Finding Swimming Pools with IBM Watson

Finding Swimming Pools with IBM Watson

A hobbyist's introduction to the IBM Watson APIs.

167e3044ed845338afdf5cc3446f5b17?s=128

Tom Robertshaw

June 30, 2016
Tweet

Transcript

  1. Finding Swimming Pools with IBM Watson Tom Robertshaw @bobbyshaw

  2. Disclaimer At most, I’m a hobbyist. I’m a little bit

    in love with IBM Watson
  3. SaaS APIs IBM Watson Google Cloud Machine Learning Amazon Machine

    Learning Microsoft Azure Machine Learning
  4. IBM Watson

  5. IBM Watson According to IBM, Watson uses "more than 100

    different techniques are used to analyze natural language, identify sources, find and generate hypotheses, find and score evidence, and merge and rank hypotheses.
  6. IBM Watson DeepQA software commercially available. Minimum Server Specs: $1

    million
  7. So what can it do?

  8. None
  9. Concept Insights Use case - Recommendation of related documents Input

    - Text - Collection of documents Output - Related documents (based on your collection) - Concepts mentioned in the input document
  10. None
  11. None
  12. Personality Insights Input - JSON, or Text or HTML written

    by one individual Output - A tree of cognitive and social characteristics in JSON or CSV format
  13. None
  14. Eliza

  15. None
  16. Natural Language Classifier Input - Mapping of text to classifications

    Output - Classification
  17. Sorting Hat

  18. None
  19. Tradeoff Analytics Use case - When shopping for a phone,

    a user wants to consider the importance of price, screen size, battery life and weight. Tradeoff analysis is a pareto optimisation of user’s preferences, e.g. fuzzy filters in ecommerce. Input - A decision problem with objectives and options Output - JSON objects that represent the optimal options and highlight the trade- offs between them.
  20. Others - Retrieve and Rank - Conversation - Tone Analyser

    - Relationship Extraction - Alchemy News - Document Conversion - Speech to Text - Text to Speech - Language Translation
  21. None
  22. Visual Recognition Content Categorisation - Took screenshots of websites -

    Training set of ecommerce categories - More accurate than Bayesian text classification
  23. None
  24. Kyero - 100,000s of properties submitted be estate agents or

    come in via an API. - Structured data gives more opportunity to optimise. - Manually - Show swimming pool - Automatically - Optimisation for category of image that increases conversion rate.
  25. None
  26. Kyero - Swimming Pool - Sea - Outside - Inside

    - Building - Blue Sky - Bathroom - Bedroom - Living Room - Dining Room - Kitchen - Photo - Drawing - Logo - Watermarked
  27. Training & Testing What proportion do you use for training?

    (What proportion do you use for validation?) What proportion do you use for testing?
  28. Training & Testing - Many different models. - Balance of

    simplicity and effectiveness is K-fold mechanism, e.g 2-fold - Train on 50%, test on 50% - Swap data around, test again - Take an average.
  29. Create Classifier curl -u "{username}":"{password}" \ -X POST \ -F

    "positive_examples=@swimming-pool-positive-training.zip" \ -F "negative_examples=@swimming-pool-negative-training.zip" \ -F "name=swimming-pool" \ "https://gateway.watsonplatform.net/visual-recognition- beta/api/v2/classifiers?version=2015-12-28"
  30. Test Positive curl -u "{username}":"{password}" \ -X POST \ -F

    "images_file=@swimming-pool-positive-validation.zip" \ -F 'classifier_ids={"classifiers":[{"classifier_id":" swimming_pool_2056231781"}]}' \ "https://gateway.watsonplatform.net/visual-recognition- beta/api/v2/classify?version=2015-12-28"
  31. Test Negative curl -u "{username}":"{password}" \ -X POST \ -F

    "images_file=@swimming-pool-negative-validation.zip" \ -F 'classifier_ids={"classifiers":[{"classifier_id":" swimming_pool_2056231781"}]}' \ "https://gateway.watsonplatform.net/visual-recognition- beta/api/v2/classify?version=2015-12-28"
  32. Kyero Category Accuracy (0.7 threshold) Swimming Pool 0.80 Outside 0.94

    Building 0.71 Blue Sky 0.80 Watermarked 0.33 Inside 0.89 Living Room 1.00 Sea 0.8
  33. None
  34. Potential Comes from combining services

  35. Combine Services

  36. Combine Services

  37. Getting Started Pricing is very attractive for prototyping - Free

    to get started - First ~1000 calls are normally free too - Pricing ranges from $0.0009 to $0.25 per call.
  38. AI XPrize Competition - Use IBM Watson to tackle the

    world’s grand challenges. - $5 million in prizes - $3 million grand prize - http://ai.xprize.org/
  39. Cognify - ML is now much easier to get good

    enough results. - Experiment with some of Watson’s “approachable APIs”. - Build smarter digital products.
  40. Thanks! Questions?