Machine Learning What, how, why? - GitHub Pages

Machine LearningWhat, how, why?

Rémi Emonet (@remiemonet)2015-09-30

Web En Vert

$ whoami

IBM BlueMix (Watson)

$ whoami

Software EngineerResearcher: machine learning, computer visionTeacher: web technologies, computing literacyGeek: deck.js slides, isochrones, …

You are shrewd, skeptical and restrained.You are independent: you have a strong desire to have time to yourself. You arecalm-seeking: you prefer activities that are quiet, calm, and safe. And you arephilosophical: you are open to and intrigued by new ideas and love to explorethem.Experiences that give a sense of prestige hold some appeal to you.You are relatively unconcerned with both tradition and taking pleasure in life.You care more about making your own path than following what others havedone. And you prefer activities with a purpose greater than just personalenjoyment.

4 / 50 − Rémi Emonet (@remiemonet) − Machine Learning What, how, why?

What is Machine Learning?


Machine Learning Basic PrincipleGiven a dataset {y , x , … , x }

Optimize the likelihood functionL = n(w, t , d)log p(w, t ∣z)p(z, t ∣d)

Or using a sparse regularizationL − λ KL(U ∣∣p(ts∣z, d))

By using a Gibbs Samplerp(W , at ∣o = o, O ) =

i i1 ip i=1n

d

∑w

∑ta

∑ a

z

∑ts

∑ r s

sparse

d

∑z

∑

ji ji ji−ji

N (w , rt , z ) + η(w , rt )∑w ,rt′ ′ ( obs−ji ′ ′

ji′ ′ )

N (W , rt , z ) + η(W , rt )obs−ji

ji ji ji ji ji


Machine Learningin the Wild

Which One of These ServicesUses Machine Learning?



Machine Learning in Future Tech?


What is Machine Learning?an example motivation

Challenge: Which Iris Species?


⇒Sepal length: 5.1Sepal width: 2.5Petal length: 4.2Petal width: 1.0

Expected Label: “Iris Setosa”

Feature Extraction

Analysis and Program Writing

IFTTT...


Predictive Machine LearningInstead of writing a program that solves a task,We

collect labeled data: input/output pairs1.

automatically generate a programthat computes an output for each new input

2.

profit!3.

The machine learns to generalizefrom a limited number of examples,

like humans do.


Different Types of TasksSupervised learning: some labels are known

classification: find the label of an exampleregression: find the target value

Unsupervised learning: no labelsclustering: group things togetherpattern mining: find recurrent eventsanomaly detection: find “outliers”


The Principle of “Overfitting”


A Lot of Different MethodsAlso called “models”

linear regression, logistic regression, SVM, kernel SVM, neural networks, k-meansclustering, collaborative filtering, bayesian networks, expectation maximization,belief propagation, multiple kernel learning, metric learning, transfer learning,decision trees, gaussian processes, random forests, boosting, ...

For different contextsdifferent tasksdifferent nature of datadifferent suppositions on the datadifferent amount of data


Different Ways to StartUse a product that uses ML

e.g. adwords, ibm bluemix, …

Use a prediction APIsend your data to the serviceget API to process new inputse.g., google pred. API, prediction.io, ...

Dive into machine learning…


Into Machine LearningUsing libraries

libraries exist in most languagesmost models already implementedtest different methods with different parameters

Learning machine learningmany online coursesget deeper understanding


Does Machine Learning ActuallyMatter?


Example: The Netflix Challenge

FAIR



Example: Facebook AI ResearchDirector: Yann LeCunScientific Leads

Léon BouttouRob FergusFlorent Perronnin

fr rest


Data, Data, Data


Data isMachine Learning's Fuel

data === power

Getting Data?


Getting Data?Collect from your services/applicationsDo it yourselfPay some people you knowUse crowd-sourcing,e.g., Amazon Mechanical Turk (MTurk)Find existing datasets (open data, etc)Work for/with a “data rich” companyCreate your “intermediation” business


What Can It DoFor Me


SearchGoogle Search, Bing, etc


AdvertisingAdWords, etc


RecommendationsNetflix, Amazon, Youtube,

app Stores, etc


Text Translation


Optical Character Recognition(postcodes, checks, book scans,

etc)


Visual Recognition (objects,plants, animals, etc)


Face DetectionSmile Detection

(embedded in Cameras)


Face Identification(Picasa, Facebook, etc)


Kinect Controller


Self Driving Cars


Voice Recognitionand Synthesis

(GoogleNow, Siri, Cortana)


Sound Recognition(birds, underwater sounds,

safety, etc)


Fraud Detection(Banking, Websites, etc)


Automated Trading…


Customer/Person ProfilingBlueMix Watson, etc


Adaptive Websites(automated A/B testing)


The “Big Data” Hype

and much more...


Where Will it Stop?

Singularity?

Thanks! Questions?

twitter: @remiemonetweb/email: http://home.heeere.comRecommended Links:

comprehensive introduction to ML modelsscikit learn (python)... 50 / 50 − Rémi Emonet (@remiemonet) − Machine Learning What, how, why?

tatadbb

-Porsupah-

ali eminov

efilpera

jannekestaaks

Asa-moya

francoisjouffroy

JD Hancock

Machine Learning What, how, why? - GitHub Pages

Documents

Transcript of Machine Learning What, how, why? - GitHub Pages