Deep Learning Lightning Talk

Deep Learning Lightning Talk

052bbdd81d5dfe4fac23ad4544ad2c0d?s=128

Mateusz Buśkiewicz

June 06, 2014
Tweet

Transcript

  1. Deep Learning Lightning Talk 1 Deep Learning State-of-the-art, powerful way

    to do machine learning Keynote Template
  2. Deep Learning Lightning Talk 2 Our goal: Build awesome data

    products User activity, business data, images, text, audio… Big Data technologies: Hadoop, Spark, Hive, Pig, Storm, Impala… Extract meaning from data and incorporate it into product: predicion, analytics, recommendations ! Technology ! Data ! Machine Learning
  3. Deep Learning Lightning Talk 3 ! Machine Learning Machine Learning

    There are tons of different machine learning algorithms for different problems Unsupervised methods: Customer segmentation (clustering), dataset visualisation, dimensionality reduction ⊞ Supervised methods: Predictions, classifications. Example product: spam filtering ⊞ Recommendations, anomaly detection… ! ⊞
  4. Deep Learning Lightning Talk 4 Difficult problems Image Recognition "

    Speech Recognition ♫ Natural Language Processing $
  5. Deep Learning Lightning Talk 5 Breakthrough: Deep Learning And now

    it works ;)
  6. Deep Learning Lightning Talk 6 Examples ! Skype Translator !

    Google+ Photo Tagging http://www.youtube.com/watch?v=eu9kMIeS0wQ + Voice recognition in Android 4.0+, Apple’s Siri, Baidu’s Image Search, and more…
  7. Deep Learning Lightning Talk 7 A bit of theory Deep

    Learning is a „bigger and badder” approach to neural networks, which are known since 80’ y= g(x ⊗ W)
  8. Deep Learning Lightning Talk 7 A bit of theory Deep

    Learning is a „bigger and badder” approach to neural networks, which are known since 80’ y= g(x ⊗ W)
  9. Deep Learning Lightning Talk 7 A bit of theory Deep

    Learning is a „bigger and badder” approach to neural networks, which are known since 80’ y= g(x ⊗ W)
  10. Deep Learning Lightning Talk 7 A bit of theory Deep

    Learning is a „bigger and badder” approach to neural networks, which are known since 80’ y= g(x ⊗ W) Now we have much more computing power to train large (and deep) networks ⊞ Now we know better regularization and optimization methods ⊞ Now we have much more labeled data ⊞ Now we can also train models with unlabeled data ⊞
  11. Deep Learning Lightning Talk 8 Why it works? Let’s consider

    the problem of face recognition That’s how we see it 0.2 0.0 0.1 1.0 1.0 0.1 0.4 0.8 1.0 ... 0.1 That’s how „machine” sees it
  12. Deep Learning Lightning Talk 9 Why it works? It’s much

    easier to infer that something is a face based on that it has two eyes and nose, than it has some black pixels in lower left corner, and white area somewhere in the middle
  13. Deep Learning Lightning Talk 10 A bit of practice GPU

    Cluster
  14. Deep Learning Lightning Talk 11 A bit of practice GPU

    Numerical operations are very efficient, up to 100x faster than CPU ⊞ Single machine, no communication overhead ⊞ Significant memory contraints, we can’t train larger models ⊟
  15. Deep Learning Lightning Talk 12 A bit of practice TASK

    one learning task, many workers different parameters for each worker PICK BEST MODEL Netflix style! GPU WORKER 1 WORKER 2 WORKER 3 WORKER 4
  16. Deep Learning Lightning Talk 13 A bit of practice Cluster

  17. Deep Learning Lightning Talk 13 A bit of practice Cluster

    WORKER 2 WORKER 1 WORKER 3 WORKER 4 + ASYNCHRONOUS PARAMETERS SERVER Google style!
  18. Deep Learning Lightning Talk 14 A bit of practice Cluster

    We can train much larger and more powerful models ⊞ Scalable ⊞ Poor resource utlization, even if we restrict connectivity ⊟ Complicated ⊟
  19. Deep Learning Lightning Talk 15 Hype NETFLIX MOVES INTO DEEP

    LEARNING RESEARCH TO IMPROVE PERSONALIZATION 10 BREAKTHROUGH TECHNOLOGIES 2013 GIGAOM GUIDE TO DEEP LEARNING: WHO’S DOING IT AND WHY IT MATTERS NYU „DEEP LEARNING” PROFESSOR LECUN WILL HEAD FACEBOOK’S NEW ARTIFICIAL INTELLIGENCE LAB Geoffrey Hinton Leading researcher in DL, his startup was acquired by Google Lookflow Deep Learning image startup, acquired by Yahoo DeepMind Deep Learning startup, acquired by Google for 400 mln USD Yan LeCun Leading researcher in DL, hired by Facebook to lead new AI lab.
  20. Deep Learning Lightning Talk 16 Geoffrey Hinton Leading researcher in

    DL, his startup was acquired by Google Lookflow Deep Learning image startup, acquired by Yahoo DeepMind Deep Learning startup, acquired by Google for 400 mln USD Yan LeCun Leading researcher in DL, hired by Facebook to lead new AI lab. Hype
  21. Deep Learning Lightning Talk 17 It’s not a silver bullet

    It’s difficult. Sometimes it’s better to use simpler method. " # Nevertheless, it’s a very powerful technique, has attention of biggest IT companies and brings us closer to real artificial intelligence It requires substantial computing power and memory. Sometimes it’s not feasible to use deep learning models, especially if we have to train them regularly ! It’s kind of `black-box` Sometimes we can’t draw conclusions from learned features
  22. ! :) THANKS $ mateusz.buskiewicz@getbase.com RESOURCES MOOC: Neural Networks for

    Machine Learning & https://www.coursera.org/course/neuralnets DL Tutorials + sample code & http://deeplearning.net/ Google+ Deep Learning Community & https://plus.google.com/u/0/communities/112866381580457264725 Deep Learning Book by Yoshua Bengio (draft) & http://www.iro.umontreal.ca/~bengioy/dlbook/ Deep Learning Libraries & Software & http://deeplearning.net/software_links/