I was testing some network architectures in Keras for classifying the MNIST dataset. I have implemented one that is similar to the LeNet.
I have seen that in the examples that I have found on the internet, there is a step of data normalization. For example:
X_train /= 255
I have performed a test without this normalization and I have seen that the performance (accuracy) of the network has decreased (keeping the same number of epochs). Why has this happened?
If I increase the number of epochs, the accuracy can reach the same level reached by the model trained with normalization?
So, the normalization affects the accuracy, or only the training speed?
The complete source code of my training script is below:
from keras.models import Sequential
from keras.layers.convolutional import Conv2D
from keras.layers.convolutional import MaxPooling2D
from keras.layers.core import Activation
from keras.layers.core import Flatten
from keras.layers.core import Dense
from keras.datasets import mnist
from keras.utils import np_utils
from keras.optimizers import SGD, RMSprop, Adam
import numpy as np
import matplotlib.pyplot as plt
from keras import backend as k
def build(input_shape, classes):
model = Sequential()
model.add(Conv2D(20, kernel_size=5, padding="same",activation='relu',input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
model.add(Conv2D(50, kernel_size=5, padding="same", activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
model.add(Flatten())
model.add(Dense(500))
model.add(Activation("relu"))
model.add(Dense(classes))
model.add(Activation("softmax"))
return model
NB_EPOCH = 4 # number of epochs
BATCH_SIZE = 128 # size of the batch
VERBOSE = 1 # set the training phase as verbose
OPTIMIZER = Adam() # optimizer
VALIDATION_SPLIT=0.2 # percentage of the training data used for
evaluating the loss function
IMG_ROWS, IMG_COLS = 28, 28 # input image dimensions
NB_CLASSES = 10 # number of outputs = number of digits
INPUT_SHAPE = (1, IMG_ROWS, IMG_COLS) # shape of the input
(X_train, y_train), (X_test, y_test) = mnist.load_data()
k.set_image_dim_ordering("th")
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
X_train = X_train[:, np.newaxis, :, :]
X_test = X_test[:, np.newaxis, :, :]
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')
y_train = np_utils.to_categorical(y_train, NB_CLASSES)
y_test = np_utils.to_categorical(y_test, NB_CLASSES)
model = build(input_shape=INPUT_SHAPE, classes=NB_CLASSES)
model.compile(loss="categorical_crossentropy",
optimizer=OPTIMIZER,metrics=["accuracy"])
history = model.fit(X_train, y_train, batch_size=BATCH_SIZE, epochs=NB_EPOCH, verbose=VERBOSE, validation_split=VALIDATION_SPLIT)
model.save("model2")
score = model.evaluate(X_test, y_test, verbose=VERBOSE)
print('Test accuracy:', score[1])
Normalization is a generic concept not limited only to deep learning or to Keras.
Why to normalize?
Let me take a simple logistic regression example which will be easy to understand and to explain normalization.
Assume we are trying to predict if a customer should be given loan or not. Among many available independent variables lets just consider Age and Income.
Let the equation be of the form:
Y = weight_1 * (Age) + weight_2 * (Income) + some_constant
Just for sake of explanation let Age be usually in range of [0,120] and let us assume Income in range of [10000, 100000]. The scale of Age and Income are very different. If you consider them as is then weights weight_1 and weight_2 may be assigned biased weights. weight_2 might bring more importance to Income as a feature than to what weight_1 brings importance to Age. To scale them to a common level, we can normalize them. For example, we can bring all the ages in range of [0,1] and all incomes in range of [0,1]. Now we can say that Age and Income are given equal importance as a feature.
Does Normalization always increase the accuracy?
Apparently, No. It is not necessary that normalization always increases accuracy. It may or might not, you never really know until you implement. Again it depends on at which stage in you training you apply normalization, on whether you apply normalization after every activation, etc.
As the range of the values of the features gets narrowed down to a particular range because of normalization, its easy to perform computations over a smaller range of values. So, usually the model gets trained a bit faster.
Regarding the number of epochs, accuracy usually increases with number of epochs provided that your model doesn't start over-fitting.
A very good explanation for Normalization/Standardization and related terms is here.
In a nutshell, normalization reduces the complexity of the problem your network is trying to solve. This can potentially increase the accuracy of your model and speed up the training. You bring the data on the same scale and reduce variance. None of the weights in the network are wasted on doing a normalization for you, meaning that they can be used more efficiently to solve the actual task at hand.
As #Shridhar R Kulkarni says, normalization is a general concept and doesn’t only apply to keras.
It’s often applied as part of data preparation for ML learning models to change numeric values in the dataset to fit a standard scale without distorting the differences in their ranges. As such, normalization enhances the cohesion of entity types within a model by reducing the probability of inconsistent data.
However, not every other dataset and use case requires normalization, it’s primarily necessary when features have different ranges. You may use when;
You want to improve your model’s convergence efficiency and make
optimization feasible
When you want to make training less sensitive to scale features, you can better
solve coefficients.
Want to improve analysis from multiple models.
Normalization is not recommended when;
-Using decision tree models or ensembles based on them
-Your data is not normally distributed- you may have to use other data pre-
processing techniques
-If your dataset comprises already scaled variables
In some cases, normalization can improve performance. However, it is not always necessary.
The critical thing is to understand your dataset and scenario first, then you’ll know whether you need it or not. Sometimes, you can experiment to see if it gives you good performance or not.
Check out deepchecks and see how to deal with important data-related checks you come across in ML.
For example, to check duplicated data in your set, you can use the following code detailed code
from deepchecks.checks.integrity.data_duplicates import DataDuplicates
from deepchecks.base import Dataset, Suite
from datetime import datetime
import pandas as pd
I think there are some issue with the convergence of the optimizer function too. Here i show a simple linear regression. Three examples:
First with an array with small values and it works as expected.
Second an array with bigger values and the loss function explodes toward infinity, suggesting the need to normalize. And at the end in model 3 the same array as case two but it has been normalized and we get convergence.
github colab enabled ipython notebook
I've use the MSE optimizer function i don't know if other optimizers suffer the same issues.
Related
I am a bit confusing with comparing best GridSearchCV model and baseline.
For example, we have classification problem.
As a baseline, we'll fit a model with default settings (let it be logistic regression):
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
baseline = LogisticRegression()
baseline.fit(X_train, y_train)
pred = baseline.predict(X_train)
print(accuracy_score(y_train, pred))
So, the baseline gives us accuracy using the whole train sample.
Next, GridSearchCV:
from sklearn.model_selection import cross_val_score, GridSearchCV, StratifiedKFold
X_val, X_test_val,y_val,y_test_val = train_test_split(X_train, y_train, test_size=0.3, random_state=42)
cv = StratifiedKFold(n_splits=5, random_state=0, shuffle=True)
parameters = [ ... ]
best_model = GridSearchCV(LogisticRegression(parameters,scoring='accuracy' ,cv=cv))
best_model.fit(X_val, y_val)
print(best_model.best_score_)
Here, we have accuracy based on validation sample.
My questions are:
Are those accuracy scores comparable? Generally, is it fair to compare GridSearchCV and model without any cross validation?
For the baseline, isn't it better to use Validation sample too (instead of the whole Train sample)?
No, they aren't comparable.
Your baseline model used X_train to fit the model. Then you're using the fitted model to score the X_train sample. This is like cheating because the model is going to already perform the best since you're evaluating it based on data that it has already seen.
The grid searched model is at a disadvantage because:
It's working with less data since you have split the X_train sample.
Compound that with the fact that it's getting trained with even less data due to the 5 folds (it's training with only 4/5 of X_val per fold).
So your score for the grid search is going to be worse than your baseline.
Now you might ask, "so what's the point of best_model.best_score_? Well, that score is used to compare all the models used when searching for the optimal hyperparameters in your search space, but in no way should be used to compare against a model that was trained outside of the grid search context.
So how should one go about conducting a fair comparison?
Split your training data for both models.
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)
Fit your models using X_train.
# fit baseline
baseline.fit(X_train, y_train)
# fit using grid search
best_model.fit(X_train, y_train)
Evaluate models against X_test.
# baseline
baseline_pred = baseline.predict(X_test)
print(accuracy_score(y_test, baseline_pred))
# grid search
grid_pred = best_model.predict(X_test)
print(accuracy_score(y_test, grid_pred))
I am trying to predict the hygrothermal response of a wall, given the interior and exterior climate. Based on literature research, I believe this should be possible with RNN but I have not been able to get good accuracy.
The dataset has 12 input features (time-series of exterior and interior climate data) and 10 output features (time-series of hygrothermal response), both containing hourly values for 10 years. This data was created with hygrothermal simulation software, there is no missing data.
Dataset features:
Dataset targets:
Unlike most time-series prediction problems, I want to predict the response for the full length of the input features time-series at each time-step, rather than the subsequent values of a time-series (eg financial time-series prediction). I have not been able to find similar prediction problems (in similar or other fields), so if you know of one, references are very welcome.
I think this should be possible with RNN, so I am currently using LSTM from Keras. Before training, I preprocess my data the following way:
Discard first year of data, as the first time steps of the hygrothermal response of the wall is influenced by the initial temperature and relative humidity.
Split into training and testing set. Training set contains the first 8 years of data, the test set contains the remaining 2 years.
Normalise training set (zero mean, unit variance) using StandardScaler from Sklearn. Normalise test set analogously using mean an variance from training set.
This results in: X_train.shape = (1, 61320, 12), y_train.shape = (1, 61320, 10), X_test.shape = (1, 17520, 12), y_test.shape = (1, 17520, 10)
As these are long time-series, I use stateful LSTM and cut the time-series as explained here, using the stateful_cut() function. I only have 1 sample, so batch_size is 1. For T_after_cut I have tried 24 and 120 (24*5); 24 appears to give better results. This results in X_train.shape = (2555, 24, 12), y_train.shape = (2555, 24, 10), X_test.shape = (730, 24, 12), y_test.shape = (730, 24, 10).
Next, I build and train the LSTM model as follows:
model = Sequential()
model.add(LSTM(128,
batch_input_shape=(batch_size,T_after_cut,features),
return_sequences=True,
stateful=True,
))
model.addTimeDistributed(Dense(targets)))
model.compile(loss='mean_squared_error', optimizer=Adam())
model.fit(X_train, y_train, epochs=100, batch_size=batch=batch_size, verbose=2, shuffle=False)
Unfortunately, I don't get accurate prediction results; not even for the training set, thus the model has high bias.
The prediction results of the LSTM model for all targets
How can I improve my model? I have already tried the following:
Not discarding the first year of the dataset -> no significant difference
Differentiating the input features time-series (subtract previous value from current value) -> slightly worse results
Up to four stacked LSTM layers, all with the same hyperparameters -> no significant difference in results but longer training time
Dropout layer after LSTM layer (though this is usually used to reduce variance and my model has high bias) -> slightly better results, but difference might not be statistically significant
Am I doing something wrong with the stateful LSTM? Do I need to try different RNN models? Should I preprocess the data differently?
Furthermore, training is very slow: about 4 hours for the model above. Hence I am reluctant to do an extensive hyperparameter gridsearch...
In the end, I managed to solve this the following way:
Using more samples to train instead of only 1 (I used 18 samples to train and 6 to test)
Keep the first year of data, as the output time-series for all samples have the same 'starting point' and the model needs this information to learn
Standardise both input and output features (zero mean, unit variance). I found this improved prediction accuracy and training speed
Use stateful LSTM as described here, but add reset states after epoch (see below for code). I used batch_size = 6 and T_after_cut = 1460. If T_after_cut is longer, training is slower; if T_after_cut is shorter, accuracy decreases slightly. If more samples are available, I think using a larger batch_size will be faster.
use CuDNNLSTM instead of LSTM, this speed up the training time x4!
I found that more units resulted in higher accuracy and faster convergence (shorter training time). Also I found that the GRU is as accurate as the LSTM tough converged faster for the same number of units.
Monitor validation loss during training and use early stopping
The LSTM model is build and trained as follows:
def define_reset_states_batch(nb_cuts):
class ResetStatesCallback(Callback):
def __init__(self):
self.counter = 0
def on_batch_begin(self, batch, logs={}):
# reset states when nb_cuts batches are completed
if self.counter % nb_cuts == 0:
self.model.reset_states()
self.counter += 1
def on_epoch_end(self, epoch, logs={}):
# reset states after each epoch
self.model.reset_states()
return(ResetStatesCallback)
model = Sequential()
model.add(layers.CuDNNLSTM(256, batch_input_shape=(batch_size,T_after_cut ,features),
return_sequences=True,
stateful=True))
model.add(layers.TimeDistributed(layers.Dense(targets, activation='linear')))
optimizer = RMSprop(lr=0.002)
model.compile(loss='mean_squared_error', optimizer=optimizer)
earlyStopping = EarlyStopping(monitor='val_loss', min_delta=0.005, patience=15, verbose=1, mode='auto')
ResetStatesCallback = define_reset_states_batch(nb_cuts)
model.fit(X_dev, y_dev, epochs=n_epochs, batch_size=n_batch, verbose=1, shuffle=False, validation_data=(X_eval,y_eval), callbacks=[ResetStatesCallback(), earlyStopping])
This gave me very statisfying accuracy (R2 over 0.98):
This figure shows the temperature (left) and relative humidity (right) in the wall over 2 years (data not used in training), prediction in red and true output in black. The residuals show that the error is very small and that the LSTM learns to capture the long-term dependencies to predict the relative humidity.
I am trying to build a model to predict house prices.
I have some features X (no. of bathrooms , etc.) and target Y (ranging around $300,000 to $800,000)
I have used sklearn's Standard Scaler to standardize Y before fitting it to the model.
Here is my Keras model:
def build_model():
model = Sequential()
model.add(Dense(36, input_dim=36, activation='relu'))
model.add(Dense(18, input_dim=36, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='mse', optimizer='sgd', metrics=['mae','mse'])
return model
I am having trouble trying to interpret the results -- what does a MSE of 0.617454319755 mean?
Do I have to inverse transform this number, and square root the results, getting an error rate of 741.55 in dollars?
math.sqrt(sc.inverse_transform([mse]))
I apologise for sounding silly as I am starting out!
I apologise for sounding silly as I am starting out!
Do not; this is a subtle issue of great importance, which is usually (and regrettably) omitted in tutorials and introductory expositions.
Unfortunately, it is not as simple as taking the square root of the inverse-transformed MSE, but it is not that complicated either; essentially what you have to do is:
Transform back your predictions to the initial scale of the original data
Get the MSE between these invert-transformed predictions and the original data
Take the square root of the result
in order to get a performance indicator of your model that will be meaningful in the business context of your problem (e.g. US dollars here).
Let's see a quick example with toy data, omitting the model itself (which is irrelevant here, and in fact can be any regression model - not only a Keras one):
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import mean_squared_error
import numpy as np
# toy data
X = np.array([[1,2], [3,4], [5,6], [7,8], [9,10]])
Y = np.array([3, 4, 5, 6, 7])
# feature scaling
sc_X = StandardScaler()
X_train = sc_X.fit_transform(X)
# outcome scaling:
sc_Y = StandardScaler()
Y_train = sc_Y.fit_transform(Y.reshape(-1, 1))
Y_train
# array([[-1.41421356],
# [-0.70710678],
# [ 0. ],
# [ 0.70710678],
# [ 1.41421356]])
Now, let's say that we fit our Keras model (not shown here) using the scaled sets X_train and Y_train, and get predictions on the training set:
prediction = model.predict(X_train) # scaled inputs here
print(prediction)
# [-1.4687586 -0.6596055 0.14954728 0.95870024 1.001172 ]
The MSE reported by Keras is actually the scaled MSE, i.e.:
MSE_scaled = mean_squared_error(Y_train, prediction)
MSE_scaled
# 0.052299712818541934
while the 3 steps I have described above are simply:
MSE = mean_squared_error(Y, sc_Y.inverse_transform(prediction)) # first 2 steps, combined
MSE
# 0.10459946572909758
np.sqrt(MSE) # 3rd step
# 0.323418406602187
So, in our case, if our initial Y were US dollars, the actual error in the same units (dollars) would be 0.32 (dollars).
Notice how the naive approach of inverse-transforming the scaled MSE would give a very different (and incorrect) result:
np.sqrt(sc_Y.inverse_transform([MSE_scaled]))
# array([2.25254588])
MSE is mean square error, here is the formula.
Basically it is a mean of square of different of expected output and prediction. Making square root of this will not give you the difference between error and output. This is useful for training.
Currently you have build a model.
If you want to train the model use these function.
mode.fit(x=input_x_array, y=input_y_array, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)
If you want to do prediction of the output you should use following code.
prediction = model.predict(np.array(input_x_array))
print(prediction)
You can find more details here.
https://keras.io/models/about-keras-models/
https://keras.io/models/sequential/
I would like to calculate NN model certainty/confidence (see What my deep model doesn't know) - when NN tells me an image represents "8", I would like to know how certain it is. Is my model 99% certain it is "8" or is it 51% it is "8", but it could also be "6"? Some digits are quite ambiguous and I would like to know for which images the model is just "flipping a coin".
I have found some theoretical writings about this but I have trouble putting this in code. If I understand correctly, I should evaluate a testing image multiple times while "killing off" different neurons (using dropout) and then...?
Working on MNIST dataset, I am running the following model:
from keras.models import Sequential
from keras.layers import Dense, Activation, Conv2D, Flatten, Dropout
model = Sequential()
model.add(Conv2D(128, kernel_size=(7, 7),
activation='relu',
input_shape=(28, 28, 1,)))
model.add(Dropout(0.20))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Dropout(0.20))
model.add(Flatten())
model.add(Dense(units=64, activation='relu'))
model.add(Dropout(0.25))
model.add(Dense(units=10, activation='softmax'))
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
model.fit(train_data, train_labels, batch_size=100, epochs=30, validation_data=(test_data, test_labels,))
How should I predict with this model so that I get its certainty about predictions too? I would appreciate some practical examples (preferably in Keras, but any will do).
To clarify, I am looking for an example of how to get certainty using the method outlined by Yurin Gal (or an explanation of why some other method yields better results).
If you want to implement dropout approach to measure uncertainty you should do the following:
Implement function which applies dropout also during the test time:
import keras.backend as K
f = K.function([model.layers[0].input, K.learning_phase()],
[model.layers[-1].output])
Use this function as uncertainty predictor e.g. in a following manner:
def predict_with_uncertainty(f, x, n_iter=10):
result = numpy.zeros((n_iter,) + x.shape)
for iter in range(n_iter):
result[iter] = f(x, 1)
prediction = result.mean(axis=0)
uncertainty = result.var(axis=0)
return prediction, uncertainty
Of course you may use any different function to compute uncertainty.
Made a few changes to the top voted answer. Now it works for me.
It's a way to estimate model uncertainty. For other source of uncertainty, I found https://eng.uber.com/neural-networks-uncertainty-estimation/ helpful.
f = K.function([model.layers[0].input, K.learning_phase()],
[model.layers[-1].output])
def predict_with_uncertainty(f, x, n_iter=10):
result = []
for i in range(n_iter):
result.append(f([x, 1]))
result = np.array(result)
prediction = result.mean(axis=0)
uncertainty = result.var(axis=0)
return prediction, uncertainty
Your model uses a softmax activation, so the simplest way to obtain some kind of uncertainty measure is to look at the output softmax probabilities:
probs = model.predict(some input data)[0]
The probs array will then be a 10-element vector of numbers in the [0, 1] range that sum to 1.0, so they can be interpreted as probabilities. For example the probability for digit 7 is just probs[7].
Then with this information you can do some post-processing, typically the predicted class is the one with highest probability, but you can also look at the class with second highest probability, etc.
A simpler way is to set training=True on any dropout layers you want to run during inference as well (essentially tells the layer to operate as if it's always in training mode - so it is always present for both training and inference).
import keras
inputs = keras.Input(shape=(10,))
x = keras.layers.Dense(3)(inputs)
outputs = keras.layers.Dropout(0.5)(x, training=True)
model = keras.Model(inputs, outputs)
Code above is from this issue.
The problem is more about the training algorithm for DNN rather than the software keras.
As far as I know, deep neural network works due to the improvement in training algorithm. From the 1980s, the BP algorithm has been used to train neural network but will result in over-fitting problem when the network is deep. About 10 years ago, Hinton improved the algorithm by first pre-traning the network using unlabeled data and then using BP algorithm. The pre-traning plays an important role to avoid over-fitting.
However, as I begin to try Keras, the example (in the below) of mnist DNN using SGD algorithm without any mention about the pre-training process leads to a very high prediction accuracy. So, I begin to wonder where has the pre-training gone? Wheter I misundertood the deep learning training algorithm (I think the classical BP is almost the same as SGD)? Or a new traning technique has replaced the pre-traning process?
Very grateful for your help!
'''Trains a simple deep NN on the MNIST dataset.
Gets to 98.40% test accuracy after 20 epochs
(there is *a lot* of margin for parameter tuning).
2 seconds per epoch on a K520 GPU.
'''
from __future__ import print_function
import numpy as np
np.random.seed(1337) # for reproducibility
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.optimizers import SGD, Adam, RMSprop
from keras.utils import np_utils
batch_size = 128
nb_classes = 10
nb_epoch = 20
# the data, shuffled and split between train and test sets
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train = X_train.reshape(60000, 784)
X_test = X_test.reshape(10000, 784)
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test = np_utils.to_categorical(y_test, nb_classes)
model = Sequential()
model.add(Dense(512, input_shape=(784,)))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(10))
model.add(Activation('softmax'))
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(),
metrics=['accuracy'])
history = model.fit(X_train, Y_train,
batch_size=batch_size, nb_epoch=nb_epoch,
verbose=1, validation_data=(X_test, Y_test))
score = model.evaluate(X_test, Y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
You are wrong.
Past vs. Today
The difference between Neural Networks in the past and the ones today is not about the training algorithm. Every DNN is trained with Backpropagation based on some SGD-based algorithm, exactly like in the past. (There are some new algorithms trying to reduce parameter-tuning with adaptive learning-rates like Adam, RMSprop and co.; but plain SGD is still the most common algorithm and was used for AlphaGo for example)
The difference is just the size = number of layers (deepness; which is possible due to GPU-based evaluation) and the choosings of activation-functions. ReLU is just working better than the classic Sigmoid or Tanh activations (regarding speed and stability).
Pre-training
I also think, that pre-training was very popular 5-10 years ago but nobody is doing that today (if you got enough data)!
Let me quote from here:
It's true that unsupervised pre-training was initially what made it possible to train deeper networks, but the last few years the pre-training approach has been largely obsoleted.
Nowadays, deep neural networks are a lot more similar to their 80's cousins. Instead of pre-training, the difference is now in the activation functions and regularisation methods used (and sometimes in the optimisation algorithm, although much more rarely).
I would say that the "pre-training era", which started around 2006, ended in the early '10s when people started using rectified linear units (ReLUs), and later dropout, and discovered that pre-training was no longer beneficial for this type of networks.
I can recommend these slides as introduction to modern Deep Learning (as starting point).
Pretraining Is actually gaining again a lot of traction in NLP community, see OpenAI's GPT: the idea is that pretraining acts as an unsupervised initialization step before fine-tuning the model with the supervised data. This is because unlabeled data is much more abundant that labeled counterpart and it can be exploited to derive sensible weights inside the model that express the hidden links inside the dataset structures.
Hope that the explanation was not too goofy :)