Machine Learning @ Amazon - Ralf Herbrich Amazon - WSDM
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Machine Learning @ Amazon
Ralf Herbrich
Amazon
2/21/17 1
Background
1992 – 1997 (Berlin, Diploma)
1997 – 2000 (Berlin, PhD)
2000 – 2009 (Microsoft Research)
2009 – 2011 (Microsoft)
2011 – 2012 (Facebook)
2012 – Present (Amazon)
2/21/17 2
Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 3
Our Customers 2/21/17 4
Devices 2/21/17 5
Amazon’s Virtuous Cycles
Lower Cost Lower
Structure Prices
Selection &
Convenience
Customer
Sellers Growth Experience
Traffic
2/21/17 2016 (c) Amazon 3
Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 7
Artificial Intelligence and Machine Learning
Science Artificial Intelligence Machine Learning
• Computer Science • Knowledge Representation • Rule Extraction from Past (Training)
• Statistics • Knowledge Extraction • Forecast of Future (Prediction)
• Neuroscience • Reasoning • Taking Actions Now (Decision
• Operations Research • Planning Making)
2/21/17 2016 (c) Amazon 8
Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 9
Machine Learning: Formal Definition • Labelled Data • Unlabelled Data • Probability is a central concept in Machine Learning! 2/21/17 10
Why Probability? 1. Mathematics of Uncertainty (Cox’ axioms) 2/21/17 11
Cox Axioms: Probabilities and Beliefs
• Design: System must assign degree of plausability to each logical
statement A.
• Axiom:
• is a real number
• is independent of Boolean rewrite
•
P must be a probability measure!
2/21/17 12Why Probability? 1. Mathematics of Uncertainty (Cox’ axioms) 2. Variables and Factors map to Memory & CPU 2/21/17 13
Factor Graphs
• Definition: Graphical representation of product structure of a
function (Wiberg, 1996)
• Nodes: = Factors = Variables
• Edges: Dependencies of factors on variables.
• Semantic: a b
• Local variable dependency of factors c
2/21/17 14Inference in a Factor Graph
s1 s2 s3 s4
t1 t2 t3
y12 y23
2/21/17 15Factor Graphs and Cloud Computing
Belief Store
(“Memory”)
ϑ1 ϑ2 ϑϑ3 ϑ4 ϑ5
Message Passing
(“Communicate”)
Data Messages
(“Compute”)
Y1 Y2 Y3 Y4 Y5 Y6 Y7
2/21/17 16Factor Graphs and MXNet 2/21/17 2016 (c) Amazon 17
Why Probability? 1. Mathematics of Uncertainty (Cox’ axioms) 2. Variables and Factors map to Memory & CPU 3. Decouple Data Modeling and Decision Making 2/21/17 18
Infer-Predict-Decide Cycle
Inference:
Decision Making: P(Parameters) + Data à
Loss(Action,Data) + P(Data) P(Parameters|Data)
à Action • Requires a (structural) model
P(Data|Parameters)
• Business-loss not learning-loss! • Allows to incorporate prior
• Often involves optimization! information
P(Parameters|Data)
Prediction:
P(Parameters) +
Data à P(Data)
• Requires
integration/summ
ation of parameter
uncertainty
• Does not change
state!
2/21/17 19Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 20Finite Resources: Time • Real-Time Prediction Service: 5ms • Number of search & ads candidates to rank: 10,000 è 500 ns/candidate • Time to read from main memory: 10ns è 50 variables • L1/L2/L3 ranking pipelines to use more complex predictors • Real-Time Learning Service: 1B examples/day • Number of second per day: 86,400 è 86,400 ns/example • Time to write to RAM: 100ns è 864 writes only (excluding disc access!) • Sparse models are the result of time constraints!
Finite Resource: Cost
Economics 101
• Profit = Revenue – Cost
• In the long run, a business that generates negative profits is not viable!
Facebook 2015 It’s power, stupid!
Annual Revenue $17,928,000,000.00*
Daily Revenue $49,117,808.22 Some constraints might not be obvious:
building new datacenters and powering
Number of DAU 1,038,000,000** them is non-trivial.
Number of Story Candidates 1,500***
Example: 1 GPU box = 20 CPU boxes
Number of Daily Stories 1.557E+12 (in terms of power consumption)
Maximum Cost per Story Candidate $0.0000315
*http://www.statista.com/statistics/277229/facebooks-annual-revenue-and-net-
income/
**http://www.statista.com/statistics/346167/facebook-global-dau/
***https://www.facebook.com/business/news/News-Feed-FYI-A-Window-Into-News-
FeedOverview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 23Locations
S9
ML Berlin
ML Seattle ML Cambridge
A2Z
A9 Ivona
Evi
ML Los Angeles
ML Bangalore
2/21/17 24Machine Learning Opportunities @ Amazon
Retail Customers Seller Catalog Digital
• Demand • Product • Fraud Detection • Browse-Node • Named-Entity
Forecasting Recommendation • Predictive Help Classification Extraction
• Vendor Lead Time • Product Search • Seller Search & • Meta-data • XRay
Prediction • Visual Search Crawling validation • Plagiarism
• Pricing • Product Ads • Review Analysis Detection
• Packaging • Shopping Advice • Hazmat Prediction • Echo Speech
• Substitute • Customer Problem Recognition
Prediction Detection • Knowledge
Acquisiion
2/21/17 25Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 26Fashion Forecasting (2014 – 2016)
Setting
• Given past sales of a fashion product, predict market demand up to 18 months into the future
Challenges
• Sparsity: Huge skew – many products sell very few items
• Size: Variable for Amazon retail in buying but not a variable for end customers
• Seasonal: Most fashion products only run for one season – often 12 months in advance
• Distributions: Future is uncertain è predictions must be distributions
• Scale: 5M+ fashion products in each market è 109 training examples
• Censored: Past sales ≠ past demand (inventory constraint)
2/21/17 2016 (c) Amazon 27Demand Forecasting
Example fashion product to illustrate the challenges of forecasting.
Training Range: Non-fashion items Missing Features or Input:
have longer training ranges that we Unexplained spikes in demand are
can leverage. Need to information likely caused by missing features or
share across new and old products. incomplete input data.
Seasonality: This item has Christmas
seasonality with higher growth over time.
This is where we need growth features in
addition to date features.
2/21/17 2016 (c) Amazon 28Learning and Prediction
sales/demand
P (zi t | ✓) ⇠
time
Learning
Forecasting
Model ParametersSlow Moving Inventory
Typical midsize dataset:
• About 5M items
• About 4.5B item-days
• About 98% zero demandSampling Predictions
P (zi t | ✓) ⇠
• 0 or ≥1 ?
Binary classification #1
• 1 or ≥2 ?
Binary classification #2
• If ≥2:
Count regression z-2x1 x2 x3 x4 x5
l1,2 l2,2 l3,2 l4,2 l5,2
Latent State
l1,1 l2,1 l3,1 l4,1 l5,1
l1,0 l2,0 l3,0 l4,0 l5,0
y1,2 y2,2 y3,2 y4,2 y5,2
Multistage
Likelihood
y1,1 y2,1 y3,1 y4,1 y5,1
y1,0 y2,0 y3,0 y4,0 y5,0
z1 z2 z3 z4 z5In Practice
x1 x2 x3 x4 x5 x1 x2 x3 x4 x5 x1 x2 x3 x4 x5
l1,2 l2,2 l3,2 l4,2 l5,2
l1,1 l2,1 l3,1 l4,1 l5,1
l1,0 l2,0 l3,0 l4,0 l5,0 l1,0 l2,0 l3,0 l4,0 l5,0
y 1,2 y 2,2 y 3,2 y 4,2 y 5,2 y 1,2 y 2,2 y 3,2 y 4,2 y 5,2
y 1,1 y 2,1 y 3,1 y 4,1 y 5,1 y 1,1 y 2,1 y 3,1 y 4,1 y 5,1
y 1,0 y 2,0 y 3,0 y 4,0 y 5,0 y 1,0 y 2,0 y 3,0 y 4,0 y 5,0 y 1,0 y 2,0 y 3,0 y 4,0 y 5,0
z1 z2 z3 z4 z5 z1 z2 z3 z4 z5 z1 z2 z3 z4 z5Modelling Out of Stock
GLM
BridgeOverview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 35Product Machine Translation (2013 – 2015)
Lifetime Profit
Human Translation
Machine Translation
Selection Gap Products
2/21/17 2016 (c) Amazon 36Machine Translation Pipeline
Input Request Input Normalization Tokenization
Detection & Escaping of
Lowercasing Sentence Segmentation
Non-translatables
Re-insertion of (converted)
Translation/Decoding Recasing
Nontranslatables
Translated Request Post-processing De-Tokenization
2/21/17 2016 (c) Amazon 37Machine Translation: Deep Dive
p ( English) ´ p (German | English)
p ( English | German) =
p (German)
µ p ( English) ´ p (German | English)
Language Translation
Model Model
• Language Model: What are fluent English sentences?
• Translation Model: What English sentences account
well for a given German sentence?
2/21/17 2016 (c) Amazon 38Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 39Automated Produce Inspection: The Goal
Current Inspection New Automated Inspection
Computer Vision
2/21/17 2016 (c) Amazon 40Challenges • Illumination • Clutter/Occlusions • Viewpoint • Scale • Intra-class variability
Computer Vision Pipeline
• Segmentation
• Defect Mining
• Produce Classification
Amazon ConfidentialPredicting Longevity
Strawberry ID
2/21/17
Age à
2016 (c) Amazon 43Age Aligned Strawberries (Test Set)
Overview
• What is Amazon?
• Machine Learning in Practise
• Probabilities
• Finite Resource
• Machine Learning @ Amazon
• Forecasting
• Machine Translation
• Visual Systems
• Conclusions and Challenges
2/21/17 45Conclusions • Machine Learning “translates” data from the past into accurate predictions about the future! • In practice, probabilistic models and finite resources matter. • Machine Learning helps to improve customer experience at Amazon! 2/21/17 2016 (c) Amazon 46
Thanks! 2/21/17 2016 (c) Amazon 47
You can also read
