Facebook ML Open House

2017-11-14 / Facebook Bldg 13

Intro

AI at facebook

25% ML engineer,

next gen challenges

• ai at massive scale: trillions of examples per models
• extreme vision: billion visual classes
• How to solve problems like false news and polarizations?
• How do we build closer and safer communities?
• connect you to the best stories?
• ad team: serving the best ads to you
• language barriers
• natural conversational interfaces
• 6 speakers and posters

1.Applying Machine Learning to Ads Integrity

Emanual Strauss/ Engineering Manager

Ads, Policies and Quality

example of low quality ads

• shock and scare
• before and after
• amount of text in images
• https://facebook.com/policies/ads

Ad Review Overview

Detecting low quality ads Ad submission -> feature engineering -> model scoring -> human computing decision -> ads ranking & delivery

Challenges

• large class imbalances (most of ads are good quality)
• limited human reviewer capacity and accuracy is not perfect
• feature engineering on ad content(text, image, video, audio, etc)
• most traing data comes from already known patterns
• ecosystem is dynamic and patters evolve over time
• global reach, internationalization and multiple languages
• scalability

Large Class Imbalance

• goal
  • measure our review system
  • identify new trends in not-sutable ads
• detecing low quality Ads

Account Quality

• good group of social connection produces good results

image and video understanding

• real photo or drawing or meme?
• who/what is in the picture>
• nearest neightbors embedding
• similarity

Amount of Text detection

• images can’t have more than 20% text

2.Tacking Misinformation, fakenews

Outline

• problems and our role in curbing misinformation
• steps towards solutions
• challenges

Misinformation

• what is truth? are we editors?

how?

• 3rd party fact checkers
  • snopes, politifact, factcheck.org le monde.fr correct!v niewuescheckers nul.nl

challenges

• hard ml problem
  • labels are sparse and expensive
  • fast
  • mostly text
  • hard to do pictures
• examples
  • images: burning flag, photoshoped
• scaling globally
  • 3 PFCs are not present in every country
  • every country has its own nature of fake claims

3.Make Every Ad Impression Meaningful at Scale with Machine Learning

Tianshi Gao/ Engineering Manager

Agenda

• ads ranking overview
• ads ranking ML problems and challenges
• scale and accelerate ML development

overview

• 2 billlion people to 6 million advertisers
• make every ad impression meaningful
• how: machine learning

Ad Ranking ML problems

• event predition
  • probability of cliking like
• value prediction
  • guess of how much money people to spend
• multiple event predictiono
  • P(view), P(click), P(add_to_cart), P(check_out)
• recommender system for dynamic ads
• reinforcement learning
  • ads ai, like or buy
  • try to learn some policy to maximize the value
• content and sequence understanding
  • oba(online behavior advertising)

ML challenges

• a data centric view,
• feature/ label
• heterogeonous features -> hetrogenours models and optimizers, joint training
• sparsity, big collection of small data
• bias
• delayed feedback
• non-stationary, seasonality

scale and accelerate ML development

• area under the curve and ML dev cycle
• hypothesis -> experiment - observation -> hypothesis

Modeling Platform

• IP[y]: notebook reader = createsomereader model.fc(100,20) # define model architecture start_some_training_workflow(model, reader) caffe 2

Auto ML

• innovations / repeated manual work
• innovations / repeated manual work + machine

4. Understaing content and understanding in newsfeed

Shiling Ding

connect content and user

• feed ranking machine -> rank stories and score
• latest stories to the top of the feed

People problem

• problem: stories a person cares about might not be shared with them
• solution: recommendation can be the primary surface for people

Products

• explore feed
• suggested videos

Recommendation: why hard?

• scale
  • 1 billion posts per day, 2 billiion MAU
• personalization
  • content: text, image, video
  • user interests
• solution
  • multi-stage approaches: retrieval and ranking
  • content and user understanding

How

• entity extracting and linking
  • human readable entities, eg elon musk, tesla, golden state warriors
  • taxonomy
• classification/ embedding
  • collaboration with applied machine learning team
  • deep text for text understanding
    • docnn/regnn
• end-to-end approach
  • pull all raw features into dl models
  • joint optimzie for the final ranking goal
• user profile
  • user interests by a list of entities with scores
  • user-text embediding
• user modeling
  • RNN: use session and contextural signals to predict user short term interest
  • RNN - applied on snapshot (short time)
  • Query 1, history sequence,
• scale
  • retrieve ccontent relevant to user interest

Entity based recommendation

• form query from user profile
• extended user interests to similar *filament = embedding methods
• learn from graph edges

[users] - engage - [content] - title and description - [words]

• subscribe - [page] - create - [content]

• similariy searching using embedding
• https://code.facebook.com/posts/

DL methods

• objective
• minimize the distance between user and storiy if user engaged the story
• neural nets translate features into embeddings

summary

• recommendation is challenging
  • connect use to the right contents
  • exploration
  • global optimization
  • reinforcement learning

5.Unlocking Meaning Across Language

AML/ LAnguage Technology and Translation team

Facebook is Multilingual

• 101 languages
• idenifiy 150 languages
• 1/2 no english

Building Multilingual Community

• vision: facebook for everyone in their preferce

45+ languages supported for translations 2K+ 600M_ daily active people 1.3B every user

Migration to Neural MT

AI in a multilingural

developing NLP Applications

• english data -> english classifier
• russian data -> russian classifier
• english > russian > japanese > german (languages of the internet, wikipedia)

Cross-lingual understanding(XLU)

• feed many languages data -> multi-lingual classifier
• universal representation of text
• universal word embedding -> learn embedding for each langues separately
• project onto common space
• require a directory

XLU through universal word embeddings

• train a classifier using the universal word embedding

XLU accuray

• sale post classification, 79~96% accuracy

XLU vs translation

XLU at Facebook

• integrated in our text understading platforms
  • visit our platforms poster

Going beyond words

• capture universal semantics at the sequence level

universal sentence representation

• learn an encoder that maps each sentence into a vector

6.Audio/Speech

Reena

Mission

build audio and speech tch to drive at scale at facebook

Video Captioning

2016 first, editors can review and edit

MZ harvard speech - captioning

• Video views per QTR day

Voice Interfaces

• oculus

Video Understanding - Beyond Speech

• language identification
• semi-supervised traning
• traning on 50k hours of audio data
• models are lightweight so we can run them piror to ASR

Audio-Visual Dataset for Videos

• why?
  • audio-visual modeling required to detect animal cruelty
  • public datasets available, but limited and not on FB data
  • labels across different modalitieis
  • visual: scenes, objects, actions
  • audio: auduio events
  • audio labels compatible with audioset
  • we are planning to release it to the public to support research on joint modeling

Posters

• caffe2
• Lumos