Image Net Classification with Deep Convolutional Neural Networks
- Slides: 30
Image. Net Classification with Deep Convolutional Neural Networks http: //people. cs. ksu. edu/~okerinde /
Introduction Objective: ü To train a large, deep convolutional neural network (on the subsets of Image. Net) to classify 1. 2 million high-resolution images into 1000 different categories. ü To learn about thousands of objects from millions of images
Introduction Dataset: ü Over 15 million labeled high-resolution images belonging to roughly 22, 000 categories ü Amazon’s Mechanical Turk crowd-sourcing tool ü 1. 2 million training images ü 50, 000 validation images ü 150, 000 testing images
Dataset
Terminology ü Deep Convolutional Neural Networks – Convolutional neural networks’ (CNNs) capacity can be controlled by varying their depth and breadth, and they also make strong and mostly correct assumptions about the nature of images (namely, stationarity of statistics and locality of pixel dependencies). ü Object recognition
Methodology ü The Architecture ü Re. LU Nonlinearity ü Local Response Normalization ü Dropout ü Data Augmentation ü Principal Component Analysis ü Stochastic Gradient with Momentum
Re. LU Nonlinearity
Aside Re. LU Nonlinearity Source: http: //cs 231 n. stanford. edu/slides/2017/cs 231 n_2017_lecture 7. pdf
Methodology
Methodology Reducing Overfitting – ü Data Augmentation ü Dropout
Methodology
Methodology
Methodology Data Augmentation Section 4. 1 …discuss
Methodology The Neural Network has: ü 60 million parameters ü 650, 000 neurons ü Five convolutional layers ü Three fully-connected layers ü 1000 -way softmax
Methodology Regularization method used to reduce overfitting – ü Dropout.
Aside Drop. Out… Source: http: //cs 231 n. stanford. edu/slides/2017/cs 231 n_2017_lecture 7. pdf
Methodology Local Response Normalization Section 3. 3. …discuss
Methodology Training Time Complexity: ü Non-saturating neurons ü Efficient GPU implementation of the convolution operation ü Network takes between five and six days to train on two GTX 580 3 GB GPUs
Methodology cs 231 n_2017_lecture 5
Methodology cs 231 n_2017_lecture 5
Methodology Overlapping Pooling Section 3. 4 …discuss
Methodology Details of Learning ü stochastic gradient descent ü Batch size of 128 examples ü Momentum of 0. 9 ü Weight decay of 0. 0005
Methodology cs 231 n_2017_lecture 7
Methodology Details of Learning Section 5 …discuss
Experiments, Result and Discussion Error rates Top-1 => 37. 5% Top-5 => 17. 0% Top-5 error rate is the fraction of test images for which the correct label is not among the five labels considered most probable by the model
Experiments, Result and Discussion
Conclusion and Future Work ü Result showed a large, deep convolutional neural network is capable of achieving record-breaking results on a highly challenging dataset using purely supervised learning ü Network degrades if a single convolutional layer is removed ü Depth is important ü Yet to match the infero-temporal pathway of human visual system ü Future work on video sequences
Resource ü https: //www. youtube. com/watch? v=40 ri. Cqv. Ro. Ms
Similar Paper
THANK YOU FOR LISTENING http: //people. cs. ksu. edu/~okerinde/
- Leon gatys
- Image super resolution using deep convolutional networks
- Deep forest towards an alternative to deep neural networks
- Visualizing and understanding convolutional networks
- Convolutional neural networks for visual recognition
- Stride in cnn
- Csrmm
- Xooutput
- Lmu cis
- Convolutional neural networks
- Convolutional neural network alternatives
- Deep convolutional networks
- Convolutional deep belief networks
- Efficient processing of deep neural networks pdf
- Mippers
- Feature map in cnn
- Modeling relational data with graph convolutional networks
- Instance segmentation
- Vc bound
- Ib psychology
- Audio super resolution using neural networks
- Least mean square algorithm in neural network
- Pixel recurrent neural networks.
- Matlab neural network toolbox
- Neural networks for rf and microwave design
- 11-747 neural networks for nlp
- Xor problem
- On the computational efficiency of training neural networks
- Tlu perceptron
- Fuzzy logic lecture
- Few shot learning with graph neural networks