Le reti neurali sono modelli computazionali costituiti da semplici unità interconnesse, dette neuroni, che processano informazioni provenienti dall'ambiente esterno in modo da individuare relazioni complesse in base alle quali produrre opportuni segnali in uscita. Sono oggetto di studio delle scienze informatiche e sono utilizzate in varie discipline come neuroscienze, matematica, statistica, fisica o ingegneria, per risolvere problemi reali di classificazione, regressione, diagnosi, clustering, controllo, automazione, ecc. La tesi si propone di confrontare i risultati ottenuti con due diversi algoritmi neurali applicati a due situazioni reali nell'ambito dell'apprendimento supervisionato. Si definiscono quindi gli elementi che caratterizzano una rete neurale e successivamente si trattano nel dettaglio i due algoritmi di addestramento: il primo è la Back-Propagation, l'algoritmo più conosciuto, che ha la caratteristica di essere iterativo; il secondo è l'algoritmo Extreme Learning Machine, più recente, che utilizza tecniche di pseudo-inversione matriciale. Di quest'ultimo si studiano anche tecniche di regolarizzazione non presenti allo stato dell'arte, per migliorarne la stabilità. Questi algoritmi di addestramento vengono utilizzati per risolvere due differenti problemi reali. La prima applicazione studiata è parte del progetto finanziato dalla Comunità Europea HoliDes ("Holistic Human Factors and System Design of Adaptive Cooperative Human-Machine Systems"), di cui il Dipartimento di Informatica dell'Università di Torino è uno dei 31 partner coinvolti. L'obiettivo è di diagnosticare lo stato di eventuale distrazione visiva di un guidatore, dovuta a disturbi esterni, al fine di inserire dispositivi di controllo in tempo reale nelle autovetture per supportare la guida. Il secondo studio è stato svolto in collaborazione con il Politecnico di Torino, Dipartimento di Ingegneria Strutturale , Edile e Geotecnica; l'obiettivo è riconoscere e classificare il livello di danno negli edifici esistenti in caso di evento sismico, secondo la normativa sismica vigente, al fine di prevedere la vulnerabilità degli stessi e indirizzare gli interventi di adeguamento. Per entrambe le applicazioni sono effettuati numerosi test volti a ottimizzare gli algoritmi studiati e sono confrontati i risultati sia in termini di accuratezza che di tempo computazionale.
Neural networks are computational models made up of an interconnected group of simple units, called, neurons, which processes information coming from the external environment to identify complex relationships to provide opportune output signals. They are the research field of computer science and they are used in various disciplines such as neuroscience, mathematics, statistics, physics or engineering, to solve real problems of classification, regression, diagnosis, clustering, control, automation, etc. The aim of this thesis is to compare the results obtained by two different algorithms applied to two real situations through the supervised learning. So, we defined the elements characterizing a neural network and then we deal in detail the two training algorithms: the first is the Back-Propagation, the most popular algorithm, which is an iterative method; the second is the Extreme Learning Machine algorithm, more recently, that uses matrix pseudo-inversion techniques. Of the latter we also studied regularization methods, not present at the actual state of the art, to improve its stability. Finally These training algorithms are used to solve two different real problems. The first application that we deepen is part of the project financed by the European Community HoliDes ("Holistic Human Factors and System Design of Adaptive Cooperative Human-Machine Systems ") and the Department of Computer Science of the University of Turin is one of 31 partners involved. The objective is to predict the state of possible visual drivers distraction, due to external disturbances, in order to insert real time control devices in the cars, to support the drivers. The second study is carried out with the Politecnico di Torino, Department of Structural, Construction and Geotechnical Engineering; the purpose is to recognize and to classify the expected level of structural damage on existing buildings in consequence to a seismic event , in according to current seismic norms, to know in advance the vulnerability of the buildings and to address interventions of adjustment to wide staircase. For both applications numerous tests to optimize the studied algorithms are carried out and the results are compared both in terms of accuracy and computational time.
TRAINING NEURALE PER APPLICAZIONI NEL MONDO REALE ATTRAVERSO PSEUDOINVERSIONE: UN APPROCCIO COMPARATIVO
DI MARTINO, MARTA
2013/2014
Abstract
Neural networks are computational models made up of an interconnected group of simple units, called, neurons, which processes information coming from the external environment to identify complex relationships to provide opportune output signals. They are the research field of computer science and they are used in various disciplines such as neuroscience, mathematics, statistics, physics or engineering, to solve real problems of classification, regression, diagnosis, clustering, control, automation, etc. The aim of this thesis is to compare the results obtained by two different algorithms applied to two real situations through the supervised learning. So, we defined the elements characterizing a neural network and then we deal in detail the two training algorithms: the first is the Back-Propagation, the most popular algorithm, which is an iterative method; the second is the Extreme Learning Machine algorithm, more recently, that uses matrix pseudo-inversion techniques. Of the latter we also studied regularization methods, not present at the actual state of the art, to improve its stability. Finally These training algorithms are used to solve two different real problems. The first application that we deepen is part of the project financed by the European Community HoliDes ("Holistic Human Factors and System Design of Adaptive Cooperative Human-Machine Systems ") and the Department of Computer Science of the University of Turin is one of 31 partners involved. The objective is to predict the state of possible visual drivers distraction, due to external disturbances, in order to insert real time control devices in the cars, to support the drivers. The second study is carried out with the Politecnico di Torino, Department of Structural, Construction and Geotechnical Engineering; the purpose is to recognize and to classify the expected level of structural damage on existing buildings in consequence to a seismic event , in according to current seismic norms, to know in advance the vulnerability of the buildings and to address interventions of adjustment to wide staircase. For both applications numerous tests to optimize the studied algorithms are carried out and the results are compared both in terms of accuracy and computational time.| File | Dimensione | Formato | |
|---|---|---|---|
|
700623_neural_training_pseudoinversion.pdf
non disponibili
Tipologia:
Altro materiale allegato
Dimensione
9.31 MB
Formato
Adobe PDF
|
9.31 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14240/43977