Deep learning and computer vision
Download as PPTX, PDF2 likes651 views
The document discusses image classification using convolutional neural networks (CNNs), detailing their biological inspiration and evolution from early models like LeNet to advanced architectures such as ResNet and Inception. It highlights the significance of large-scale datasets like ImageNet and challenges in biomedical image analysis, emphasizing the need for methods like U-Net for segmentation tasks. Additionally, it addresses various advances in model regularization techniques and their impact on performance metrics in image recognition.
![20
Critical Feautures (Simonyan, K., & Zisserman, A., 2014 ):
• Kernels with small receptive fields: 3x3 which is the smallest size to capture the notion of
left/right up/down, center. It is easy to see that a stack of two 3×3 conv. layers (without spatial
pooling in between) has an effective receptive field of 5×5, and so on.
• Small size Receptive Field is a way to increase the nonlinearity of the decision function fields of
the conv. layers.
• Increasing depth architectures: VGG-16 (2xConv3-64, 2xConv3-128, 3xConv3-256, 6xConv3-512,
3xFC), VGG-19 (same as VGG-16 but with 8xConv3-512).
• Upside: less complex topology, outperforms GoogleNet on single-network classification accuracy
• Downside: 138 millions
parameters for VGG-16 !
Results:
• Multi ConvNet model :
(D/[256;512]/256,384,51
2),
(E/[256;512]/256,384,51
2), multi-crop & dense
eval: 23.7% Top-1 Error,
6.8% Top-5 Error.](https://image.slidesharecdn.com/deeplearningandcomputervision-170505232822/85/Deep-learning-and-computer-vision-20-320.jpg)
Deep learning and computer vision
- 1. Francesco Pugliese, PhD Italian National Institute of Statistics, Division "Information and Application Architecture“, Directorate for methodology and statistical design Matteo Testi, MsC Data Science at Sferanet S.r.l Email Francesco Pugliese : francesco.pugliese@istat.it Email Matteo Testi: testi@sferaspa.com
- 2. • Image classification is the task of taking an input image and outputting a class (a cat, dog, etc) or a probability of those classes that better describe the image. For humans, this task of recognition is one of the first skills we learn. • When we see an image or just when we look at the world around us, most of the time we are able to immediately characterize the scene and give each object a label, all this without even consciously noticing it. 2
- 3. 3 What we see What computers see
- 4. 4 Convolutional Neural Networks (CNNs) are biologically-inspired variants of MLPs. From Hubel and Wiesel’s early work on the cat’s visual cortex, we know the visual cortex contains a complex arrangement of cells. These cells are sensitive to small sub-regions of the visual field, called a receptive field.
- 5. 5 • For example, some neurons fired when exposed to vertical edges and some when shown horizontal or diagonal edges. • Feature identifiers
- 6. mercoledì 3 Maggio 2017 6 *
- 7. 7 LeNet was one of the very first convolutional neural networks which helped to propel the field of Deep Learning. This pioneering work by Yann LeCun was named LeNet5 after many previous successful iterations since the year 1988.
- 8. 8 MNIST database is handwritten digits composed of 60.000 pattern. • 60.000 training set • 10,000 test set LeNet has been applied to this dataset with accuracy of 0.95%.
- 9. 9
- 10. 10
- 11. 11
- 12. 12 So far, in tens of experiments, the resulting performances were many magnitudes better than other machine learning techniques today available. GPU •The advent of GPUs makes possible the training of very large neural networks with even more than 150 millions of parameters. BIG DATA • A new generation of larger training and test sets. Dropou t • Better model regularization techniques have been discovered such as “Dropout” or “Data Augmentation”
- 13. 13 - A new study proves the relationship between Vision capabilities and Intelligence (Tsukahara et al., 2016). - Computer Vision needs human-like abilities. EVERYDAY LIFE BIOMEDICAL IMAGES
- 14. 14 •A new generation of machines might accomplish typical human tasks such as recognizing and moving objects, driving cars, cultivating fields, cleaning streets, city garbage collecting, etc.
- 15. ImageNet: ImageNet is a dataset of over 15 million labeled high-resolution images belonging to roughly 22,000 categories. • Since 2010 a competition called «ImageNet Large-Scale Visual Recognition Challenge (ILSVRC)» uses a subset of ImageNet with roughly 1000 images in each of 1000 categories. Train Set: 1.2 million training images, Validation Set: 50,000 images Test set: 150,000 images. 15
- 16. Kaggle: In 2010, Kaggle was founded as a platform for predictive modeling and analytics competitions on which companies and researchers post their data. • Statisticians and data scientists from all over the world compete to produce the best models. • Data Science Bowl 2017 was the biggest competition focused on “Lung Cancer Detection”. The competition was founded by Arnold Foundation and awarded $1 million in prizes (1st ranked $500,000). Train Set: around 150 CT labelled scans images per patient from 1200 patients encoded in DICOM format. Stage 1 test set: 190 patients CT scans. Stage 2 test : 500 patients CT scans. 16 Grand Challenges in Biomedical Image Analysis: This is a web- site hosting new competitions in the Biomedicine field. Specifically, LUNA (LUng Nodule Analysis) focuses on a large- scale evaluation of automatic nodule detection algorithms. Train Set: LIDC/IDRI database consisting of 888 CT Scans labelled by 4 expert radiologists.
- 17. Each neuron in the convolutional layer is connected only to a local region in the input volume spatially. In this case there are 5 neurons along the depth all looking at the same region. 17 Convolutional Neural Networks (CNN) are biologically-inspired variants of MLPs. We know the visual cortex contains a complex arrangement of cells (Hubel, D. and Wiesel, T., 1968). These cells are sensitive to small sub-regions of the visual field, called a receptive field. Other layers are: RELU layer, Pool Layer. Typical CNNs settings are: a) Number of Kernels (Filters), b) Receptive Field size, b) Padding, c) Stride. These parameters are tied by the following equation:
- 18. 18 Conv Nets AlexNet (2012) GoogleNet (2014) Residual Nets (2015) VGG Net (2° ranked in 2014) Traditional issues with Convolutional Layers: •Wide convolutional layers determine overfitting and vanishing gradient problem with the Solver (SGD, Adam, etc). •Low depth architectures produce raw features (need to push the depth forward). Model Regularization : •Dropout - Co-adaptation and Models Ensemble (Srivastava, N. et al., 2014). •Weight penalty L1/L2 •Data Augmentation (crop, flip, rotation, ZCA whitening, etc.)
- 19. 19 Critical Feautures (Krizhevsky, A. et al, 2012) • 8 trainable layers: 5 convolutional layers and 3 fully connected layers. • Max pooling layers after 1st, 2nd and 5th layer. • Rectified Linear Units (ReLUs) (Nair, V., & Hinton, G. E. 2010). • Local Response Normalization. • 60 millions parameters, 650 thousands neurons. • Regularizations: Dropout (prob 0.5 in the first 2 fc layers, Data Augmentation (translactions, horizontal reflections, PCA on RGB). • Trained on 2 GTX 580 3 GB GPUs. Results: • 1 CNNs: 40.7% Top-1 Error, 18.2% Top-5 Error • 5 CNNs: 38.1% Top-1 Error,16.4% Top-5 Error • SIFT+FVs: 26.2% Top-5 Error (Sánchez, J., et al., 2013).
- 20. 20 Critical Feautures (Simonyan, K., & Zisserman, A., 2014 ): • Kernels with small receptive fields: 3x3 which is the smallest size to capture the notion of left/right up/down, center. It is easy to see that a stack of two 3×3 conv. layers (without spatial pooling in between) has an effective receptive field of 5×5, and so on. • Small size Receptive Field is a way to increase the nonlinearity of the decision function fields of the conv. layers. • Increasing depth architectures: VGG-16 (2xConv3-64, 2xConv3-128, 3xConv3-256, 6xConv3-512, 3xFC), VGG-19 (same as VGG-16 but with 8xConv3-512). • Upside: less complex topology, outperforms GoogleNet on single-network classification accuracy • Downside: 138 millions parameters for VGG-16 ! Results: • Multi ConvNet model : (D/[256;512]/256,384,51 2), (E/[256;512]/256,384,51 2), multi-crop & dense eval: 23.7% Top-1 Error, 6.8% Top-5 Error.
- 21. 21 Critical Feautures (Szegedy, C., et al., 2015): • Computationally Effective Deep architecture: 22 layers • Why the name inception, you ask? Because the module represents a network within a network. If you don't get the reference, go watch Christopher Nolan's “INCEPTION”, computer scientists are hilarious. • Inception: it isbasically the parallel combination of 1×1 3×3, and 5×5 convolutional filters. • Bottleneck layer: The great insight of the inception module is the use of 1×1 conv- olutional blocks (NiN) to reduce the number of features before the expensive parallel blocks. • Upside: 4 millions parameters! • Downside: Not scalable! Results: • 7 Models Ensemble : 6.67% Top-5 Error.
- 22. 22 Critical Feautures (He, K., et al., 2016). : • Degradation Problem: Stacking more and more layers IS NOT better. With the network depth increasing, accuracy gets saturated and then degrades rapidly! It’s an issue of “solvers”. • Solves the “Degradation problem”: by fitting a residual mapping which is easier to optimize. • Shortcut connections: • Very deep architecture: up to 1202 layers with WideResnet with only 19.4 million parameters! • Upside: Increasing accuracy with more depth • Downside: They don’t consider other architectures breakthroughs. Results: • ResNet : 3.57% Top-5 Error. • CNNs show superhuman abilities at Image Recognition! 5% Human estimated Top-5 error (Johnson, R. C., 2015).
- 23. 23
- 24. 24 Problems: • Feature extraction: In biomedicine feature extraction is not as easy as in an Imagenet competition with general images. A previous Image Preprocessing is needed. This is called Segmentation. • On Kaggle website there are whole competitions just regarding Segmentation. One of these was called «Ultrasound Nerve Segmentation».
- 25. 25 Critical Feautures (Ronneberger, O., et al., 2015): • U-NET can be trained end-to-end from very few images and outperforms the prior best methods. • It consists of a contracting path (left side) to capture context and an symmetric expansive path (right side) enabling precise localization. • Upsampling part (repeating rows and cols) has a large number of feature channels which allow the network to propagate context information to higher resolution layers. • Spatial Dropout: feature maps dropout. • Upside: Small training set. • Downside: Risk of overfitting.
- 26. 26 Candidate Nodule Selection via UNET Dilation, Erosion, Nodules Distance Merging False Positive Reduction via WideResNet
- 27. 27
- 28. Tsukahara, J. S., Harrison, T. L., & Engle, R. W. (2016). The relationship between baseline pupil size and intelligence. Cognitive Psychology, 91, 109-123. Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional architecture of monkey striate cortex. The Journal of physiology, 195(1), 215-243. Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1), 1929-1958. Nair, V., & Hinton, G. E. (2010). Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10) (pp. 807-814). Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105). Sánchez, J., Perronnin, F., Mensink, T., & Verbeek, J. (2013). Image classification with the fisher vector: Theory and practice. International journal of computer vision, 105(3), 222-245. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. 28 .
- 29. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 1-9). He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 770-778). Johnson, R. C. (2015). Microsoft, Google beat humans at image recognition. EE Times. Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 234-241). Springer International Publishing. 29 .
- 30. Thank you for attention. Francesco Pugliese Matteo Testi 30