Recurrent

Neural Networks
 Deep Learning

● We’ve used Neural Networks to solve

Classification and Regression problems,
but we still haven’t seen how Neural
Networks can deal with sequence
information.
 Deep Learning

● Just as CNNs were more effective for use

with 2D image data, RNNs are more
effective for sequence data (e.g. time-
stamped sales data, sequence of text,
heart beat data, etc…)
 Deep Learning

● Time Series
 Deep Learning

● Time Series
 Deep Learning

● Sequences

Hello how are you

Hello how are you today?

 Deep Learning

● Section Overview
○ RNN Theory
○ LSTMs and GRU Theory
○ Basic Implementation of RNN
○ Time Series with an RNN
○ Exercise and Solution
Let’s get started!
Recurrent Neural
Networks Theory
 Deep Learning

● Examples of Sequences
○ Time Series Data (Sales)
○ Sentences
○ Audio
○ Car Trajectories
○ Music
 Deep Learning

● Let’s imagine a sequence:

○ [1,2,3,4,5,6]
● Would you be able to predict a similar
sequence shifted one time step into
the future?
○ [2,3,4,5,6,7]
 Deep Learning

● To do this properly, we need to somehow

let the neuron “know” about its previous
history of outputs.
● One easy way to do this is to simply feed
its output back into itself as an input!
 Deep Learning

● Normal Neuron in Feed Forward Network

Output

Activation Function

Σ Aggregation of Inputs

Input
 Deep Learning

● Recurrent Neuron - Sends output

Output
back to itself!
- Let’s see what
this looks like
Σ
over time!
Input
 Deep Learning

● Recurrent Neuron
Output
Output at t-1

Σ = Σ

Input Input
at t-1
 Deep Learning

● Recurrent Neuron
Output
Output at t-1

Σ = Σ

Input Input
at t-1
Time
 Deep Learning

● Recurrent Neuron
Output Output
Output at t-1 at t

Σ = Σ Σ

Input Input Input

at t-1 at t
Time
 Deep Learning

● Recurrent Neuron
Output Output Output
Output at t-1 at t at t+1

Σ = Σ Σ Σ

Input Input Input Input

at t-1 at t at t+1
Time
 Deep Learning

● Cells that are a function of inputs from

previous time steps are also known as
memory cells.
● RNN are also flexible in their inputs and
outputs, for both sequences and single
vector values.
 Deep Learning

● We can also create entire layers of

Recurrent Neurons...
 Deep Learning

● ANN Layer with 3 Neurons:

X y
 Deep Learning

● RNN Layer with 3 Neurons:

X y
 Deep Learning

● “Unrolled” layer.
Output Output Output
at t=0 at t+1 at t+2

Input Input Input

at t=0 at t+1 at t+2

Time
 Deep Learning

● RNN are also very flexible in their inputs

and outputs.
● Let’s see a few examples.
 Deep Learning

● Sequence to Sequence (Many to Many)

Y0 Y1 Y2 Y3 Y4

X0 X1 X2 X3 X4
 Deep Learning

● Given 5 previous words, predict the next 5

Y0 Y1 Y2 Y3 Y4

X0 X1 X2 X3 X4
 Deep Learning

● Sequence to Vector (Many to One)

Y0 Y1 Y2 Y3 Y4

X0 X1 X2 X3 X4
 Deep Learning

● Given 5 previous words, predict next word

Y0 Y1 Y2 Y3 Y4

X0 X1 X2 X3 X4
 Deep Learning

● Vector to Sequence (One to Many)

Y0 Y1 Y2 Y3 Y4

X0 0 0 0 0
 Deep Learning

● Given 1 word predict the next 5 words

Y0 Y1 Y2 Y3 Y4

X0 0 0 0 0
 Deep Learning

● A basic RNN has a major disadvantage,

we only really “remember” the previous
output.
● It would be great it we could keep track of
longer history, not just short term history.
 Deep Learning

● Another issue that arises during training

is the “vanishing gradient”.
● Let’s explore vanishing gradients in more
detail before moving on to discussing
LSTM (Long Short Term Memory Units).
 Exploding and
Vanishing Gradients
 Deep Learning

● As our networks grow deeper and more

complex, we have 2 issues arise:
○ Exploding Gradients
○ Vanishing Gradients
● Recall that the gradient is used in our
calculation to adjust weights and
biases in our network.
 Deep Learning

● Let’s think about a network.

Hidden Layers
 Deep Learning

● For complex data we need deep networks

Hidden Layers
 Deep Learning

● Issues can arise during backpropagation

Hidden Layers
 Deep Learning

● Backpropagation goes backwards from

the output to the input layer, propagating
the error gradient.
● For deeper networks issues can arise from
backpropagation, vanishing and
exploding gradients!
 Deep Learning

● As you go back to the “lower” layers,

gradients often get smaller, eventually
causing weights to never change at lower
levels.
● The opposite can also occur, gradients
explode on the way back, causing issues.
 Deep Learning

● Let’s discuss why this might occur and

how we can fix it.
● Then in the next lecture we’ll discuss how
these issues specifically affect RNN and
how to use LSTM and GRU to fix them.
 Deep Learning

● Why does this happen?

1
Output

0
0

z = wx + b
 Deep Learning

● Why does this happen?

1
Output

0
0

z = wx + b
 Deep Learning

● The derivative can be much smaller!

 Deep Learning

● When n hidden layers use an activation

like the sigmoid function, n small
derivatives are multiplied together.
● The gradient could decrease
exponentially as we propagate down to
the initial layers.
 Deep Learning

● Using Different Activation Functions

Output

z = wx + b
 Deep Learning

● The ReLu doesn’t saturate positive values.

Output

z = wx + b
 Deep Learning

● “Leaky” ReLU

Output

z = wx + b
 Deep Learning

● Exponential Linear Unit (ELU)

1
Output 0

-1
0

z = wx + b
 Deep Learning

● Another solution is to perform batch

normalization, where your model will
normalize each batch using the batch
mean and standard deviation.
 Deep Learning

● Choosing different initialization of

weights can also help alleviate these
issues (Xavier Initialization).
 Deep Learning

● Apart from batch normalization,

researchers have also used “gradient
clipping”, where gradients are cut off
before reaching a predetermined limit
(e.g. cut off gradients to be between -1
and 1)
 Deep Learning

● RNN for Time Series present their own

gradient challenges, let’s explore special
LSTM (Long Short Term Memory) neuron
units that help fix these issues!
LSTM and GRU Units
 Deep Learning

● Many of the solutions previously presented

for vanishing gradients can also apply to
RNN: different activation functions, batch
normalizations, etc…
● However because of the length of time
series input, these could slow down
training
 Deep Learning

● A possible solution would be to just

shorten the time steps used for prediction,
but this makes the model worse at
predicting longer trends.
 Deep Learning

● Another issue RNN face is that after awhile

the network will begin to “forget” the first
inputs, as information is lost at each step
going through the RNN.
● We need some sort of “long-term memory”
for our networks.
 Deep Learning

● The LSTM (Long Short-Term Memory) cell

was created to help address these RNN
issues.
● Let’s go through how an LSTM cell works!
 Deep Learning

● A typical RNN
Output Output Output
at t-1 at t at t+1

Input Input Input

at t-1 at t at t+1
 Deep Learning

● RNN
Output
at t

Input
at t
 Deep Learning

● RNN

Output t-1
Output t

Input t
 Deep Learning

● RNN

Ht-1 Ht

Xt
 Deep Learning

● RNN

Ht-1 Ht=tanh(W[Ht-1,Xt]+b) Ht

Xt
 Deep Learning

● LSTM
 Deep Learning

● LSTM Output at time t

Long Term New Long

Memory Term Memory

Short Term New Short

Memory Term Memory

Input at time t
 Deep Learning

● LSTM Output at time t

Long Term New Long

Memory Forget Output Term Memory
Gate Gate

Input Update
Gate Gate New Short
Short Term
Memory Term Memory

Input at time t
 Deep Learning

● Gates optionally let information through

σ X

0
 Deep Learning

● LSTM Output at time t

Long Term New Long

Memory Term Memory

Short Term New Short

Memory Term Memory

Input at time t
 Deep Learning

● LSTM ht

Ct-1 Ct

ht-1
ht

xt
 An LSTM Cell
Output at t
Here we can
X + see the entire
tanh
X
LSTM cell.
X
σ σ tanh σ Let’s go
through the
process!
Input at t

http://colah.github.io/posts/2015-08-Understanding-LSTMs/
 An LSTM Cell
ht
Here we can
Ct-1 Ct
X + see the entire
tanh
X
LSTM cell.
X

ht-1
σ σ tanh σ Let’s go
ht
through the
xt
process!
 An LSTM Cell
ht

Ct-1 Ct
X +
tanh
X
X

ht-1
σ σ tanh σ
ht

xt
 An LSTM Cell
ht

Ct-1 Ct
X +
tanh
ft X
X

ht-1
σ σ tanh σ
ht

xt
 An LSTM Cell
ht

Ct-1 Ct
X +
tanh
it X
~
Ct X

ht-1
σ σ tanh σ
ht

xt
 An LSTM Cell
ht

Ct-1 Ct
X +
tanh
ft it X
~
Ct X

ht-1
σ σ tanh σ
ht

xt
 An LSTM Cell
ht

Ct-1 Ct
X +
tanh
X
X

ht-1
σ σ tanh σ
ht

xt
 An LSTM Cell with “peepholes”
ht

Ct-1 Ct
X +
tanh
X
X

ht-1
σ σ tanh σ
ht

xt
 Gated Recurrent Unit (GRU)
ht

ht-1 Ct
X
+
X rt 1- zt X

σ σ tanh

xt
 Deep Learning

● Fortunately with our deep learning python

library , we simply need to call the import
for RNN or LSTM instead of needing to
code all of this ourselves!
● Let’s explore how to use LSTMs with
Python code!
Basic RNN
 Deep Learning

● Let’s now explore how to use RNN on a

basic time series, such as a sine wave.
● Before we jump to the notebook, let’s
quickly discuss what RNN sequence
batches look like.
 Deep Learning

● Let’s imagine a simple time series:

○ [0,1,2,3,4,5,6,7,8,9]
● We separate this into 2 parts:
○ [0,1,2,3,4,5,6,7,8] [9]
● Given training sequence, predict the
next sequence value.
 Deep Learning

● Keep in mind we can usually decide how

long the training sequence and predicted
label should be:
○ [0,1,2,3,4] [5,6,7,8,9]
 Deep Learning

● We can also edit the size of the training

point, as well as how many sequences to
feed per batch:
○ [0,1,2,3] [4]
○ [1,2,3,4] [5]
○ [2,3,4,5] [6]
 Deep Learning

● So how do we decide how long the

training sequence should be?
○ There is no definitive answer, but it
should be at least long enough to
capture any useful trend information.
 Deep Learning

● For example, if dealing with seasonal data:

 Deep Learning

● If this is monthly, we should include at

least 12 months in the training sequence
 Deep Learning

● This often takes domain knowledge and

experience, as well as simply
experimenting and using RMSE to
measure error of forecasted predictions.
● Typically a good starting choice for the
label is just one data point into the future.
 Deep Learning

● How do we forecast with RNNs?

● Let’s imagine all our data is:
○ [0,1,2,3,4,5,6,7,8,9]
● And we trained on sequences such as:
○ [0,1,2,3] [4]
○ [1,2,3,4] [5]
○ [2,3,4,5] [6]
 Deep Learning

● Then our forecasting technique is to

predict a time step ahead, and then
incorporate our prediction into the next
sequence we predict off of.
● Let’s walk through a quick example!
 Deep Learning

● How do we forecast with RNNs?

● Let’s imagine all our data is:
○ [0,1,2,3,4,5,6,7,8,9]
● And we trained on sequences such as:
○ [0,1,2,3] [4]
○ [1,2,3,4] [5]
○ [2,3,4,5] [6]
 Deep Learning

● [6,7,8,9] [10] Forecast prediction!

 Deep Learning

● [6,7,8,9] [10] Forecast prediction!

● Then to keep predicting further:
 Deep Learning

● [6,7,8,9] [10] Forecast prediction!

● Then to keep predicting further:
● [7,8,9,10] [11.2]
 Deep Learning

● [6,7,8,9] [10] Forecast prediction!

● Then to keep predicting further:
● [7,8,9,10] [11.2]
● [8,9,10,11.2] [12.4]
 Deep Learning

● [6,7,8,9] [10] Forecast prediction!

● Then to keep predicting further:
● [7,8,9,10] [11.2]
● [8,9,10,11.2] [12.4]
● [9,10,11.2,12.4] [14]
 Deep Learning

● [6,7,8,9] [10] Forecast prediction!

● Then to keep predicting further:
● [7,8,9,10] [11.2]
● [8,9,10,11.2] [12.4]
● [9,10,11.2,12.4] [14]
● [10,11.2,12.4,14] Completed Forecast
 Deep Learning

● Let’s explore this further with Python!

Basic RNN on a
Sine Wave
 Deep Learning

● Let’s now train and evaluate our RNN

● Recall our original data:
 Deep Learning

● We split this into train_set and test_set:

 Deep Learning

● During training, we could evaluate

performance on the test data.
 Deep Learning

● Let’s zoom in on that range!

 Deep Learning

● As we train, we will forecast on this range

to visually see the training.
 Deep Learning

● As we train, we will forecast on this range

to visually see the training.
 Deep Learning

● As we train, we will forecast on this range

to visually see the training.
 Deep Learning

● Eventually once we are satisfied with the

results on the test portion, it will be time to
forecast into the unknown future.
● This means retraining on all available data
(both train and test) and forecasting
beyond the scope of the original data.
 Deep Learning

● Keep in mind, for real data, we would no

longer have values to compare these
predictions to!
 RNN on a
Time Series
RNN Exercises
RNN Exercises
Solutions
Multivariate Time Series
with LSTM RNNs
 Deep Learning

● As a quick bonus, we’ll discuss how to use

LSTMs and RNNs to predict multivariate
time series.
● Keep in mind, there are a few cons to using
LSTMs for this approach!
 Deep Learning

● As with all neural networks, the model is

essentially a black box, difficult to interpret.
● Also there are many well-studied
alternatives that are simpler, such as
SARIMAX and VARMAX models.
 Deep Learning

● We highly recommend you try those more

conventional approaches before settling
on LSTMs or RNNs for multivariate time
series data.
● Fortunately, setting up for multivariate
data only requires 2 main changes.
 Deep Learning

● Multivariate Time Series:

○ Change input shape in LSTM layer to
reflect 2-D structure
○ Final dense layer should have a
neuron per feature/variable.
● Let’s explore these changes!