Keras

Question 1

steps to specify NN in keras

Answer 1

import tensorflow.keras as keras

f = keras.sequential([
keras.layers.Dense(100, activation = “relu”, input_shape(2,)),
keras.layers.Dense(100, activation = “relu”),
keras.layers.Dense(1, activation = “linear”)
])

Question 2

steps to compile NN in keras

Answer 2

(assuming f if a keras.sequential that has been fully specified)

f.compile(optimizer = “adam”, loss = “mean_squared_error”)

Question 3

train a NN in batches using Keras

Answer 3

(assuming f if a keras.sequential that has been fully specified)

f.fit(X, Y, batch_size = 100, epochs = 5)

Question 4

predict using sequential nn in keras

Answer 4

(assuming f if a keras.sequential that has been fully specified and training is complete)

Y_hat = f.predict(X_new)

Question 5

limitations of sequential model in keras

Answer 5

restrictive for certain applications eg networks that have shared layers or non-standard routing

these can then be handled with the keras FUNCTIONAL APi

Question 6

non linear regression example

Answer 6

page 49 on (edited ln) printed page num is 56,

Question 7

how to combat overfitting

Answer 7

• stop training early
  -regularisation

Question 8

dropout

Answer 8

regularisation method

introduce a dropout layer, where each input gets randomly replaced during training by a zero value with fixed probability

results in smoother fit