[Revised] Intro to CNN
Download as PPTX, PDF•0 likes•180 views
Vincent gives an introductory presentation on convolutional neural networks (CNNs) for image recognition. He covers: 1) The principles of CNNs including convolution, ReLU activation, and max pooling for extracting features from images. 2) How CNN stacks are used along with a fully connected layer to generate predictions from feature maps. 3) Techniques for avoiding overfitting like data augmentation, dropout, and transfer learning by leveraging pretrained models.
![Principle of CNN: Convolution (1-D)
Which order of the element where I see there is a change from 1 to 0?
[1,1,1,1,0,0]
Hint: answer is [0,0,0,1,0]](https://image.slidesharecdn.com/googletechsprinttalkcnnintro-200730141604/85/Revised-Intro-to-CNN-23-320.jpg)
[Revised] Intro to CNN
- 1. Google Tech Sprint Intro to CNN By Vincent Tatan Google Data Analyst (Machine Learning)
- 2. Wait, who am I?
- 3. Meet Vincent Data Analyst (Machine Learning) Trust & Safety Google Medium : towardsdatascience.com/@vincentkernn Linkedin : linkedin.com/in/vincenttatan/ Data Podcast: https://datacast.simplecast.com/
- 4. Disclaimer This disclaimer informs readers that the views, thoughts, and opinions expressed in the text belong solely to the author, and not necessarily to the author’s employer, organization, committee or other group or individual. This article was made purely as the author’s initiatives and in no way driven by any other hidden agenda.
- 5. Path to Google Visa Business Intelligence Intern May 16 - Jul 16 Lazada Group Data Scientist Intern Dec 16 - Apr 17B.Sc., Management Information Systems and Services Aug 13 - July 17 Visa Data & Architecture Engineer Jun 17 - Aug 19 Google Data Analyst, Machine Learning Aug 19 - Present
- 6. Google: Trust and Safety To prevent phishing @ scale with Data Analytics and ML Google Data Analyst, Machine Learning Aug 19 - Present
- 10. Colab Demonstration Cat v.s. Dog using CNN https://colab.sandbox.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/images/classification.ipynb
- 11. Colab Demonstration Cat v.s. Dog using CNN https://colab.sandbox.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/images/classification.ipynb
- 12. Proprietary + Confidential Today’s Agenda 1 2 3 4 Machine Learning in Image Recognition Principles behind CNN: Convolution, ReLU, Max Pooling CNN stacks + Fully Connected Layer Avoiding Overfitting + Transfer Learning
- 13. 1 Machine Learning in Image Recognition Classify objects in images1
- 14. Image Recognition is hard -- Are all of these cats? My search for the word “cat”.
- 15. Warning -- Not all of these are cats. We need an automatic way to determine critical features to identify cats or other objects. Black color Sharp ears
- 16. We need a way to determine the features scalably
- 17. 2 Principles behind CNN Convolution, ReLU, Max Pooling
- 18. Introducing Convolutional Neural Network (CNN) Type of neural network comprising of different layers of convolutions and max pooling mainly for image recognition 1. Convolution 2. ReLU Activation 3. Max Pooling
- 19. Intuition of Convolutional Neural Network What do you see? Young lady or old grandma or both?
- 20. Intuition of Convolutional Neural Network Psychologist Edwin Boring introduced the painting of “My Wife and My Mother-in-Law” where the figure seems to morph from young lady to old woman, to the public in 1930. (Source) These boxes determined the features of the object that you would classify. ● The curve that changed into (ears/eyes) ● The straight line that changed into (choker/mouth)
- 21. Intuition of Convolutional Neural Network This means you are relying on ● Prior knowledge (your scrolling habits/understanding of faces) ● Data (strokes of the painting) To identify Features
- 22. Intuition of Convolutional Neural Network CNN distinguishes meaningful features from a sequence of pixels rather than pixel per pixel to classify the image 1 1 1 1 0 0 1 1 0 1 0 1 1 1 1 1 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0
- 23. Principle of CNN: Convolution (1-D) Which order of the element where I see there is a change from 1 to 0? [1,1,1,1,0,0] Hint: answer is [0,0,0,1,0]
- 24. Principle of CNN: Convolution (1-D) Which order of the element where I see 1 becomes -1? 1 -11 1 1 1 0 0 1st element : 1*1 + 1*-1 = 0 2nd element : 1*1 + 1*-1 = 0 3rd element : 1*1 + 1*-1 = 0 4th element : 1*1 + 0*-1 = 1 5th element : 0*1 + 0*-1 = 0 End Result: 0 0 0 1 0
- 25. Principle of CNN: Convolution (2-D) Same principle also applies to 2-D where you derive the feature map using the kernels (acts as weights in NN). With this kernel you can find diagonal patterns >
- 26. Principle of CNN: Convolution (2-D), 3 stacks Feature Maps
- 27. Principle of CNN: Activation Function ReLU The ReLU function introduces non linearity to activate on a “big enough” stimulus (sparse activation). This method has been effective to solve vanishing gradients due to its constant 1 for weight more than 1. Weights that are very small will remain as 0 after the ReLU activation function.
- 28. Principle of CNN: Maxpooling(2-D) CNN uses max pooling to replace output with a max summary to reduce data size and processing time. This allows you to determine features that produce the highest impact and reduces the risk of overfitting.
- 29. Recap Understanding CNN 1. Convolution 2. ReLU Activation Function 3. Maxpooling
- 30. 3 CNN Stacks + Fully Connected Layer Generating Real Results
- 31. CNN Stacks + Fully Connected Layer(2-D) Take the Feature maps from CNN as the inputs and insert it into NN models which will calculate the probabilities of the image to the labels assigned. Finally, we will serve the convolutional and max pooling feature map outputs with Fully Connected Layer (FCL). We flatten the feature outputs to column vector and feed-forward it to FCL. Every node in the previous layer is connected to the last layer and represents which distinct label to output. FCL represents Artificial Neural Network and has the final layer as MultiLayer Perceptron (MLP)
- 32. CNN Stacks: Cats vs Dogs The discovery of curves
- 33. Fully Connected Layer Activation: Softmax softmax activation function squashes the probability scores to add up to 1.0.
- 34. CNN Big Picture: Cats vs Dogs The sensors of the neural network might start by understanding what edges and colors it is looking at. The last layer, called the fully connected layer, summarizes this analysis and gives the probability of whether it is looking at a dog. A third layer is trying to conceptualize what these shapes represent (for instance, is it an animal or something else?). Another layer is trying to understand what shapes it is looking at.
- 35. Using Tensorflow + Keras
- 36. Create and Compile our CNN Change to ‘categorical_crossentropy’ for softmax CNN Stacks Fully Connected Layer
- 37. Train our CNN
- 38. Results: Cats vs Dogs
- 39. Results: Cats vs Dogs Overfitting! loss: 0.0134 acc: 0.9965 val_loss: 3.5137 val_acc: 0.7290
- 40. 4 Avoiding Overfitting + Transfer Learning Improve Generalization and leverage pretrained model
- 41. Overfitting Mistakes: CNN Common Mistakes Solutions Using test set as the validation set to test the model Use validation sets as validation tests and test models as final tests Dataset is relatively small Image augmentations to add new variants of images Over Memorization Dropout/ reduce epochs/layers/neurons per layer to increase generalizations
- 42. Overfitting Solutions: Data Augmentation In the case of small training data set, you need to artificially boost the diversity and number of training examples. One way of doing this is to add image augmentations and creating new variants. These include translating images and creating dimension changes such as zoom, crop, flips, etc.
- 43. Overfitting Solutions: Data Augmentation
- 44. Overfitting Solutions: Dropout You could also use regularization techniques such as Dropout to remove activation unit in every gradient step training. Each epoch training deactivates different neurons. Since the number of gradient steps is usually high, all neurons will averagely have same occurrences for dropout. Intuitively, the more you drop out, the less likely your model memorizes.
- 47. Results: Cats vs Dogs Much better loss: 0.5622 acc: 0.7145 val_loss: 0.5090 val_acc: 0.7650
- 48. Results: Cats vs Dogs Overfitting CNN CNN with Dropout and Image Augmentation loss: 0.5622 acc: 0.7145 val_loss: 0.5090 val_acc: 0.7650 loss: 0.0134 acc: 0.9965 val_loss: 3.5137 val_acc: 0.7290
- 49. Transfer Learning: ImageNet Transfer learning is a technique that reuses an existing model to the current model. You could produce on top of existing models that were carefully designed by experts and trained with millions of pictures. You can freeze the model, then only apply the training for the high level layers.
- 50. Important Resources (Teachable Machine) https://teachablemachine.withgoogle.com/train/image
- 51. Proprietary + Confidential Today’s Summary 1 2 3 4 Machine Learning in Image Recognition Principles behind CNN: Convolution, ReLU, Max Pooling CNN stacks + Fully Connected Layer Avoiding Overfitting + Transfer Learning
- 52. Proprietary + Confidential Today’s Summary Machine Learning in Image Recognition Principles behind CNN: Convolution, ReLU, Max Pooling CNN stacks + Fully Connected Layer Transfer Learning Avoiding Overfitting + Transfer Learning 04 ● Data Augmentation: artificially boost diversity and number of training examples ● Dropout: Regularization Techniques to remove activation unit in every gradient step training. Each epoch training deactivates different neurons. ● Transfer learning is a technique that reuses an existing model to the current model. You could produce on top of existing models that were carefully designed by experts and trained with millions of pictures. CNN Stacks +Fully Connected Layer 03 ● Fully Connected Layer: Take the Feature maps from CNN as the inputs and insert it into NN models which will calculate the probabilities of the image to the labels assigned. ● Softmax activation function which assign decimal probabilities for each possible label which add up to 1.0. Principles behind CNN: Convolution, ReLU, and Max Pooling02 ● Convolution: Same principle also applies to 2-D where you derive from the feature map using the kernels (acts as weights in NN). ● The ReLU function introduces non linearity to activate on a “big enough” stimulus (sparse activation). ● Max pooling to replace output with a max summary to reduce data size and processing time. Machine Learning in Image Recognition 01 ● Classical Image Recognition Allows you to extract features and derive characteristics. ● Convolution Neural Network (CNN) Deep learning techniques using convolutions, max pooling to share weights, produce critical features, and classify labels. Commonly used in image classifications
- 53. Colab Demonstration Cat v.s. Dog using CNN https://colab.sandbox.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/images/classification.ipynb
- 54. Reach out to me :) Medium : towardsdatascience.com/@vincentkernn Linkedin : linkedin.com/in/vincenttatan/ Survey: tiny.cc/vincent-survey