| RUS ENG |
| Full version |
|
|
|
| SEMINARS |
|
Seminar by V. V. Kozlov, S. V. Kozyrev, A. S. Trushechkin and I. V. Volovich "Quantum mathematical physics"
|
|||
|
|
|||
|
Loss Function Landscape is Not What We Expected V. Goloschapov |
|||
|
Abstract: The loss function landscape differs from conventional expectations, explaining unexpected phenomena in neural networks. Excessively large, overtrained networks can suddenly achieve perfect generalization with 100% validation accuracy. Currently, this effect is mainly reproduced on synthetic datasets under controlled conditions. |
|||