Deep learning for Industries
0 likes163 views
The document discusses deep learning and its mechanisms, comparing them to how the human brain functions, emphasizing their ability to identify patterns and simplify complex information. It covers the foundational concepts of neural networks, particularly convolutional neural networks (CNNs), and their applications in various real-world scenarios like computer vision and image classification. Additionally, the document mentions the use of TensorFlow for implementing deep learning models and highlights the potential power and risks associated with these technologies.
![[224x224x3]
f 1000 numbers,
indicating class scores
Feature
Extraction
vector describing
various image statistics
[224x224x3]
f 1000 numbers,
indicating class scores
training
training
@hellorahulk](https://image.slidesharecdn.com/deeplearningforindustriesv2-180629114530/85/Deep-learning-for-Industries-14-320.jpg)
![DNN Approach
CNN Features off-the-shelf: an Astounding Baseline for Recognition
[Razavian et al, 2014]
@hellorahulk](https://image.slidesharecdn.com/deeplearningforindustriesv2-180629114530/85/Deep-learning-for-Industries-16-320.jpg)
![ConvNets are everywhere…
e.g. Google Photos search
Face Verification, Taigman et al. 2014 (FAIR)
Self-driving cars
[Goodfellow et al. 2014]
Ciresan et al. 2013
@hellorahulk](https://image.slidesharecdn.com/deeplearningforindustriesv2-180629114530/85/Deep-learning-for-Industries-25-320.jpg)
![[224x224x3]
f 1000 numbers,
indicating class scores
training
Only two basic operations are involved throughout:
1. Dot products wT x
2. Max operations max(.)
@hellorahulk](https://image.slidesharecdn.com/deeplearningforindustriesv2-180629114530/85/Deep-learning-for-Industries-28-320.jpg)
Deep learning for Industries
- 1. Deep Learning for Industries Rahul Kumar Chief AI Scientist @BotSupply.ai / Jatana.ai Demystifying Deep Learning Hands on word2vec
- 3. Deep Learning for Industries DE3p Larenn1g mhica3ns wrok smliair to hOw biarns wrok. Tehse mahcnies wrok by s33nig f22Uy pa773rns and cnonc3t1ng t3Hm t0 fU22y cnoc3tps. T3hy wRok l4y3r by ly43r, j5ut lK1e a f1L37r, t4k1NG cmopl3x scn33s aNd br3k41ng tH3m dwon itno s1pmLe iD34s. @hellorahulk
- 4. Deep Learning for Industries Deep Learning mechanism work similar to how brain work. These machines works by seeing funny patterns and connecting them to funny concepts. They work layer by layer, just like a filter, taking complex scenes and breaking them down into simple ideas. @hellorahulk
- 6. So far... Some input vector (very few assumptions made). Will discuss in detail @hellorahulk
- 7. In many real-world applications input vectors have structure. Spectrograms ImagesText @hellorahulk
- 8. Neural Networks: A pinch of history @hellorahulk
- 9. Hubel & Wiesel, 1959 RECEPTIVE FIELDS OF SINGLE NEURONES IN THE CAT'S STRIATE CORTEX 1962 RECEPTIVE FIELDS, BINOCULAR INTERACTION AND FUNCTIONAL ARCHITECTURE IN THE CAT'S VISUAL CORTEX 1968... @hellorahulk
- 13. Computer Vision 2011 Page 3+ code complexity :( @hellorahulk
- 14. [224x224x3] f 1000 numbers, indicating class scores Feature Extraction vector describing various image statistics [224x224x3] f 1000 numbers, indicating class scores training training @hellorahulk
- 15. “Run the image through 20 layers of 3x3 convolutions and train the filters with SGD.” @hellorahulk
- 16. DNN Approach CNN Features off-the-shelf: an Astounding Baseline for Recognition [Razavian et al, 2014] @hellorahulk
- 17. e.g. with TensorFlow The power is easily accessible. # Python 2 $ sudo pip install --upgrade tensorflow # Python 3 $ sudo pip3 install --upgrade tensorflow @hellorahulk
- 18. TensorFlow : Programming Paradigm Load library and MNIST data @hellorahulk
- 19. Design neural network architecture TensorFlow : Programming Paradigm @hellorahulk
- 20. TensorFlow : Programming Paradigm Select optimization algorithm @hellorahulk
- 21. TensorFlow : Programming Paradigm Initialize the session and variables @hellorahulk
- 22. TensorFlow : Programming Paradigm Train the model @hellorahulk
- 23. TensorFlow : Programming Paradigm 95.3% @hellorahulk
- 25. ConvNets are everywhere… e.g. Google Photos search Face Verification, Taigman et al. 2014 (FAIR) Self-driving cars [Goodfellow et al. 2014] Ciresan et al. 2013 @hellorahulk
- 26. ConvNets are everywhere… Whale recognition, Kaggle Challenge Satellite image analysis Mnih and Hinton, 2010 Galaxy Challenge Dielman et al. 2015 WaveNet, van den Oord et al. 2016 Image captioning, Vinyals et al. 2015 @hellorahulk
- 27. ConvNets are everywhere… DeepDream reddit.com/r/deepdream NeuralStyle, Gatys et al. 2015 deepart.io, Prisma, etc. @hellorahulk
- 28. [224x224x3] f 1000 numbers, indicating class scores training Only two basic operations are involved throughout: 1. Dot products wT x 2. Max operations max(.) @hellorahulk
- 29. e.g. 200K numbers e.g. 10 numbers @hellorahulk
- 31. 32 32 3 Convolution Layer 5x5x3 filter 32x32x3 image Convolve the filter with the image i.e. “slide over the image spatially, computing dot products” @hellorahulk
- 32. 32 32 3 Convolution Layer 5x5x3 filter 32x32x3 image Convolve the filter with the image i.e. “slide over the image spatially, computing dot products” Filters always extend the full depth of the input volume @hellorahulk
- 33. 32 32 3 Convolution Layer 32x32x3 image 5x5x3 filter 1 number: the result of taking a dot product between the filter and a small 5x5x3 chunk of the image (i.e. 5*5*3 = 75-dimensional dot product + bias) @hellorahulk
- 34. 32 32 3 Convolution Layer 32x32x3 image 5x5x3 filter convolve (slide) over all spatial locations activation map 1 28 28 @hellorahulk
- 35. 32 32 3 Convolution Layer 32x32x3 image 5x5x3 filter convolve (slide) over all spatial locations activation maps 1 28 28 consider a second, green filter @hellorahulk
- 36. 32 32 3 Convolution Layer activation maps 6 28 28 For example, if we had 6 5x5 filters, we’ll get 6 separate activation maps: We stack these up to get a “new image” of size 28x28x6! @hellorahulk
- 37. ConvNet is a sequence of Convolution Layers, interspersed with activation functions Sigmoid Tanh ReLU @hellorahulk
- 38. ConvNet is a sequence of Convolution Layers, interspersed with activation functions 32 32 3 CONV, ReLU e.g. 6 5x5x3 filters 28 28 6 CONV, ReLU e.g. 10 5x5x6 filters CONV, ReLU …. 10 24 24 @hellorahulk
- 39. two more layers to go: POOL/FC @hellorahulk
- 40. Pooling layer - makes the representations smaller and more manageable - operates over each activation map independently: @hellorahulk
- 41. 1 1 2 4 5 6 7 8 3 2 1 0 1 2 3 4 Single depth slice x y max pool with 2x2 filters and stride 2 6 8 3 4 MAX POOLING @hellorahulk
- 42. Fully Connected Layer (FC layer) - Contains neurons that connect to the entire input volume, as in ordinary Neural Networks @hellorahulk
- 44. Addressing other tasks... image CNN feature s 224x224x3 A block of compute with a few million calculations. 7x7x512 @hellorahulk
- 45. Addressing other tasks... image CNN feature s 224x224x3 A block of compute with a few million calculations. 7x7x512 predicted thing desired thing @hellorahulk
- 46. Addressing other tasks... image CNN feature s 224x224x3 A block of compute with a few million parameters. 7x7x512 predicted thing desired thing this part changes from task to task @hellorahulk
- 47. Image Classification thing = a vector of probabilities for different classes image CNN feature s 224x224x3 7x7x512 e.g. vector of 1000 numbers giving probabilities for different classes. fully connected layer @hellorahulk
- 48. Image Captioning image CNN feature s 224x224x3 7x7x512 A sequence of 10,000-dimensional vectors giving probabilities of different words in the caption. RNN @hellorahulk
- 49. Localization image CNN feature s 224x224x3 7x7x512 fully connected layer Class probabilities (as before) 4 numbers: - X coord - Y coord - Width - Height @hellorahulk
- 50. Reinforcement Learning image CNN feature s 160x210x3 fully connected e.g. vector of 8 numbers giving probability of wanting to take any of the 8 possible ATARI actions. Mnih et al. 2015 @hellorahulk
- 51. Segmentation image CNN feature s 224x224x3 7x7x512 deconv layers 224x224x20 array of class probabilities at each pixel. image class “map” @hellorahulk
- 52. Hands on word2vec Colab link : https://goo.gl/7H7mVo @hellorahulk
- 53. Thank you! "Deep learning" offer us great power - and pose unique risks. Can we Vectorise them? @hellorahulk