EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM TRACKING WITH GOOGLE ANALYTICS.pptx
Bol.com
1. 25-9-2017
Deep Learning for E-commerce
recognize content in images
Big Data Expo ‘17
Barrie Kersbergen
#krsbrgn
Computer Science
Machine Learning?
Formal definition:
Machine learning (ML) is a subfield of computer science … that
gives computers the ability to learn without being
explicitly programmed
1. Taken from https://en.wikipedia.org/wiki/Machine_learning at 2016-08-05
Machinelearning
2. 25-9-2017
Deep Learning?
Formal definition:
Deep learning (DL) is a branch of machine learning …
…(computer) attempt to model different levels
abstractions in data (automatically).
sourcehttps://en.wikipedia.org/wiki/Deep_Learning at 2016-08-05
Computer Science
Machinelearning
Deep learning
ML: (Statistical) classification
Classifier
software
“Offline/Batch”
Every week
train
Classifier
software
“Offline/Batch”
predict
Training phase
Prediction phase
petal_length petal_width
5.6 1.6
class
?
class
Iris-virginica
petal_length petal_width class
5.5 1.8 Iris-virginica
4.8 1.8 Iris-versicolor
1.4 0.2 Iris-setosa
1.4 0.2 Iris-setosa
train.csv
prediction.csv
ML: (Statistical) classification
petal_length petal_width class
5.5 1.8 Iris-virginica
4.8 1.8 Iris-versicolor
1.4 0.2 Iris-setosa
1.4 0.2 Iris-setosa
petal_length petal_width
5.6 1.6
class
?
train.csv
prediction.csv
• Onsite bot detection
• Customer churning
• Fraud detection
• Customer problem detection
• Catalog attribute prediction
• …
• but …
3. 25-9-2017
Data’s rarely structured in rows and columns
Product images
Product reviews
Product descriptions
Customer support
Logfiles from our applications
Data scientists/analysts spend 80 percent of their time sourcing, cleaning, and preparing data.
(source Data Science Report Crowdflower 2016)
‘Common’ machine learning
Traditional
ML software
(human performance)
Deep learning software
“automatic feature engineering”
(superhuman performance)
Unstructured dataStructured data
Data scientist/analystspetal_length petal_width class
5.5 1.8 Iris-virginica
4.8 1.8 Iris-versicolor
1.4 0.2 Iris-setosa
1.4 0.2 Iris-setosa
“feature engineering:
painful black art,
transforming data”
Features are key to statistics and machine learning
Deep Learning
• Deep learning f.k.a. Artificial Neural Networks. (1950’s)
• Loosely inspired by how our brain functions. The software can teach
themselves to understand things like images, text and speech.
• Major advances in 1980’s 1990’s. The DL quality improves when:
• Massive data
• (Massive) Cheap computing horse power.
Popular Deep Learning ‘network configs’
Super human results when working
on unstructured data
4. 25-9-2017
image class
Long sleeve
Perfume
Flower
Short sleeve
Deep Convolutional Neural Network
Classifier
software
(Tensorflow)
“Offline” (Batch)
Every week
train
Classifier
software
(Tensorflow)
“Offline” (Batch)
predict
Training phase
Prediction phase
class
Long sleeve
image class
?
Deep Convolutional Neural Network
Deep learning best practice:
For every distinct “class” you need at least 1.000
training images
1000 classes * 1000 images=
1.000.000 training images
‘training’ tensorflow will take 1 year on a modern pc *
* FireCaffe: near-linear acceleration of deep neural network training on compute clusters, 2016
It is common to train a dozen times to see which settings work best
Deep Convolutional Neural Network
* FireCaffe: near-linear acceleration of deep neural network training on compute clusters, 2016
machine learning
* i7 CPU GPU
Clock speed 4.2 GHz 1.7 GHz
# of processor ‘cores’ 8 2560
Duration (days) 365 21
5. 25-9-2017
Transfer learning: reduce training from 21 days to 2 minutes
• Download a model that’s already trained on millions of images.
• Retrain that model with our images.
• Advantages:
• Still good results if lesser training images per class
• (Re)training done in minutes instead of 21 days
• Get really good results
Use-case: Predict t-shirt sleeve length
What do you need to install?
(opensource software)
• Python 3 (64bit)
• Google Tensorflow (CPU or GPU edition). Runs native on
Windows 7 and Windows 10.
Deep Convolutional Neural Network with transfer learning
6. 25-9-2017
• Tensorflow example is
was missing code to do
predictions
Deep Convolutional Neural Network with transfer learning DCNN-TL project:
Each directory now contains 100+ distinct images
23
Place training images in ‘train’ directory
10. 25-9-2017
Deep Convolutional Neural Network with Transfer Learning
• Predict from structured data using RandomForest
• Predict from unstructured data using DCNN with TL
• Improve training duration
• Applications
DL: What’s the catch?
• DL
• CPU 365 days = 365*24hrs*91J*3600sec/h= 2869.8 MJ
• GPU 21 days = 21days*24hrs*(91J+280J)*3600sec/h= 673.1MJ
• Traditional ML (quality will not match DL)
• 20 hours on CPU=20hrs*91J*3600sec/h= 6.5 MJ
* Usually we only train once in a week, month.
That’s 104 times more power consumption! *
(When not using transfer learning)
• Barrie Kersbergen
Deep Learning for E-commerce
You got Questions we’ve got Answers
#krsbrgn
recognize content in images