Deep Learning approach in Machine learning
- 1. DEEP LEARNING Supervised Deep Learning Unsupervised Deep Learning Semi supervised Deep Learning
- 2. LECTURE 5 SMALLER NETWORK: CNN We know it is good to learn a small model. From this fully connected model, do we really need all the edges? Can some of these be shared?
- 3. CONSIDER LEARNING AN IMAGE: Some patterns are much smaller than the whole image “beak” detector Can represent a small region with fewer parameters
- 4. PLACES: THEY CAN BE COMPRESSED! WHAT ABOUT TRAINING A LOT OF SUCH “SMALL” DETECTORS AND EACH DETECTOR MUST “MOVE AROUND”. “upper-left beak” detector “middle beak” detector They can be compressed to the same parameters.
- 5. A CONVOLUTIONAL LAYER A filter A CNN is a neural network with some convolutional layers (and some other layers). A convolutional layer has a number of filters that does convolutional operation. Beak detector
- 6. CONVOLUTION 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 … … These are the network parameters to be learned. Each filter detects a small pattern (3 x 3).
- 7. CONVOLUTION 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 3 -1 stride=1 Dot product
- 8. CONVOLUTION 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 3 -3 If stride=2
- 9. CONVOLUTION 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 3 -1 -3 -1 -3 1 0 -3 -3 -3 0 1 3 -2 -2 -1 stride=1
- 10. CONVOLUTION 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 3 -1 -3 -1 -3 1 0 -3 -3 -3 0 1 3 -2 -2 -1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 -1 -1 -1 -1 -1 -1 -2 1 -1 -1 -2 1 -1 0 -4 3 Repeat this for each filter stride=1 Two 4 x 4 images Forming 2 x 4 x 4 matrix Feature Map
- 11. COLOR IMAGE: RGB 3 CHANNELS 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 1 -1 -1 -1 1 -1 -1 -1 1 1 -1 -1 -1 1 -1 -1 -1 1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 Color image
- 12. 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 image convolution -1 1 -1 -1 1 -1 -1 1 -1 1 -1 -1 -1 1 -1 -1 -1 1 1 x 2 x … … 36 x … … 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 Convolution v.s. Fully Connected Fully- connected
- 13. 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 1 2 3 … 8 9 … 13 14 15 … Only connect to 9 inputs, not fully connected 4 : 10: 16 1 0 0 0 0 1 0 0 0 0 1 1 3 fewer parameters!
- 14. 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 1 : 2 : 3 : … 7 : 8 : 9 : … 13 : 14 : 15 : … 4 : 10: 16 : 1 0 0 0 0 1 0 0 0 0 1 1 3 -1 Shared weights 6 x 6 image Fewer parameters Even fewer parameters
- 15. THE WHOLE CNN Fully Connected Feedforward network cat dog …… Convolution Max Pooling Convolution Max Pooling Flattened Can repeat many times
- 16. MAX POOLING 3 -1 -3 -1 -3 1 0 -3 -3 -3 0 1 3 -2 -2 -1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 -1 -1 -1 -1 -1 -1 -2 1 -1 -1 -2 1 -1 0 -4 3 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1
- 17. WHY POOLING Subsampling pixels will not change the object Subsampling bird bird We can subsample the pixels to make image smaller fewer parameters to characterize the image
- 18. A CNN COMPRESSES A FULLY CONNECTED NETWORK IN TWO WAYS: Reducing number of connections Shared weights on the edges Max pooling further reduces the complexity
- 19. MAX POOLING 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 3 0 1 3 -1 1 3 0 2 x 2 image Each filter is a channel New image but smaller Conv Max Poolin g
- 20. THE WHOLE CNN Convolution Max Pooling Convolution Max Pooling Can repeat many times A new image The number of channels is the number of filters Smaller than the original image 3 0 1 3 -1 1 3 0
- 21. THE WHOLE CNN Fully Connected Feedforward network cat dog …… Convolution Max Pooling Convolution Max Pooling Flattened A new image A new image
- 22. FLATTENING 3 0 1 3 -1 1 3 0 Flattened 3 0 1 3 -1 1 0 3 Fully Connected Feedforward network
- 23. Only modified the network structure and input format (vector -> 3-D tensor) CNN in Keras Convolution Max Pooling Convolution Max Pooling input 1 -1 -1 -1 1 -1 -1 -1 1 -1 1 -1 -1 1 -1 -1 1 -1 There are 25 3x3 filters. … … Input_shape = ( 28 , 28 , 1) 1: black/white, 3: RGB 28 x 28 pixels 3 -1 -3 1 3
- 24. Only modified the network structure and input format (vector -> 3-D array) CNN in Keras Convolution Max Pooling Convolution Max Pooling Input 1 x 28 x 28 25 x 26 x 26 25 x 13 x 13 50 x 11 x 11 50 x 5 x 5 How many parameters for each filter? How many parameters for each filter? 9 225= 25x9
- 25. Only modified the network structure and input format (vector -> 3-D array) CNN in Keras Convolution Max Pooling Convolution Max Pooling Input 1 x 28 x 28 25 x 26 x 26 25 x 13 x 13 50 x 11 x 11 50 x 5 x 5 Flattened 1250 Fully connected feedforward network Output
- 26. ALPHAGO Neural Network (19 x 19 positions) Next move 19 x 19 matrix Black: 1 white: -1 none: 0 Fully-connected feedforward network can be used But CNN performs much better
- 27. ALPHAGO’S POLICY NETWORK Note: AlphaGo does not use Max Pooling. The following is quotation from their Nature article:
- 28. CNN IN SPEECH RECOGNITION Time Frequency Spectrogram CNN Image The filters move in the frequency direction.
- 29. CNN IN TEXT CLASSIFICATION Source of image: http://citeseerx.ist.psu.ed u/viewdoc/download?doi=10.1.1.703.6858 &rep=rep1&type=pdf ?