Time Matters in Regularizing Deep Networks: Weight Decay and Data Augmentation Affect Early Learning Dynamics, Matter Little Near Convergence.
Aditya GolatkarAlessandro AchilleStefano SoattoPublished in: CoRR (2019)
Keyphrases
- prior knowledge
- data sets
- image data
- data analysis
- data collection
- network structure
- database
- learning process
- synthetic data
- connectionist networks
- hidden variables
- learning models
- human experts
- background knowledge
- learning systems
- data points
- data sources
- end users
- data structure
- raw data
- experimental data
- recurrent networks
- small number
- sensory data
- probabilistic model
- convergence rate
- original data
- learning algorithm
- knowledge discovery
- wireless networks
- xml documents
- high dimensional data
- statistical analysis
- computer systems
- metadata
- data processing
- supervised learning