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

Dreams, Drugs and ConvNets

Dreams, Drugs and ConvNets

Artefacts of human and artificial cognition.
at Studenckie Koło Naukowe Neurobiologii, the University of Warsaw

Piotr Migdał

March 02, 2017
Tweet

More Decks by Piotr Migdał

Other Decks in Science

Transcript

  1. Dreams, Drugs
    & ConvNets
    Piotr Migdał, PhD
    deepsense.io / freelancer
    http://p.migdal.pl/
    2 Mar 2017,
    Studenckie Koło Naukowe Neurobiologii UW

    View full-size slide

  2. Deep Learning progress
    • image recognition, neural style, word analogies,
    per-char translations, playing ATARI games, Go,
    [no idea what’s next]
    • fast-paced (more than my quantum physics PhD):

    6 month ago a breakthrough, now a baseline
    • (no questions about Singularity please!)

    View full-size slide

  3. Artificial neural networks
    • Multidimensional arrays
    • Simple matrix operations (add, multiply)
    • Floats (not spikes)
    • Simple activation functions (sigmoid, ReLU)
    • Cost function and back-propagation
    • A lot of weights (e.g. 100M)

    View full-size slide

  4. Many layers
    https://research.facebook.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/

    View full-size slide

  5. Convolutions
    http://setosa.io/ev/image-kernels/

    View full-size slide

  6. Patterns activating channels
    https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html

    View full-size slide

  7. Big questions
    • creating general AI or sentient beings
    • transferring our minds (mind uploading)
    “I mean, if we're able to save even just a small piece of
    ourselves, why wouldn't we do that?” - SOMA
    https://www.youtube.com/watch?v=BZTfi1jv-EE

    View full-size slide

  8. Flesh vs machines:

    practical questions
    • Cross-inspiration:
    • Machine Learning to Human Cognition
    • Human Cognition to Machine Learning
    • Common abstractions
    • Learning in general
    • Hierarchical signal processing

    View full-size slide

  9. NSFW, 18+, etc

    View full-size slide

  10. http://www.techrepublic.com/article/why-microsofts-tay-ai-bot-went-wrong/
    Tay.AI (+ 4chan)

    View full-size slide

  11. The biggest lie of every

    parent, pet owner

    and neural network trainer:
    “I’ve never shown it that,
    it must have learned it by itself!”

    View full-size slide

  12. Stereotypes
    http://p.migdal.pl/2017/01/06/king-man-woman-queen-why.html

    View full-size slide

  13. If you’ve never had a trip…
    https://youtu.be/cPKq7JuQDvg?t=256

    View full-size slide

  14. If you’ve never had a trip…
    https://youtu.be/z7_U6y8kJaY?t=72

    View full-size slide

  15. Deep dreams: forcing ConvNet

    to see things
    https://research.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html

    View full-size slide

  16. • What should I use:

    THC, LSD, DMT?
    • A good GPU!
    http://alexgrey.com/shop/st-al.html

    View full-size slide

  17. Grocery trip
    https://www.youtube.com/watch?v=DgPaCWJL7XI

    View full-size slide

  18. Forcing networks
    to see things
    https://research.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html

    View full-size slide

  19. Neural style transfer
    also video: https://www.youtube.com/watch?v=Khuj4ASldmU

    View full-size slide

  20. With own photos
    to play with https://deepart.io/ or https://prisma-ai.com/

    View full-size slide

  21. What is sexual?
    http://blog.clarifai.com/what-convolutional-neural-networks-see-at-when-they-see-nudity/#.WLhGmRIrLLh

    View full-size slide

  22. What is sensual?
    http://blog.clarifai.com/what-convolutional-neural-networks-see-at-when-they-see-nudity/#.WLhGmRIrLLh

    View full-size slide

  23. And this?
    (yes, ConvNet is also fooled)

    View full-size slide

  24. Sketch to picture
    http://affinelayer.com/pixsrv/

    View full-size slide

  25. Sketch to picture:
    let’s get rogue!
    http://affinelayer.com/pixsrv/

    View full-size slide

  26. https://raw.githubusercontent.com/Newmu/dcgan_code/master/images/faces_arithmetic_collage.png
    https://github.com/255BITS/HyperGAN
    Picture analogies

    View full-size slide

  27. https://open_nsfw.gitlab.io/
    Image generation

    View full-size slide

  28. Learning what is important
    • Object consistency
    • Toddler needs to learn, so do ConvNets
    http://blogs.scientificamerican.com/illusion-chasers/what-little-babies-see-that-you-no-longer-can/

    View full-size slide

  29. Uncertainty and learning
    http://www.nature.com/articles/ncomms11609

    View full-size slide

  30. Retrain your own network:

    McCollough effect
    http://www.michaelbach.de/ot/col-McCollough/

    View full-size slide

  31. Trypophobia

    (don’t google it!)
    visit here instead: https://en.wikipedia.org/wiki/Trypophobia

    View full-size slide

  32. Trypophobia detector
    https://github.com/grzegorz225/trypophobia-detector
    (a 5-day workshop was enough for 88% accuracy!)

    View full-size slide

  33. And if you

    don’t want to code…
    http://www.sciencemag.org/news/2015/11/pigeons-spot-cancer-well-human-experts
    https://www.youtube.com/watch?v=flzGjnJLyS0

    View full-size slide

  34. Questions?
    • If you want to learn:
    • http://lumiverse.io/series/neural-networks-demystified
    • https://gist.github.com/stared/7de2908b9bcba01c39ee3c591875a23c
    • http://www.deeplearningbook.org/
    • Keras (on top of TensorFlow) in Jupyter Notebook
    • https://medium.com/@mateuszsieniawski/keras-with-gpu-on-amazon-ec2-a-step-by-
    step-instruction-4f90364e49ac#.4i1norivh
    • https://www.erowid.org/
    http://p.migdal.pl/ + I do like emails: [email protected]

    View full-size slide