ML-DL Notes

Notes for machine learning and deep learning models, design and training techniques
Mainly includes: CNN, RNN(LTSM)
ensemble techniques
activation functions

1. deep learning concept

1. LTSM,

a type of RNN, most successful RNNs are LTSM RNNs
1. use a vector as cell state to record state
2. use three gates to erase, update, output the cell state.

Example: Rakuten data challenge
Character-level tokenization
Ensembling:
Bidirectional training

2. Ensemble learning

  1. Bagging(bootstrap aggregating):
    randomly select N training sets, then vote
    example: random forest

  2. Boosting
    try to learn from a bunch of weak models, to learn from mistakes
    example: gradient boosting, gradient boosting decistion tree

  3. model stacking
    training different models first, then stack them together (with different weights)

3. Activation functions

https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6
Sigmoid, R:(0,1)
Tanh, R:(-1, 1)
ReLu, y=max(0,x) R:(0, z), f(z) = 0 if z <= 0, otherwise z = z (most used, used in almost all teh convolutional neural networks)
Leaky ReLu: solve ReLu’s range problem
Softplus, y=log(1+ex)
Softmax

2. deep learning packages:

Keras:

TensorFlow:

4. Other models to be learned

Siamese network
triplet loss neural network

Written with StackEdit.