Continual Learning (CL) represents a crucial challenge in the field of artificial intelligence, addressing the need to develop learning models capable of adapting to new tasks and data without forgetting previously acquired knowledge. This thesis provides an overview of the state-of-the-art Continual Learning techniques, examining various application scenarios and the existing strategies to mitigate catastrophic forgetting, such as replay, regularization, and parameter isolation. One area where CL plays a crucial role is in Anomaly Detection, particularly in clinical settings, where the accurate and adaptive identification of anomalies can enhance diagnoses and treatments. In this context, BMAD medical dataset is introduced to test and validate anomaly detection models in real-world scenarios. More specifically, this thesis focuses on training PatchCore on BMAD, adapting it to a Continual Learning framework. Through a critical literature review and experimentation on BMAD, this research compares the effectiveness of various CL strategies using PatchCore, with the aim of identifying the approach that ensures an optimal balance between stability and plasticity, allowing the model to progressively learn new tasks without compromising its performance on previous ones.
Il Continual Learning (CL) rappresenta una sfida cruciale nel campo dell’intelligenza artificiale, affrontando la necessità di sviluppare modelli di apprendimento in grado di adattarsi a nuovi compiti e dati senza dimenticare le conoscenze precedentemente acquisite. Questa tesi fornisce una panoramica dello stato dell’arte delle tecniche di Continual Learning, esaminando i vari scenari di applicazione e le strategie esistenti per mitigare il catastrophic forgetting, come il replay, la regolarizzazione e l’isolamento dei parametri. Un ambito in cui il CL riveste un ruolo cruciale è l’Anomaly Detection (AD), in particolare nei contesti clinici, dove l’identificazione accurata e adattiva delle anomalie può migliorare diagnosi e trattamenti. In questa direzione, si introduce BMAD, un dataset sviluppato per testare e validare modelli di anomaly detection in scenari reali. Più nello specifico, questa tesi ha come focus l’addestramento di PatchCore su BMAD, adattandolo a un framework di Continual Learning. Attraverso un’analisi critica della letteratura e una sperimentazione su BMAD, questa ricerca mette a confronto l’efficacia di varie strategie di CL utilizzando PatchCore, in modo tale da individuare quale approccio garantisca un equilibrio ottimale tra stabilità e plasticità, consentendo al modello di apprendere progressivamente nuovi task senza compromettere le prestazioni su quelli precedenti.
Rilevamento Continuo di Anomalie nel Dominio dell'Imaging Medico
BEDA, NICOLA
2024/2025
Abstract
Continual Learning (CL) represents a crucial challenge in the field of artificial intelligence, addressing the need to develop learning models capable of adapting to new tasks and data without forgetting previously acquired knowledge. This thesis provides an overview of the state-of-the-art Continual Learning techniques, examining various application scenarios and the existing strategies to mitigate catastrophic forgetting, such as replay, regularization, and parameter isolation. One area where CL plays a crucial role is in Anomaly Detection, particularly in clinical settings, where the accurate and adaptive identification of anomalies can enhance diagnoses and treatments. In this context, BMAD medical dataset is introduced to test and validate anomaly detection models in real-world scenarios. More specifically, this thesis focuses on training PatchCore on BMAD, adapting it to a Continual Learning framework. Through a critical literature review and experimentation on BMAD, this research compares the effectiveness of various CL strategies using PatchCore, with the aim of identifying the approach that ensures an optimal balance between stability and plasticity, allowing the model to progressively learn new tasks without compromising its performance on previous ones.File | Dimensione | Formato | |
---|---|---|---|
Beda_Nicola.pdf
accesso aperto
Dimensione
8.94 MB
Formato
Adobe PDF
|
8.94 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/82591