logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-02022021-230156


Tipo di tesi
Tesi di laurea magistrale
Autore
ROSASCO, ANDREA
URN
etd-02022021-230156
Titolo
Distilled Replay: Mitigating Forgetting through Dataset Distillation
Dipartimento
INFORMATICA
Corso di studi
INFORMATICA
Relatori
relatore Prof. Bacciu, Davide
relatore Dott. Carta, Antonio
relatore Dott. Cossu, Andrea
Parole chiave
  • continual learning
  • dataset distillation
  • machine learning
Data inizio appello
05/03/2021
Consultabilità
Non consultabile
Data di rilascio
05/03/2091
Riassunto
Continual Learning refers to a Machine Learning setup where data is presented to the model in a sequential fashion. The main problem faced by Continual Learning is the catastrophic forgetting of existing knowledge when acquiring new information.
We combined a replay-based Continual Learning approach with the Dataset Distillation technique, to mitigate forgetting. Replay approaches store a memory of previous patterns and interleave them with incoming data at training time. Dataset Distillation allows to compress an entire dataset into a small set of informative examples. We used the distilled patterns as replay memory and showed superior performance with respect to traditional replay on three Continual Learning benchmarks.
File