Few-Bit Backward: Quantized Gradients of Activation Functions for Memory Footprint Reduction.
Georgii S. NovikovDaniel BershatskyJulia GusakAlex ShonenkovDenis DimitrovIvan V. OseledetsPublished in: CoRR (2022)
Keyphrases
- memory footprint
- activation function
- neural network
- memory usage
- artificial neural networks
- neural architecture
- feed forward
- back propagation
- hidden layer
- basis functions
- radial basis function
- multilayer perceptron
- neural nets
- learning rate
- significant bit
- fuzzy neural network
- prediction model
- neural network model
- probabilistic model
- feature vectors
- pattern recognition
- training data