Residual connections encourage iterative inference

Tutkimustuotos: Artikkeli kirjassa/konferenssijulkaisussaKonferenssiesitysvertaisarvioitu

Tutkijat

  • Stanisław Jastrz Ebski
  • Devansh Arpit
  • Nicolas Ballas
  • Vikas Verma

  • Tong Che
  • Yoshua Bengio

Organisaatiot

  • Jagiellonian University in Kraków
  • University of Montreal
  • Facebook Inc
  • CIFAR

Kuvaus

Residual networks (Resnets) have become a prominent architecture in deep learning. However, a comprehensive understanding of Resnets is still a topic of ongoing research. A recent view argues that Resnets perform iterative refinement of features. We attempt to further expose properties of this aspect. To this end, we study Resnets both analytically and empirically. We formalize the notion of iterative refinement in Resnets by showing that residual connections naturally encourage features of residual blocks to move along the negative gradient of loss as we go from one block to the next. In addition, our empirical analysis suggests that Resnets are able to perform both representation learning and iterative refinement. In general, a Resnet block tends to concentrate representation learning behavior in the first few layers while higher layers perform iterative refinement of features. Finally we observe that sharing residual layers naively leads to representation explosion and counterintuitively, overfitting, and we show that simple existing strategies can help alleviating this problem.

Yksityiskohdat

AlkuperäiskieliEnglanti
Sivumäärä1-14
TilaJulkaistu - 1 tammikuuta 2018
OKM-julkaisutyyppiEi oikeutettu
TapahtumaInternational Conference on Learning Representations - Vancouver, Kanada
Kesto: 30 huhtikuuta 20183 toukokuuta 2018
Konferenssinumero: 6
https://iclr.cc/Conferences/2018

Conference

ConferenceInternational Conference on Learning Representations
LyhennettäICLR
MaaKanada
KaupunkiVancouver
Ajanjakso30/04/201803/05/2018
www-osoite

ID: 38842838