Teaching machines to see the process of designing (datasets) with ai
2 likes127 views
This document summarizes a presentation about teaching machines to see and recognize images using artificial intelligence. It discusses Clarifai, a company that provides image and video recognition services using convolutional neural networks (CNNs). The presentation explains how CNNs can learn representations of images without needing manual feature engineering. It demonstrates how CNNs can be trained to detect safe versus unsafe image content. Finally, it discusses how CNNs have progressed over time to surpass traditional computer vision techniques for tasks like object detection.
![Deconvolutional Networks
• Provides way to map activations at high
layers back to the input
• Same operations as Convnet, but in reverse:
– Unpool feature maps
– Convolve unpooled maps
• Filters copied from Convnet
– Used here purely as a probe
– Originally proposed as unsupervised learning
method
– No inference, no learning
– Compare with sparse coding Input Image
Convolution (learned)
Unpooling
Feature maps
Non-linearity
[Zeiler et al. CVPR’10, ICCV’11]](https://image.slidesharecdn.com/teachingmachinestosee-theprocessofdesigningdatasetswithai-170511022454/85/Teaching-machines-to-see-the-process-of-designing-datasets-with-ai-39-320.jpg)
Teaching machines to see the process of designing (datasets) with ai
- 1. Teaching Machines to See The process of designing (datasets) with AI Ryan Compton [email protected]
- 2. Clarifai - Clarifai.com - 40 people • $40M • Imagenet ‘13 - Image + Video Recognition - NYC • SF
- 3. Current product applications - Navigating and understanding internal image libraries - Auto-proposing tags upon upload to a service - ‘You may also like’ carousel - Filtering user generated content
- 4. Framework - mysterious but not magical
- 5. Demo
- 6. Easy to use API
- 7. How simple is it to Search?
- 8. How simple is it to Train?
- 9. So why is it so easy ?
- 14. Object detection in 1978
- 15. Object detection in 1996 Jain, A. K., Zhong, Y., & Lakshmanan, S. (1996). Object matching using deformable templates. IEEE Transactions on pattern analysis and machine intelligence, 18(3), 267-278.
- 16. Learning templates in 2009 https://arxiv.org/abs/1104.2123
- 17. Disclaimer: This presentation will contain visualizations of nudity for scientific purposes.
- 18. 1996
- 25. Eventually someone will get real good at feature engineering and the problem will be solved, right?
- 28. But how ?
- 31. convnet Demo network for this talk (modified ZF net): 5 convolutional layers 3 fully-connected layers 2 softmax outputs (‘sfw’ and ‘nsfw’) 69,961,504 parameters
- 32. Train Test
- 34. Train Test Deploy to prod and everything breaks
- 35. And how does it do that ?
- 39. Deconvolutional Networks • Provides way to map activations at high layers back to the input • Same operations as Convnet, but in reverse: – Unpool feature maps – Convolve unpooled maps • Filters copied from Convnet – Used here purely as a probe – Originally proposed as unsupervised learning method – No inference, no learning – Compare with sparse coding Input Image Convolution (learned) Unpooling Feature maps Non-linearity [Zeiler et al. CVPR’10, ICCV’11]
- 40. Sparse Coding
- 41. wavelets, curvelets, framelets, shapelets, shearlets...
- 42. Objective for a single layer in a deconvnet
- 45. Reversible Max Pooling Pooled Feature Maps Max Locations “Switches” Pooling Unpooling Feature Map Reconstructed Feature Map
- 46. 2014
- 47. First Layer Image from Zeiler, Matthew D., and Rob Fergus. "Visualizing and understanding convolutional networks." European Conference on Computer Vision. Springer International Publishing, 2014.
- 48. Second Layer
- 49. Third Layer
- 50. Fourth Layer
- 51. Fifth Layer
- 52. conv1
- 53. conv2
- 54. conv3
- 55. conv4
- 56. conv5
- 57. fc8
- 59. Oh and safe stuff too...
- 61. Oh and safe stuff too...
- 67. Recap 20161996
- 68. The Future...
- 69. Context matters
- 70. Context matters
- 71. Context matters
- 72. Context matters
- 75. DEVELOPERS 10,000 Free API Credits Get started now! developer.clarifai.com