Statements (50)
| Predicate | Object |
|---|---|
| gptkbp:instance_of |
gptkb:technique
|
| gptkbp:bfsLayer |
3
|
| gptkbp:bfsParent |
gptkb:television_channel
gptkb:Deep_Learning |
| gptkbp:applies_to |
deep learning
fully connected layers recurrent neural networks convolutional layers sequence models |
| gptkbp:benefits |
very deep networks
|
| gptkbp:can_be_used_with |
dropout
residual networks |
| gptkbp:can_lead_to |
better generalization
|
| gptkbp:can_provide |
overfitting
|
| gptkbp:developed_by |
gptkb:Christian_Szegedy
gptkb:Sergey_Ioffe |
| gptkbp:enhances |
gradient flow
layer inputs |
| gptkbp:form |
layer normalization
|
| https://www.w3.org/2000/01/rdf-schema#label |
Batch Normalization
|
| gptkbp:improves |
training speed
|
| gptkbp:includes |
scaling factor
bias term |
| gptkbp:introduced |
gptkb:2015
|
| gptkbp:is_effective_against |
vanishing gradients
|
| gptkbp:is_implemented_in |
gptkb:Graphics_Processing_Unit
gptkb:Keras gptkb:Py_Torch |
| gptkbp:is_often_associated_with |
gptkb:microprocessor
|
| gptkbp:is_often_used_in |
gptkb:generative_adversarial_networks
convolutional neural networks |
| gptkbp:is_part_of |
gptkb:microprocessor
|
| gptkbp:is_related_to |
group normalization
instance normalization |
| gptkbp:is_used_in |
natural language processing
image classification object detection transfer learning |
| gptkbp:is_utilized_in |
training phase
inference phase |
| gptkbp:key |
modern architectures
|
| gptkbp:reduces |
internal covariate shift
|
| gptkbp:requires |
computational overhead
|
| gptkbp:sensor |
batch size
|
| gptkbp:technique |
accelerating convergence
improves model stability normalizes the output of a previous layer. normalizing activations reduces sensitivity to initialization |
| gptkbp:uses |
mini-batch statistics
|