Download to read offline
Day 3 take up convolutional neural network
- 7. ICEBREAKER Phạm Nguyễn AnhThư @thu.phamhcmut Let’s get to know each other!
- 8. Record shop _____ _______ Thật bất ngờ ___ ____ ____ Em dạo này Tình đắng như ly cà phê ____ _____ ____ _ __ ____ ___ ___ ___ ___ ___ ______ Để Mị nói cho mà nghe ____ ____ ____ ___ __ ____ ____ Đâu cần một bài ca tình yêu
- 9. how-to-AI Series: UnlockPotential Day 3: Take UP Convolutional Neural Network Nguyễn Luật Gia Khôi @giakhoi.nguyenluat Nguyễn Thế Bình @binh.nguyen288
- 10. Outline 1. Break theice 2. Convolutional Neural Network 3. Why using CNN instead of NN? 4. Demo code 5. Kaggle Challenge
- 11.
- 12. 1 -1 -1-1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 -1 -1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 -1 -1 1 1 ● -1: black ● 1: white Matrix representation
- 13. 1 -1 -1-1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 -1 -1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 -1 -1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 1 1 1 -1 1 1 1 -1 1 -1 1 1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 -1 1 1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 ANN can solve this variety in digits
- 14. 1 -1 -1-1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 -1 -1 1 1 1 -1 -1 -1 1 1 -1 ... 1 -1 1 1 -1 -1 1 1 x2 x33 x32 x3 x4 x34 x1 x35 ... h2 x33 x32 h3 h4 x34 h1 x35 ... h2 h8 h3 h9 h1 h10 ... 1 9 8 2 3 0 0.01 0.92 0.003 0.008 0.015 0.02
- 15. WAIT A SECOND,how about this?
- 16. Image size =1920 x 1080 x 3 Input layer #neurons = 1920 x 1080 x 3 = 6 million Hidden layer #neurons = Say you keep it = 4 million Weights between input and first hidden layer = 6 * 4 = 24 trillion parameters
- 18. Disadvantages of ANNin image classification - Computationally expensive. - Treat local pixels same as pixels far apart => no locality concern - Sensitive to location of the object in the image.
- 19. How does humanrecognize images?
- 20.
- 21.
- 22. How can wemake computers recognize these patterns?
- 23. 1 -1 -1-1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 1 -1 -1 -1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 Circular pattern filter Diagonal line filter Diagonal line filter
- 24. 1 -1 -1-1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 -1 -1 -1 -1 1 -1 -1 -1 -1 * = -1 9 -1 -5 1 -3 -3 3 -3 -5 -1 -5 -3 -5 -3 Circular pattern filter -1 x -1 + -1 x -1 + … + -1 x -1 + 1 x 1 = 9
- 25. x -1 -1 -1 -11 -1 -1 -1 -1 * Circular pattern filter = x -1 -1 -1 -1 1 -1 -1 -1 -1 * Circular pattern filter = x x -1 -1 -1 -1 1 -1 -1 -1 -1 * Circular pattern filter =
- 26.
- 27. x x * Eye filter = x * Nosefilter = x x * Ear filter = * Head filter x = x x
- 28. x x X ... X ... h2 x8 h3 x9 h1 x10 ... 0.01 (Giraffe) 0.80 (Koala) 0.003(Horse) 0.008 (Buffalo) 0.15 (Bear) 0.02 (Squirrel)
- 29.
- 30. 1 -1 -1-1 1 1 -1 1 -1 1 1 -1 -1 -1 1 1 1 1 -1 1 1 1 1 -1 1 1 1 -1 1 1 1 -1 1 1 1 -1 -1 -1 -1 1 -1 -1 -1 -1 * = -1 9 -1 -5 1 -3 -3 3 -3 -5 -1 -5 -3 -5 -3 Circular pattern filter ReLU 0 9 0 0 1 0 0 3 0 0 0 0 0 0 0
- 31. Stride & Padding stride= 2, padding = “valid” stride = 1, padding = “same” input output input output
- 33. 0 9 00 0 1 0 0 0 3 0 0 0 0 0 0 0 0 0 0 Max pooling pool_size = (2,2) stride = 2 9 0 3 0 output_shape = math.floor((input_shape - pool_size) / strides) + 1
- 34. Benefits of maxpooling - Reduce feature map size => Reduce computational cost. - Fewer parameters needed => Reduce overfitting. - Make model tolerant towards variations and distortions.
- 35. Complete pipeline Conv +ReLU Pooling Conv + ReLU Pooling w1 x h1x c1 w2 x h2 x c1 w2 x h2 x c2 w3 x h3 x c2 Flatten
- 36. Why using CNNinstead of NN?
- 37. - Can capturespatial features => Connection sparsity => Reduce overfitting x x * Eye filter = x * Nose filter = x x * Ear filter =
- 38. 0 0 90 0 1 0 0 0 3 0 0 0 0 0 0 0 0 0 0 - Translation invariant. 1 9 3 0 0 0 0 0 0 0 9 0 0 3 0 0 0 0 0 0 0 0 0 0 0 9 3 0 Conv Conv Pooling Pooling
- 39. - Location invariantfeature detection.
- 40.
- 41.
- 42.
- 43.
- 44. ● Challenge: Builda model that can automatically classifies images to their correct labels. ● Registration form is now open! ● Timeline: 12/12/2021 - 30/12/2021 ● Rewards: ○ First prize: 1.000.000 VND ○ Second prize: 700.000 VND ○ Third prize: 500.000 VND ○ 2 AI potential prizes: 400.000 VND GDSC AI CHALLENGE
- 45.
- 46.
- 47.