Deep Learning for Sales Professionals

Dr. Miguel Fierro
Data Scientist at Microsoft
@miguelgfierro
https://miguelgfierro.com
Intelligent Cloud GBB Ready | Lisbon Sept 2017

Deep learning solutions Deep learning tecniques

Intelligent Cloud GBB Ready Lisbon 2017 - Dr. Miguel Fierro @miguelgfierro
source: wikipedia
“Our strategy is to
build best-in-class
platforms and services
for an intelligent cloud
and an intelligent
edge infused with
artificial intelligence”
source: annual report 2017

source: PwC AI analysis 2017
BIGGEST
OPPORTUNITY$15.7Trillion by 2030 ~ 14% GPD
Productivity gains ($6.6T)
Automation
Increased demand ($9.1T)
Augmentation Higher quality products

source: PwC AI analysis 2017

DNNs (learn complex patterns
on HUGE amounts of data)
Intelligent applications … such as….
Cloud (flexibility, scalability, efficiency)

Speech
recognition
Natural Language
Processing
Computer
Vision

“The more data,
The better it works*”
* always read the small print
source: https://xkcd.com/1838/

Best bet to solve intelligence

wait, wait, wait…
What makes deep learning deep?
input hidden output

input hidden hidden hidden hidden output
…
…
…

source: http://www.iro.umontreal.ca/~bengioy/yoshua_en/
“Learning multiple levels of
representation to help a learner
accomplish a task of interest,
with higher levels capturing
more abstract concepts
through a deeper composition
of computations”
Yoshua Bengio

low level features high level featuresmedium level features
Interesting paper about representations: https://arxiv.org/abs/1411.1792

Convolutional Neural Networks
(CNN)
Recurrent Neural Networks
(RNN)

MNIST with CNNs
(source here)

source: https://www.cs.toronto.edu/~kriz/cifar.html
CIFAR with ResNet CNN
(source here)

Object detection with FasterRCNN
(source here)
dog

Transfer Learning with CNNs
(source here)

Video classification with 3D CNNs
(source here)

Text classification with
CNNs (source here)
Input text: “It was a breeze to configure and worked straight away”
Books Electronics
Home &
Kitchen
Clothing Movies Health Sports

Speech recognition
with LSTMs
(source here)

Time series forecasting
with LSTMs (source here)

Autoencoders (source here)
Original Learned
Encoder Decoder
Low dimensional
representation

Reinforcement Learning
tutorial (basic & advanced)
source: wikipedia

Generative Adversarial Networks (GANs)
tutorial (source here)

Diagnosis augmentation for lung cancer detection
3D CNN for nodule segmentation +
CNN for nodule classification (info1, info2)
CT scan: horizontal section of lung
Patient health monitoring with sensors (IoT):
Autoencoder on normal class +
reconstruction error abnormal class (info)

Predictive maintenance with sensors:
LSTM for classification
Visual fault detection in components:
Transfer learning with CNNs

Generative
Asset wealth management and prediction:
Reinforcement Learning for optimal action
policy
Bots for insurance policy management:
LSTM&CNN for speech recognition+
CNN for recommendation system

Route optimization:
Reinforcement Learning
Inventory optimization:
LSTM

Content recommendation:
CNNs featurization + clustering
Personalized marketing:
CNN featurization + clustering

Personalized fashion design and
production:
CNNs for image classification +
GANs for image generation
Customer demand forecasting:
LSTM

Demand forecasting:
LSTM
Predictive maintenance:
LSTM

Supply chain defect reduction
based on images:
Transfer learning with CNNs
On-demand production:
LSTM

source: https://github.com/Azure/fast_retraining
Degradation of accuracy
with new data
Airline delay prediction
You need to retrain!!!

Learning can be viewed as finding
a and b in:
y = a.x + b
y (outputs)
x (inputs)
a,b (parameters)

y = a.x + b
1 = a.1 + b
5 = a.5 + b
= a. + b
a=1 b=0
e = ytrue – ypred
ytrue=1
ypred depends on a,b
optimize a,b so e is
minimum

source: R. Rojas: Neural Networks, Springer-Verlag, Berlin, 1996

Sharpening filter
Laplacian filter
Sobel x-axis filter

Convolution Pooling PoolingConvolution
Fully
connected
Fully
connected
Input image Output
predictions
"tabby cat"

Max pooling with 2x2 kernel and stride of 2x2
1 5 3 2
2 6 0 9
1 5 0 1
5 7 2 4
6 9
7 4

input hidden output

SoftmaxReLUtanh

source: Introduction to Natural Language Processing with FastText
Mappings of discrete objects to a high
dimensional space
bird: (1.072, 4.245, …)
dove: (1.134, 4.079, …)
woman: (-1.385, -2.104, …)
man: (-2.207, 2.118, …)

input
hidden
output
hidden
hidden
ti ti+1 ti+2 ti+3
…
…
…

number of layers

source: https://en.wikipedia.org/wiki/Long_short-term_memory

CNTK
TensorFlow
https://github.com/ilkarman/DeepLearningFrameworks: 5 frameworks, same problem

GIVE ME DATA!!!!
CONCEPT DRIFT IS YOUR FRIEND
MASTER THE CORE PRINCIPLES
KNOW THE FRAMEWORKS&TOOLS

Intelligent Cloud GBB Ready Lisbon 2017
The code from these slides is published at:
https://github.com/ilkarman/DeepLearningFrameworks
Attribution of some images:
• http://morguefile.com
• https://unsplash.com
• Ana Corrales Photography
• http://wikipedia.org
Dataset citation:
• CIFAR-10. Learning Multiple Layers of Features from Tiny
Images, Alex Krizhevsky, 2009.
https://www.cs.toronto.edu/~kriz/cifar.html
• IMDB. https://s3.amazonaws.com/text-datasets/imdb.npz

Deep Learning for Sales Professionals

More Related Content

Viewers also liked (8)

Similar to Deep Learning for Sales Professionals (20)

More from Miguel González-Fierro (10)

Recently uploaded (20)

Deep Learning for Sales Professionals