Neural Networks

Fred

Transcript

1. Neural Networks Brief introduction and sample case
2. "Artificial Intelligence"
3. AI in Popular Culture
4. AI Effect "AI is whatever hasn't been done yet." Douglas Hofstadter "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'" Rodney Brooks
5. Types of AI • Machine learning ◦ Neural networks ▪ Perceptron ▪ Recurrent neural network ▪ Convoluted neural network ◦ Support Vector Machines (SVM) • Fuzzy systems • Evolutionary algorithms ◦ Genetic algorithm ◦ Differential evolution • Swarm Intelligence • Probabilistic methods
6. Neural Networks
7. What is this ?
8. Perceptrons
9. Perceptrons
10. Perceptrons • x 1 – Is it raining? • x 2 – Does your girlfriend/boyfriend want to go? • x 3 – Is it near public transportation?
11. Perceptrons
12. Perceptrons
13. Perceptrons
14. Perceptrons
15. Perceptrons
16. Perceptrons
17. Neurons
18. Perceptrons
19. Neurons
20. Why is this relevant?
21. Learning!
22. Learning
23. Learning
24. Learning
25. Real world problem
26. Recognizing Handwritten Digits
27. Database The MNIST (Modified National Institute of Standards and Technology) database Contains 60,000 training images and 10,000 testing images.
28. Database
29. Neural Network Architecture
30. Neural Network Architecture
31. Neural Network Architecture
32. Neural Network Architecture
33. Neural Network Architecture
34. Measuring Outcome
35. Neural Network Architecture
36. Cost Function
37. Learning
38. Gradient Descent
39. Gradient Descent
40. Gradient Descent
41. Gradient Descent
42. Gradient Descent
43. Gradient Descent
44. Backpropagation
45. Backpropagation
46. Let's CODE!
47. Repository
48. __init__()
49. feedforward()
50. sigmoid()
51. sigmoid()
52. Example use
53. Training Our Network
54. Learning
55. SGD()
56. update_mini_batch()
57. backprop()
58. cost_derivative()
59. Performance
60. 95%
61. Current Record The current (2013) record is classifying 9,979 of 10,000 images correctly (99,79%).
62. How to improve?
63. Improving our Neural Network • Improved Cost Function ◦ Faster Training ◦ Reduced Overfitting • Regularization • Dropout • Artificially expanding training data • Convolutional Networks
64. Thank you!
65. Acknowledgement Michael Nielsen NeuralNetworksAndDeepLearning.com
66. Do we have time?
67. Convolutional Neural Networks
68. Regular Networks
69. Convolutional Neural Networks
70. Convolutional Neural Networks
71. Convolutional Neural Networks
72. Convolutional Neural Networks
73. Feature Maps (Kernels)
74. Pooling Layers
75. Max-Pooling
76. Architecture