Speaker
Elena Agliari
(Sapienza Università di Roma)
Description
We consider a $K$-layer hetero-associative neural-network and we carry out a statistical mechanical analysis. Our findings show that these networks exhibit spontaneous information processing capabilities that go far beyond those of auto-associative counterparts. In particular, they can perform frequency modulation and, when presented with a spurious state (e.g., a symmetric mixture made of $K$ patterns), disentangle and retrieve the individual components. Furthermore, we show that this capacity is enhanced by introducing controlled perturbations—either in the training data or neuronal dynamics.
Primary author
Elena Agliari
(Sapienza Università di Roma)
