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

Over Mt. Stupid to Deep Learning

Over Mt. Stupid to Deep Learning

My slides for my talk at Code.Talks 2018 about MAchine Learning, Deep Learning and my journey through this topics. Video of talk may come later.

Carsten Sandtner

October 18, 2018
Tweet

More Decks by Carsten Sandtner

Other Decks in Technology

Transcript

  1. OVER MT. STUPID TO DEEP LEARNING… Carsten Sandtner \\ @casarock

    Photo by Archie Binamira from Pexels https://www.pexels.com/photo/man-wearing-white-shirt-brown-shorts-and-green-backpack-standing-on-hill-672358/
  2. ABOUT://ME My name is Carsten and I’m Technical Director at

    mediaman GmbH in Mayence (Mainz). I’m a moz://a Techspeaker and I love the open web and everything about open standards for the web! I’m tweeting as @casarock 
 
 … and I asked myself: What is this AI thing?
  3. „ “ –Amy Webb - SXSW 2018 The robots are

    going to come and kill us all but not before they take over all of our jobs
  4. „ “ –Amy Webb - SXSW 2018 The artificial intelligence

    ecosystem —  flooded with capital, hungry for commercial applications, and yet polluted with widespread, misplaced optimism and fear — will continue to swell
  5. HISTORY 1950: Neuronal Networks! ~1980: Machine Learning Today: Deep Learning

    Photo by Dick Thomas Johnson https://www.flickr.com/photos/31029865@N06/14810867549/
  6. MACHINE LEARNING Statistics! Correlation Regression a => weights Const =>

    Offset/Bias Y = Const + aX1 + bX2 + cX3 + ... + zXn
  7. ACTIVATION FUNCTIONS 0 2 4 6 -4 -6 -2 0.0

    0.2 0.4 0.6 0.8 1.0 0 2 4 6 -4 -6 -2 0 1 2 3 4 5 Sigmoid ReLu R(z) = max(0, z) sig(z) = 1/(1+e-z) Rectifier Linear Unit
  8. SIMPLE EXAMPLE Machine Sensors x1, x2 and x3
 Weights a,

    b and c Y = Const + ax1 + bx2 + cx3 Const: Value when x1, x2 and x3 are 0 Training data: 100.000 Datasets. Keep 25.000 for validation Train with 75.000 -> Vary weights until result (Y) is ok Verify your model with the 25.000 sets for validation
  9. SUPERVISED Useful for predictions and classifications Popular use case: 


    Image recognition Needs classified training sets.
  10. REINFORCEMENT has not a defined result for training data. Using

    rewards for good results - if it isn’t good do it never again, bad boy! Example: Learn how to play a game just while analyse every pixel Popular Example: 
 Alpha Go
  11. MY JOURNEY! Confidence Wisdom Mt. Stupid Valley of despair Slope

    of enlightenment Plateau of sustainability
  12. ME. Confidence Wisdom Mt. Stupid Valley of despair Slope of

    enlightenment Plateau of sustainability
  13. NEURON Neuron Activation Function Output Input Weight Offset z =

    ∑wixi w0 x0 +b y = f(z) y0 x1 xn w1 wn y1 yn LAYER
  14. . . . . Input Layer Output Layer Hidden Layers

    Forward pass Init with random weights and offsets
  15. . . . . Input Layer Output Layer Hidden Layers

    Back propagation Propagate the error into every neuron err err err err err err err err err err err
  16. BACKPROPAGATION For every neuron from output to input Check neurons

    share in the error (using weights) Fine tune weights (in very small steps eg. 0.001) This is a Hyperparameter: Learning Rate! Applies to the whole network Every weight adjusted? Start a new run This is called an new Epoch (next Hyperparameter!)
  17. CONVERGENCE Always take a look at your error. Is it

    minimal? Stop learning! Validate your model with other data (not the learning data!) Try to avoid overfitting!
  18. ME. Confidence Wisdom Mt. Stupid Valley of despair Slope of

    enlightenment Plateau of sustainability
  19. CONVOLUTIONAL NEURAL NETWORK Conv ReLu Pool Conv Conv Conv Conv

    Conv ReLu ReLu ReLu ReLu ReLu Pool Pool Dog Cat Mouse FC
  20. ME. Confidence Wisdom Mt. Stupid Valley of despair Slope of

    enlightenment Plateau of sustainability
  21. THOUGHT ABOUT MY GOALS Confidence Wisdom Mt. Stupid Valley of

    despair Slope of enlightenment Plateau of sustainability
  22. SOME BETTER EXAMPLES Categorization of products in an online shop

    using their images. Using cognitive services for Natural Language Processing (NLP) voice or text based Using a cloud based AI Service!