Paper Image

Daydreaming neural networks effectively store patterns

Published on:

14 May 2024

Primary Category:

Disordered Systems and Neural Networks

Paper Authors:

Ludovica Serricchio,

Dario Bocchi,

Claudio Chilin,

Raffaele Marino,

Matteo Negri,

Chiara Cammarota,

Federico Ricci-Tersenghi

Bullets

Key Details

Daydreaming training avoids destructive effects of excessive unlearning

Networks store random patterns at maximal capacity

Correlations spontaneously utilized to boost capacity further

Learns class-average attractors on MNIST data

AI generated summary

Daydreaming neural networks effectively store patterns

Researchers developed an iterative neural network training procedure called Daydreaming that continuously reinforces memories to store while removing spurious memories. On both synthetic random data and real-world MNIST images, Daydreaming networks matched or exceeded state-of-the-art storage capacity and retrieval quality, even exploiting correlations to improve performance. The networks developed high-quality attractors matching unseen examples and class averages.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up