangent propagation, which make use of these invariances to improve performance.

Transformation Invariance in Pattern Recognition-Tangent Distance and Tangent Propagation

In pattern recognition, statistical modeling, or regression, the amount of data is a critical factor affecting the performance. If the amount of data and computational resources are unlimited, even trivial algorithms will converge to the optimal solution. However, in the practical case, given limited data and other resources, satisfactory performance requires sophisticated methods to regularize the problem by introducing a priori knowledge. Invariance of the output with respect to certain transformations of the input is a typical example of such a priori knowledge. In this chapter, we introduce the concept of tangent vectors, which compactly represent the essence of these transformation invariances, and two classes of algorithms, “tangent distance” and “tangent propagation”, which make use of these invariances to improve performance.

A STATIC SIGNATURE VERIFICATION SYSTEM BASED ON A COOPERATING NEURAL NETWORKS ARCHITECTURE

We are applying neural networks to the problem of handwritten signature verification. Our system is working on checks, so we can only use the static information (the image). This static information is used in three representations: geometrical parameters, outline and image. Our system is composed of several neural networks which cooperate together during the learning and decision phases. The performances in generalization, obtained with a large-scale database of 6000 signatures from real checks on random forgeries, are False Acceptance Rate (FAR) = 2% and False Rejection Rate (FRR) = 4%.

Hierarchy in lexical organisation of natural languages

The organization of lexical units in natural languages can be compared in many respects to social organizations. Graphs of lexical relations share important specific properties with graphs of social relations (“small world” graphs). As regards hierarchical structuring, we are confronted with the same problem: words generally belong to more than one semantic class, due to the pervasive phenomenon of polysemy (words with several related meanings). Nevertheless, the existence of some form of hierarchical organization is unquestionable: some words have very wide meanings while others are more precise, and one may expect a fractal structure when dealing with the whole lexicon. We shall present here a method for bringing out this hierarchical organization by analyzing a graph of synonymy The main idea is to compute from the graph a geometrical representation in which each word is associated with a more or less extended area in a ‘semantic space. We show that regions of high connectivity on the graph correspond to high density places in the geometrical representation. Thus the lexical structure can be visualized at different scales, revealing the relevant dimensions of the semantic space at each level, from the global macrostructure to the most fine-grained local organization.

Vers une modélisation continue de la structure prosodique: le cas des proéminences syllabiques

Lobjectif de cet article est de presenter un outil developpe en vue de modeliser semi-automatiquement la structure prosodique du francais. Sur la base dun alignement en phonemes, notre systeme procede a la detection des syllabes proeminentes en prenant en consideration des criteres acoustiques basiques tels que la f0, la duree et la presence de pauses. A partir des mesures ainsi prises, le systeme attribue un degre de proeminence a chacune des syllabes identifiees comme saillante. Nous illustrons ensuite les resultats de lanalyse dextraits du corpus PROSO_FR. Plus precisement, nous comparons lanalyse prosodique de phrases que lon pourrait faire avec les regles traditionnelles de la phonologie prosodique avec lanalyse conduite par notre logiciel. Nous discutons ainsi de trois regles: la regle de dominance droite, la regle de clash accentuel et la regle des sept syllabes.

Analogy between language and biology: a functional approach

We adopt here a functional approach to the classical comparison between language and biology. We first parallel events which have a functional signification in each domain, by matching the utterance of a sentence with the release of a protein. The meaning of a protein is then defined by analogy as “the constant contribution of the biochemical material composing the protein to the effects produced by any release of the protein”. The proteome of an organism corresponds to an I-language (the idiolect of an individual), and the proteome of a species is equivalent to an E-language (a language in the common sense). Proteins and sentences are both characterized by a complex hierarchical structure, but the language property of ‘double articulation’ has no equivalent in the biological domain in this analogy, contrary to previous proposals centered on the genetic code. Besides, the same intimate relation between structure and meaning holds in both cases (syntactic structure for sentences and three-dimensional conformation for proteins). An important disanalogy comes from the combinatorial power of language which is not shared by the proteome as a whole, but it must be noted that the immune system possesses interesting properties in this respect. Regarding evolutionary aspects, the analogy still works to a certain extent. Languages and proteomes can be both considered as belonging to a general class of systems, that we call “productive self-reproductive systems”, characterized by the presence of two dynamics: a fast dynamics in an external domain where functional events occur (productive aspect), and a slow dynamics responsible for the evolution of the system itself, driven by the feed-back of events related to the reproduction process.

Nouvelles perspectives en extraction d'information

Lextraction dinformation (EI) est une technologie visant a reconnaitre dans un corpus de documents textuels un ensemble dinformations specifiques, a les extraire et a les structurer dans un format predefini. LEI a connu un essor considerable ces dix dernieres annees et devrait conduire a des applications industrielles dans un avenir proche. Apres une presentation des principes de cette technologie, cet article decrit les travaux menes dans notre groupe sur ce theme. Un systeme operationnel, ayant permis danalyser un corpus de constats daccidents y est presente. Nous montrons ensuite comment les techniques de lEI peuvent etre exploitees pour de nouvelles tâches de linformatique documentaire : encodage semantique, aide a la lecture, structuration de documents composites. Enfin nous presentons des travaux en semantique susceptibles dameliorer les performances des systemes actuels.

Artificial neural networks architecture for handwritten signature authentication

It is frequently asked to individuals to prove their identity when writing official documents. This is done to avoid the use of someone elses signature and also to avoid that someone disowns a document that he has previously acknowledged. Texts are often typed, so it is not possible to authenticate these documents from handwriting. However, it is customary to append a mark authenticating the author of the document, thus showing that he agrees with the text of the document. Nowadays this mark is generally a handwritten signature, so it is interesting to devise an automatic and reliable system for the authentication of handwritten signatures appended on the numerous documents which are produced daily. The difficulty of the signature authentication problem is linked to the high number of writers, to the diversity of signatures to store, and also to the important variations between signatures from the same writer [Sabourin 90]. The authentication problem is different from the identification problem because the latter consists in determining the writer from his signature. In the authentication case, we know the writer who is supposed to have signed, as his name is written on the document, for example a check. So it is possible to access in a database to the signatures given by the writer to be used as reference signatures. Then, the authentication process consists in comparing the signature to the reference ones in order to judge if the supposed writer is really the author of the tested signature. The signature authentication can be used in several applications ; let us now focus on the verification of checks from the French Post Office. Our goal is to detect rough forgeries, which are signatures written by someone who is not imitating a genuine signature. Those rough forgeries are the most commonly found forgeries. Systems based on dynamic information (duration, speed of the signing, ...) are able to detect good imitations. In our application however, this dynamic information is lost because the image of the check contains only static information. Without major modifications, the authentication module of our system can be used by authentication systems based on other types of data such as digital fingerprints or dynamic information about the signatures.