Natalino BusaHead of Applied Data Science

Data Scientist, Big and Fast Data ArchitectCurrently at Teradata

Previously: Enterprise Data Architect at INGSenior Researcher at Philips Research

Interests: Spark, Flink, Cassandra, Akka, Kafka, MesosAnomaly Detection, Time Series, Deep Learning

Data Science: approachesSupervised:- you know what the outcome must be

Unsupervised:- you don’t know what the outcome must be

Semi-Supervised:- You know the outcome only for some samples

Popularity of Neural Networks: “The cat neuron”Andrew Ng, Jeff Dean et al:

1000 Machines10 Million images

1 Billion connectionsTrain for 3 days

http://research.google.com/archive/unsupervised_icml2012.html

Popularity of Neural Networks: “AI at facebook”Yann LeCunnDirector of AI research at Facebook

Ask the AI what it sees in the image

“Is there a baby?”Facebook’s AI: “Yes.”

“What is the man doing?”Facebook’s AI: “Typing.”

“Is the baby sitting on his lap?”Facebook’s AI: “Yes.”

http://www.wired.com/2015/11/heres-how-smart-facebooks-ai-has-become/

Data Science: approachesSupervised:- you know what the outcome must be

Unsupervised:- you don’t know what the outcome must be

Semi-Supervised:- You know the outcome only for some samples

Unsupervised Learning- Clustering, Feature extraction

Imagining, Medical data, Genetics, Crime patterns,Recommender systems, Climate hot spots analysis, anomaly detection

… Given a set of items, it answers the question “how can we efficiently describe the collection?It defines a measure of “similarity” between items.

Supervised Learning- Classification

Marketing Churn, Credit Loan, Success rateInsurance Defaulting, Health conditions and patologiesCategorization of wine, real estates,

… Given the values of some properties, it answers the question “to which class/group does this item belong?”

Classification: Dimensionality matters- Number of dimensions or features of your input data- Statistical relations, smoothness of the data- Embedded space

input : 784 dimensionsoutput: 10 classes

input : 4 dimensionsoutput: 3 classes

28x28 pixels

AI, complexity and models

Does it do well on Training Data ?

Does it do well on Test Data ?

Bigger Neural Network(rocket engine)

More Data(rocket fuel)

yes yes

nono

Done?

Different Architecture(new rocket)

no

https://www.youtube.com/watch?v=CLDisFuDnog

Evolution of Machine Learning

Input

Hand Designed Program

Rule-based System

Output

Prof. Yoshua Bengio - Deep Learninghttps://youtu.be/15h6MeikZNg

Evolution of Machine Learning

Input

Hand Designed Program

Input

Rule-based System

Output

Hand Designed Features

Mapping from features

Output

Classic Machine Learning

Prof. Yoshua Bengio - Deep Learninghttps://youtu.be/15h6MeikZNg

Evolution of Machine Learning

Input

Hand Designed Program

Input Input

Rule-based System

Output

Hand Designed Features

Mapping from features

Output

Learned Features

Mapping from features

Output

Classic Machine Learning

RepresentationalMachine Learning

Prof. Yoshua Bengio - Deep Learninghttps://youtu.be/15h6MeikZNg

Evolution of Machine Learning

Input

Hand Designed Program

Input Input

Rule-based System

Output

Hand Designed Features

Mapping from features

Output

Learned Features

Mapping from features

Output

Classic Machine Learning

Input

Learned Features

LearnedComplex features

Output

Mapping from features

RepresentationalMachine Learning

Deep Learning

Prof. Yoshua Bengio - Deep Learninghttps://youtu.be/15h6MeikZNg

“dendrites”

Axon’s response

Activation function

From Biology to a Mathematical Model

Logit model: Perceptron 1 Layer Neural NetworkTakes: n-input features: Map them to a soft “binary” space

∑

x1x2

xn

f

Multiple classes: Softmax From soft binary space to predicting probabilities:Take n inputs, Divide by the sum of the predicted values

∑x1x2

xn

f

∑ f softm

ax Cat: 95%

Dog: 5% Values between 0 and 1Sum of all outcomes = 1

It behaves like a probability, But it’s just an estimate!

Cost function: Supervised LearningThe actual outcome is different than the desired outcome

We measure the difference!This measure can be done in various ways:

- Mean absolute error (MAE) - Mean squared error (MSE)

- Categorical Cross-EntropyCompares estimated probability vs actual probability

Minimize cost: How to Learn?

The cost function depends on:

- Parameters of the model- How the model “composes”

Goal : modify the parameters to reduce the error!

Vintage math from last century

Build deeper networksStack layers of perceptrons

- “Sequential Network”- Back propagate the error SOFTMAX

Input parameters

Classes (estimated probabilities)

Feed

-forw

ard

Cost function

supervised : actual output

Correct

parameters

Some problems- Calculating the derivative of the Cost function

- can be error prone- Automation would be nice!

- Complex network graph = complex derivative

- Dense Layers (Fully connected)- Harder to converge

- Number of parameters grows fast!

- Overfitting and Parsimony- Learn “well”, generalization capacity- Be efficient in the number of parameters

Some Solutions- Calculating the derivative of the Cost function

- Software libraries

- GPU support for computing vectorial and tensorial data

- New Layers Types - Convolution Layers 2D/3D- Dropout layer

- Fast activation functions

- Faster learning methods- Derived from Stochastic Gradient Descend (SGA)- Weight initializations with Auto-Encoders and RBM

Convolutional Networks

Idea 1: reuse the weights across while scanning the imageIdea 2: subsampling results from layers to layers

Fast Activation Functions

Idea: don’t use complex exponential functions, linear functions are fast to compute, and easy to differentiate !

Dropout Layer, Batch Weight NormalizationDropout:Set randomly some of the input to zero.It improves generalization and makes the network function more robust to errors.

Batch Weight Normalization:Normalize the activations of the previous layer at each batch.

Efficient Symbolic DifferentiationThere are good libraries which calculate the derivatives symbolically of an arbitrary number of stacked layers

● efficient symbolic differentiation ● dynamic C code generation ● transparent use of a GPU

CNTK

Efficient Symbolic Differentiation (2)There are good libraries which calculate the derivatives symbolically of an arbitrary number of stacked layers

● efficient symbolic differentiation ● dynamic C code generation ● transparent use of a GPU

>>> import theano

>>> import theano.tensor as T

>>> from theano import pp

>>> x = T.dscalar('x')

>>> y = x ** 2

>>> gy = T.grad(y, x)

>>> f = theano.function([x], gy)

pp(f.maker.fgraph.outputs[0])'(2.0 * x)'

Higher Abstraction Layer: Keras

Keras: Deep Learning library for Theano and TensorFlow

- Easier to stack layers- Easier to train and test- More ready-made blocks

http://keras.io/

Example 1: Iris classificationCategorize Iris flowers based on

- Sepal length/width- Petal length/width

3 classes,Dataset is quite small (150 samples)

- Iris Setosa - Iris Versicolour - Iris Virginica

input : 4 dimensionsoutput: 3 classes

Iris classification: Network

model = Sequential()

model.add(Dense(15, input_shape=(4,)))model.add(Activation('relu'))model.add(Dropout(0.1))

model.add(Dense(10))model.add(Activation('relu'))model.add(Dropout(0.1))

model.add(Dense(nb_classes))model.add(Activation('softmax'))

SOFTMAX

RELU

RELU

Setosa Versicolour Virginica

Dropout 10%

Dropout 10%

Train- Test split 80% - 20%Test accuracy: 96%

Example 2: telecom customer marketingSemi-synthetic datasetThe "churn" data set was developed to predict telecom customer churn based on information about their account. The data files state that the data are "artificial based on claims similar to real world". These data are also contained in the C50 R package.

1 classes (churn)Dataset is quite small (about 3000 samples)

17 input dimensions:

State, account length, area code, phone number,international plan,voice mail plan,number vmail messages,total day minutes,total day calls,total day charge,total eve minutes,total eve calls,total eve charge,total night minutes,total night calls,total night charge,total intl minutes,total intl calls,total intl charge,number customer service calls

Churn telecom: Networkmodel = Sequential()

model.add(Dense(50, input_shape=(17,)))model.add(Activation("hard_sigmoid"))model.add(BatchNormalization())model.add(Dropout(0.1))

model.add(Dense(10))model.add(Activation("hard_sigmoid"))model.add(BatchNormalization())model.add(Dropout(0.1))

model.add(Dense(1))model.add(Activation(sigmoid))

SOFTMAX

RELU

RELU

Churn No-Churn

Dropout 10%

Dropout 10%

Train- Test split 80% - 20%Test accuracy: 82%

Models: Small Data, Big Data- Not all domains have large amount of data

- Think of Clinical Tests, or Lengthy/Costly Experimentations

- Small specialized data set and Neural Networks- Good for complex non-linear separation of classes

Interesting Read:https://medium.com/@ShaliniAnanda1/an-open-letter-to-yann-lecun-22b244fc0a5a#.ngpal1ojx

Conclusions- Neural Networks can be used for small data as well- Other methods might be more efficient in this scenario’s

- Neural Networks are an extension to GLMs and linear regression- Learn Linear Regression, GLM, SVM as well- Random Forests and Boosted Trees are an alternative

- More data = Bigger and better Neural Networks- We have some tools to jump start analysis

Connect on Twitter and Linkedin !

Thanks!