19 lines
680 B
Markdown
19 lines
680 B
Markdown
|
Deep Learning:
|
||
|
|
==============
|
||
|
|
Creating a model such that we don't have to hand engineer features, instead
|
||
|
|
architecting the model such that it is capable of inferring the features
|
||
|
|
on its own with large number of datasets and layers.
|
||
|
|
|
||
|
|
Input of softmax layer is called logits( classifier )
|
||
|
|
|
||
|
|
Optimization Momentum:
|
||
|
|
======================
|
||
|
|
using averaged gradients computed in previous iterations to identify how much
|
||
|
|
weight is given to the gradient descent.
|
||
|
|
|
||
|
|
Weight initialization:
|
||
|
|
======================
|
||
|
|
create a smaller network -> compute weights
|
||
|
|
use the weights and add new layer and -> compute weights
|
||
|
|
iterate and grow the network by using precomputed weights for deeper networks.
|