Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Machine Learning For The Curious But Confused

B27e5bc114b24f86625025d4dae10184?s=47 ellenkoenig
November 21, 2018

Machine Learning For The Curious But Confused

Presented at Codemotion Berlin 2018

B27e5bc114b24f86625025d4dae10184?s=128

ellenkoenig

November 21, 2018
Tweet

Transcript

  1. MACHINE LEARNING FOR THE CURIOUS BUT CONFUSED ELLEN KÖNIG //

    @ELLEN_KOENIG Senior Data Scientist @ Native Instruments
  2. OVERWHELMED? BUZZWORD BINGO!

  3. SO, EXACTLY WHAT DOES IT MEAN WHEN A MACHINE „LEARNS“?

  4. WHAT IS LEARNING? BEING ABLE TO DEAL WITH NEW SITUATIONS

    BASED ON THE PAST
  5. EXAMPLES WHAT CAN YOU DO WITH MACHINE LEARNING? Self-driving cars

    Price prediction Gene sequence identification
  6. OF HUMANS AND MACHINES WHAT HAPPENS DURING LEARNING? DATA MACHINE

    LEARNING ALGORITHM MODEL FUNCTION Input about the world Processing resources Learned representation „DOG“ Neural association Eyes + brain Outside world
  7. MACHINE FRIENDLY REPRESENTATIONS OF EXPERIENCE HOW DO WE PUT THE

    OUTSIDE WORLD INTO A MACHINE? Input about the world 1 person, 2 trees, 1 animal, lots of grass, 1 path Different grayscale pixels Extracted relevant information People Trees Animals Grass Paths 1 2 1 Yes 1 Numerical representation ( 12 1 1 1 ) Data vector representation Describe or capture Remove context Summarize with numbers
  8. MACHINE FRIENDLY REPRESENTATIONS OF LEARNINGS SO WHAT EXACTLY DO MACHINES

    LEARN? A function is a relation between a set of inputs and a set of permissible outputs with the property that each input is related to exactly one output. (Wikipedia) f ( ) = 1 MACHINES LEARN PREVIOUSLY UNKNOWN FUNCTIONS MAPPING FROM GIVEN INPUT TO GIVEN RESULTS MODEL FUNCTION f ( ) = 0 f (x) = ?
  9. WHAT DOES THAT LOOK LIKE IN PRACTICE? EXAMPLES Example Input

    data Learned Model Terrain data (slope, roughness, etc.) Function mapping terrain to speed Customer & market data and past prices Function mapping input to future prices Gene sequence identificatio Lots and lots of genome data Clusters of re-occuring gene sequence patterns
  10. COMPONENTS OF A COMPLETE MACHINE LEARNING SYSTEM WHAT DOES A

    MACHINE NEED TO LEARN? INPUT DATA ML ALGORITHM MODEL FUNCTION RESULT Unsupervised Learning
  11. COMPONENTS OF A COMPLETE MACHINE LEARNING SYSTEM WHAT DOES A

    MACHINE NEED TO LEARN? TRAINING DATA INPUT DATA ML ALGORITHM MODEL FUNCTION RESULT Supervised Learning
  12. COMPONENTS OF A COMPLETE MACHINE LEARNING SYSTEM WHAT DOES A

    MACHINE NEED TO LEARN? TRAINING DATA INPUT DATA ML ALGORITHM MODEL FUNCTION RESULT FEEDBACK Reinforcement Learning
  13. THE TWO BASIC KINDS OF MACHINE LEARNING SUPERVISED VS UNSUPERVISED

    LEARNING User tastes User 1 likes The Clash User 23 likes Die Ärzte User 42 likes Helene Fischer User 1 likes The Sex Pistols User 42 likes Heino Rain Wind Umbrella? heavy light yes none light no light strong no light light yes none strong no Supervised Unsupervised
  14. LINEAR REGRESSION A SIMPLE SUPERVISED LEARNING ALGORITHM Fitting a line:

    https://towardsdatascience.com/linear-regression-using-least-squares-a4c3456e8570 PRICE IN 10,000 € SIZE IN SQ. METERS Apartment Price Prediction
  15. K-MEANS A SIMPLE UNSUPERVISED LEARNING ALGORITHM Car Model Clustering WEIGHT

    SPEED
  16. SUMMARY CORE INTUITIONS FOR MACHINE LEARNING ▸ Machine learning works

    in a very similar way to human learning! ▸ Learning: Pattern recognition, dealing with unfamiliar situations based on experience ▸ Situations and experience can be abstracted into data to be accessible to machines ▸ Machines learn previously unknown functions from data ▸ A ML system consists of input data, ML algorithms, model functions, results and optionally feedback and training data
  17. WHERE TO GET STARTED RECOMMENDED RESOURCES FOR BEGINNERS (IN ORDER

    OF RECOMMENDATION) ▸ Tutorial for the “Kaggle Titanic Competition” (using R): http://trevorstephens.com/post/72916401642/titanic-getting-started-with-r ▸ More advanced Tutorial based on the same dataset using Python (Scikit-learn, Pandas, Tensorflow): https://blog.socialcops.com/ technology/data-science/machine-learning-python/ ▸ Online courses (MOOCs): ▸ Udacity: Intro to Machine Learning: https://www.udacity.com/course/intro-to-machine-learning--ud120 (Excellent intro to applied ML using sci-kit learn and Python) ▸ Coursera: Machine Learning: https://www.coursera.org/learn/machine-learning (Friendly intro to the theory behind common ML algorithm) ▸ Machine Learning Mastery: Lots of self-study guides for ML learners http://machinelearningmastery.com/ ▸ UCI ML Repository: Collection of “Toy problems” for ML http://archive.ics.uci.edu/ml/datasets.html ▸ Toolkits: ▸ Scikit-Learn (Python, great online documentation): http://scikit-learn.org/stable/ ▸ stats package (many simple ML algorithms), pre-installed (R) Examples: http://www.statmethods.net/stats/regression.html ▸ Book: Abu-Mostafa, Magdon-Ismail, Lin: Learning From Data - A Short Course (AMLbook.com ) (Good intro to more academic perspectives, notation and vocabulary on ML)
  18. BONUS: SO, HOW CAN I GET STARTED IN TEACHING A

    MACHINE TO LEARN?
  19. THE STARTING POINT A BASIC WORKFLOW FOR WORKING ON MACHINE

    LEARNING PROBLEMS 1. Understand the problem and context 2. Understand & clean the data, create some features 3. For supervised learning: Split into training and test data 4. Evaluate different algorithms with default parameters 5. Optimize the parameters and compute the results 6. Interpret the results 7. Repeat with different features until you get useful results
  20. THE STARTING POINT MORE ART THAN SCIENCE

  21. LEARN BY REPEATING THE WORKFLOW RINSE AND REPEAT PICK ONE

    TOOL TRY THE WORKFLOW PICK A (“TOY”) PROBLEM PICK A TYPE OF ALGORITHM
  22. LICENSE: CREATIVE COMMONS “ATTRIBUTION - SHARE ALIKE” 4.0 HTTPS:// CREATIVECOMMONS.ORG/LICENSES/BY-SA/4.0/

    IMAGE CREDITS ▸ Slide 1 : http://work.caltech.edu/dex1.html at 5:20 of the video ▸ Slide 3 & 18: http://www.thebluediamondgallery.com/highlighted/l/learning.html ▸ Slide 3: All https://pixabay.com/ ▸ Slide 4 & 8: ▸ https://commons.wikimedia.org/wiki/File:Human_genome.png ▸ https://commons.wikimedia.org/wiki/File:Hubbert_US_Lower_48_Gas_Prediction_-_1962.png ▸ https://commons.wikimedia.org/wiki/File:Waymo_self-driving_car_front_view.gk.jpg ▸ Slides 5 -7: https://en.wikipedia.org/wiki/Consciousness#/media/File:Neural_Correlates_Of_Consciousness.jpg ▸ Slide 7: pixabay.com ▸ Slide 9-11: Based on https://commons.wikimedia.org/wiki/File:Machine_Learning_Technique..JPG ▸ Slide 13: ▸ https://commons.wikimedia.org/w/index.php?curid=11967659 ▸ https://commons.wikimedia.org/wiki/File:Residuals_for_Linear_Regression_Fit.png ▸ Slide 14: Based on https://commons.wikimedia.org/wiki/File:Kmeans_animation_withoutWatermark.gif ▸ Slide 20: https.//pixabay.com