guest ::
| Home > Publications database > Optimal signal propagation in ResNets through residual scaling > print |
| Home > Publications database > Optimal signal propagation in ResNets through residual scaling > print |
| 001 | 1010660 | ||
| 005 | 20240313103122.0 | ||
| 024 | 7 | _ | |a 10.48550/ARXIV.2305.07715 |2 doi |
| 024 | 7 | _ | |a 10.34734/FZJ-2023-03175 |2 datacite_doi |
| 037 | _ | _ | |a FZJ-2023-03175 |
| 100 | 1 | _ | |a Fischer, Kirsten |0 P:(DE-Juel1)180150 |b 0 |e Corresponding author |
| 245 | _ | _ | |a Optimal signal propagation in ResNets through residual scaling |
| 260 | _ | _ | |c 2023 |b arXiv |
| 336 | 7 | _ | |a Preprint |b preprint |m preprint |0 PUB:(DE-HGF)25 |s 1695103673_21286 |2 PUB:(DE-HGF) |
| 336 | 7 | _ | |a WORKING_PAPER |2 ORCID |
| 336 | 7 | _ | |a Electronic Article |0 28 |2 EndNote |
| 336 | 7 | _ | |a preprint |2 DRIVER |
| 336 | 7 | _ | |a ARTICLE |2 BibTeX |
| 336 | 7 | _ | |a Output Types/Working Paper |2 DataCite |
| 520 | _ | _ | |a Residual networks (ResNets) have significantly better trainability and thus performance than feed-forward networks at large depth. Introducing skip connections facilitates signal propagation to deeper layers. In addition, previous works found that adding a scaling parameter for the residual branch further improves generalization performance. While they empirically identified a particularly beneficial range of values for this scaling parameter, the associated performance improvement and its universality across network hyperparameters yet need to be understood. For feed-forward networks (FFNets), finite-size theories have led to important insights with regard to signal propagation and hyperparameter tuning. We here derive a systematic finite-size theory for ResNets to study signal propagation and its dependence on the scaling for the residual branch. We derive analytical expressions for the response function, a measure for the network's sensitivity to inputs, and show that for deep networks the empirically found values for the scaling parameter lie within the range of maximal sensitivity. Furthermore, we obtain an analytical expression for the optimal scaling parameter that depends only weakly on other network hyperparameters, such as the weight variance, thereby explaining its universality across hyperparameters. Overall, this work provides a framework for theory-guided optimal scaling in ResNets and, more generally, provides the theoretical framework to study ResNets at finite widths. |
| 536 | _ | _ | |a 5232 - Computational Principles (POF4-523) |0 G:(DE-HGF)POF4-5232 |c POF4-523 |f POF IV |x 0 |
| 536 | _ | _ | |a 5234 - Emerging NC Architectures (POF4-523) |0 G:(DE-HGF)POF4-5234 |c POF4-523 |f POF IV |x 1 |
| 536 | _ | _ | |a RenormalizedFlows - Transparent Deep Learning with Renormalized Flows (BMBF-01IS19077A) |0 G:(DE-Juel-1)BMBF-01IS19077A |c BMBF-01IS19077A |x 2 |
| 536 | _ | _ | |a MSNN - Theory of multi-scale neuronal networks (HGF-SMHB-2014-2018) |0 G:(DE-Juel1)HGF-SMHB-2014-2018 |c HGF-SMHB-2014-2018 |f MSNN |x 3 |
| 536 | _ | _ | |a ACA - Advanced Computing Architectures (SO-092) |0 G:(DE-HGF)SO-092 |c SO-092 |x 4 |
| 536 | _ | _ | |a neuroIC002 - Recurrence and stochasticity for neuro-inspired computation (EXS-SF-neuroIC002) |0 G:(DE-82)EXS-SF-neuroIC002 |c EXS-SF-neuroIC002 |x 5 |
| 536 | _ | _ | |a GRK 2416 - GRK 2416: MultiSenses-MultiScales: Neue Ansätze zur Aufklärung neuronaler multisensorischer Integration (368482240) |0 G:(GEPRIS)368482240 |c 368482240 |x 6 |
| 588 | _ | _ | |a Dataset connected to DataCite |
| 650 | _ | 7 | |a Disordered Systems and Neural Networks (cond-mat.dis-nn) |2 Other |
| 650 | _ | 7 | |a Machine Learning (cs.LG) |2 Other |
| 650 | _ | 7 | |a Machine Learning (stat.ML) |2 Other |
| 650 | _ | 7 | |a FOS: Physical sciences |2 Other |
| 650 | _ | 7 | |a FOS: Computer and information sciences |2 Other |
| 700 | 1 | _ | |a Dahmen, David |0 P:(DE-Juel1)156459 |b 1 |u fzj |
| 700 | 1 | _ | |a Helias, Moritz |0 P:(DE-Juel1)144806 |b 2 |u fzj |
| 773 | _ | _ | |a 10.48550/ARXIV.2305.07715 |
| 856 | 4 | _ | |u https://arxiv.org/abs/2305.07715 |
| 856 | 4 | _ | |u https://juser.fz-juelich.de/record/1010660/files/2305.07715.pdf |y OpenAccess |
| 909 | C | O | |o oai:juser.fz-juelich.de:1010660 |p openaire |p open_access |p VDB |p driver |p dnbdelivery |
| 910 | 1 | _ | |a Forschungszentrum Jülich |0 I:(DE-588b)5008462-8 |k FZJ |b 0 |6 P:(DE-Juel1)180150 |
| 910 | 1 | _ | |a Forschungszentrum Jülich |0 I:(DE-588b)5008462-8 |k FZJ |b 1 |6 P:(DE-Juel1)156459 |
| 910 | 1 | _ | |a Forschungszentrum Jülich |0 I:(DE-588b)5008462-8 |k FZJ |b 2 |6 P:(DE-Juel1)144806 |
| 913 | 1 | _ | |a DE-HGF |b Key Technologies |l Natural, Artificial and Cognitive Information Processing |1 G:(DE-HGF)POF4-520 |0 G:(DE-HGF)POF4-523 |3 G:(DE-HGF)POF4 |2 G:(DE-HGF)POF4-500 |4 G:(DE-HGF)POF |v Neuromorphic Computing and Network Dynamics |9 G:(DE-HGF)POF4-5232 |x 0 |
| 913 | 1 | _ | |a DE-HGF |b Key Technologies |l Natural, Artificial and Cognitive Information Processing |1 G:(DE-HGF)POF4-520 |0 G:(DE-HGF)POF4-523 |3 G:(DE-HGF)POF4 |2 G:(DE-HGF)POF4-500 |4 G:(DE-HGF)POF |v Neuromorphic Computing and Network Dynamics |9 G:(DE-HGF)POF4-5234 |x 1 |
| 914 | 1 | _ | |y 2023 |
| 915 | _ | _ | |a OpenAccess |0 StatID:(DE-HGF)0510 |2 StatID |
| 920 | _ | _ | |l yes |
| 920 | 1 | _ | |0 I:(DE-Juel1)INM-6-20090406 |k INM-6 |l Computational and Systems Neuroscience |x 0 |
| 920 | 1 | _ | |0 I:(DE-Juel1)IAS-6-20130828 |k IAS-6 |l Theoretical Neuroscience |x 1 |
| 920 | 1 | _ | |0 I:(DE-Juel1)INM-10-20170113 |k INM-10 |l Jara-Institut Brain structure-function relationships |x 2 |
| 980 | 1 | _ | |a FullTexts |
| 980 | _ | _ | |a preprint |
| 980 | _ | _ | |a VDB |
| 980 | _ | _ | |a UNRESTRICTED |
| 980 | _ | _ | |a I:(DE-Juel1)INM-6-20090406 |
| 980 | _ | _ | |a I:(DE-Juel1)IAS-6-20130828 |
| 980 | _ | _ | |a I:(DE-Juel1)INM-10-20170113 |
| 981 | _ | _ | |a I:(DE-Juel1)IAS-6-20130828 |
| Library | Collection | CLSMajor | CLSMinor | Language | Author |
|---|