Machine Learning @ Amazon Ralf Herbrich Amazon 6/29/17 1
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 2
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 3
Machine Learning: Formal Definition • Labelled Data • Unlabelled Data • Probability is a central concept in Machine Learning! 6/29/17 4
Why Probability? 1. Mathematics of Uncertainty (Cox’ axioms) 6/29/17 5
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! 6/29/17 6
Why Probability? 1. Mathematics of Uncertainty (Cox’ axioms) 2. Variables and Factors map to Memory & CPU 6/29/17 7
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 6/29/17 8
Factor Graphs and Cloud Computing Belief Store (“Memory”) ϑ 1 ϑ 2 ϑ 3 ϑ ϑ 4 ϑ 5 Message Passing (“Communicate”) Data Messages (“Compute”) Y 1 Y 2 Y 3 Y 4 Y 5 Y 6 Y 7 6/29/17 9
Factor Graphs and MXNet 6/29/17 10
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 11
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* Some constraints might not be obvious: Daily Revenue $49,117,808.22 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- Feed
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 13
Locations S9 ML Berlin ML Seattle ML Cambridge A2Z Ivona A9 Evi ML Los Angeles ML Bangalore 6/29/17 14
Machine Learning Opportunities @ Amazon Retail Customers Seller Catalog Digital • Demand • Product • Fraud Detection • Browse-Node • Named-Entity Forecasting Recommendation Classification Extraction • Predictive Help • Vendor Lead Time • Product Search • Meta-data • XRay • Seller Search & Prediction validation • Visual Search Crawling • Plagiarism • Pricing • Review Analysis Detection • Product Ads • Packaging • Hazmat Prediction • Echo Speech • Shopping Advice Recognition • Substitute • Customer Problem Prediction • Knowledge Detection Acquisiion 6/29/17 15
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 16
Demand 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. 6/29/17 17
Learning and Prediction P ( z i t | θ ) ∼ sales/demand time Learning Forecasting Model Parameters
Slow Moving Inventory Typical midsize dataset: • About 5M items • About 4.5B item-days • About 98% zero demand
Sampling Predictions P ( z i t | θ ) ∼ • 0 or ≥1 ? Binary classification #1 • 1 or ≥2 ? Binary classification #2 • If ≥2: Count regression z-2
x 1 x 2 x 3 x 4 x 5 l 1,2 l 2,2 l 3,2 l 4,2 l 5,2 Latent State l 1,1 l 2,1 l 3,1 l 4,1 l 5,1 l 1,0 l 2,0 l 3,0 l 4,0 l 5,0 y 1,2 y 2,2 y 3,2 y 4,2 y 5,2 Multistage Likelihood 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 z 1 z 2 z 3 z 4 z 5
In Practice x 1 x 2 x 3 x 4 x 5 x 1 x 2 x 3 x 4 x 5 x 1 x 2 x 3 x 4 x 5 l 1,2 l 2,2 l 3,2 l 4,2 l 5,2 l 1,1 l 2,1 l 3,1 l 4,1 l 5,1 l 1,0 l 2,0 l 3,0 l 4,0 l 5,0 l 1,0 l 2,0 l 3,0 l 4,0 l 5,0 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 z 1 z 2 z 3 z 4 z 5 z 1 z 2 z 3 z 4 z 5 z 1 z 2 z 3 z 4 z 5
Modelling Out of Stock GLM Bridge
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 24
Product Machine Translation (2013 – 2015) Lifetime Profit Human Translation Machine Translation Products Selection Gap 6/29/17 25
Sockeye • Sequence-to-sequence Neural Machine Translation package build on MXNet: https://github.com/awslabs/sockeye • Support both CPU and GPU encoding and decoding • Training • Translating 6/29/17 26
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 27
Automated Produce Inspection: The Goal New Automated Inspection Current Inspection Computer Vision 6/29/17 28
Challenges • Illumination • Clutter/Occlusions • Viewpoint • Scale • Intra-class variability
Predicting Longevity Strawberry ID Age à 6/29/17 2016 (c) Amazon 30
Age Aligned Strawberries (Test Set)
Overview • Machine Learning in Practise • Probabilities • Finite Resource • Machine Learning @ Amazon • Forecasting • Machine Translation • Visual Systems • Conclusions and Challenges 6/29/17 32
Conclusions • 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! 6/29/17 33
Thanks! 6/29/17 2016 (c) Amazon 34
Recommend
More recommend
