Deep Learning And Business Models (VNITC 2015-09-13)
8 likes4,735 views
Deep Learning and Business Models Tran Quoc Hoan discusses deep learning and its applications, as well as potential business models. Deep learning has led to significant improvements in areas like image and speech recognition compared to traditional machine learning. Some business models highlighted include developing deep learning frameworks, building hardware optimized for deep learning, using deep learning for IoT applications, and providing deep learning APIs and services. Deep learning shows promise across many sectors but also faces challenges in fully realizing its potential.
![RNN (3/3)
Neural Turing Machine [A. Graves, 2014]
Neural network which has the capability of coupling the external memories
Applications: COPY, PRIORITY SORT
35
NN with
parameters to
coupling to
external
memories](https://image.slidesharecdn.com/deeplearningbusinessmodelsvnitc-2015-09-13-150912180100-lva1-app6891/85/Deep-Learning-And-Business-Models-VNITC-2015-09-13-35-320.jpg)
![40
Deep Reinforcement Learning (RL)
Q - learning: need to know future expectation reward Q(s, a)
of action a at state s (Bellman update equation)
Q(s, a) <— Q(s, a) + ß ( r + µ maxa’ Q(s’, a’) - Q(s, a) )
Deep Q-Learning Network [V. Mnih, 2015]
Get Q(s, a) by a deep neural network Q(s, a, w)
Maximize: L(w) = E[ (r + µ max Q(s’, a’, w) - Q(s, a, w))2 ]
Useful when there are many states](https://image.slidesharecdn.com/deeplearningbusinessmodelsvnitc-2015-09-13-150912180100-lva1-app6891/85/Deep-Learning-And-Business-Models-VNITC-2015-09-13-40-320.jpg)
Deep Learning And Business Models (VNITC 2015-09-13)
- 1. Text Deep Learning and Business Models Tran Quoc Hoan The University of Tokyo VNITC@2015-09-13
- 2. 2 Today agendas Deep Learning boom Essentials of Deep Learning Deep Learning - the newest Deep Learning and Business Models
- 3. 3 What is Deep Learning? Deep Learning is a new trend of machine learning that enables machines to unravel high level abstraction in large amount of data Image recognition Speech recognition Customer Centric Management Natural Language Processing Drug Discovery & Toxicology
- 4. Deep Learning Boom –Albert Einstein “If we knew what it was we were doing, it would not be called research, would it?”
- 5. Big News - ILSVRC 2012 ILSVRC 2012 (ImageNet Large Scale Visual Recognition) SuperVision ISI OXFORD_VGG XRCE/INRIA UnivofAmsterdam 29.576% 27.058%26.979%26.172% 15.315% Error rate Image source: http://cs.stanford.edu/people/karpathy/cnnembed/ 5
- 6. 6 Big News - Google brain Self-taught learning with unlabelled youtube videos and 16,000 computers AI grandmother cell (2012) (Cat)(Human)
- 7. 7 Data and Machine Learning Most learning algorithms New AI method (Deep Learning) Amount of data Performance Image source: http://cs229.stanford.edu/materials/CS229-DeepLearning.pdf
- 8. 8 New boom for the Giants & Startups Google brain project (2012) Facebook AI Research Lab (2013 Dec.) Project Adam (2014) Enlitic Ersatz Labs MetaMind Nervana Systems Skymind Deepmind One of 10 breakthrough technologies 2013 (MIT Technology Review) PFN (Japan)
- 9. 9 Pioneers Geoffrey Hinton (Toronto, Google) Yoshua Bengio (Montreal) Yann LeCun (NewYork, Facebook) Andrew Ng (Stanford, Baidu) Jeffrey Dean (Google) Le Viet Quoc (Stanford, Google) Theory + algorithms Implementation
- 10. 10 Sub-summary ・More than 2000 papers in 2014 ~ 2015 with 47 Google services.
- 11. Essentials of Deep Learning –Albert Einstein “If we knew what it was we were doing, it would not be called research, would it?”
- 12. 12 The basic calculation x1 x2 x3 +1 z f Active function z = f(x1w1 + x2w2 + x3w3 + w4) w1 w2 w3 w4 Input Forward propagation wi: weight parameters
- 13. 13 Neural network (1/2) x1 x2 x3 +1 Input layer +1 +1 y Hidden layers Output layer f f f f h1 h2 z1 z2 Put deeper The deeper layer displays for comprehensive combination of small parts Prob(cat) = 99%
- 14. 14 Neural network (2/2) x1 x2 x3 +1 Input layer +1 +1 y1 Hidden layers • Loss function E measures how output labels are similar to true labels. • For N input data {w1ij} {w2ij} {w3ij} Supervisor learning y2 yk Output (Ex. probability of each in k class in classification task) Find parameters {w} to minimize E → How?
- 15. 15 Backward propagation x1 x2 x3 +1 Input layer r s y Output layer y* E w ∂E ∂y ∂E ∂y ∂y ∂s ∂E ∂y ∂y ∂s ∂s ∂w ∂E ∂w = Backward from loss to update parameters (true label) How E change when move w How E change when move s How E change when move y
- 16. 16 Vanishing gradient problem x1 x2 x3 +1 +1 +1 y y* E … Error vanish with back-propagation at low layer ・Neural network architecture reached the limit before deep learning boom
- 17. 17 Features representation ・Deep learning reduced the human’s time-consuming of features selection process in classification task Features selection Parameter learning Answer Features selection + Parameter learning Answer Previous methods Deep Learning
- 18. Convolutional Neural Network ConvNet diagram from Torch Tutorial Sub sampling Emphasis special features The most popular kind of deep learning model 18
- 19. How to train parameters Trainable parameters Values in each kernel of convolutional layer Weigh values and bias in fully connection layer Stochastic Gradient Descent (SGD) (others: AdaGrad, Adam,…) Parameters update W ← W - ß ∂E ∂W Backward propagation Learning rate
- 20. SGD - Mini batch Batch learning: update all samples in each update (over-fitting) Mini batch: update parameters with some samples 20
- 21. Training trick Data normalization Reduce learning rate after some iterations Momentum SGD 21
- 22. 22 Auto-encoder (1/2) y … How to initialize these parameters
- 23. 23 Auto-encoder (2/2) The ability of reconstruct input makes a good initialization for parameters x x’ h(x) Reconstruct error E = ||x - x’|| Encode Decode
- 24. 24 Robust algorithms Data augmentation: more noisy data → more robust model Nodes in network don’t need to active all (mimic human brain) → Dropout concept …
- 25. Deep Learning - The Newest –Albert Einstein “If we knew what it was we were doing, it would not be called research, would it?”
- 26. 26 In Services Google Translate App Deep learning inside (work offline) July 29, 2015 version
- 27. 27 In Services My implemented service http://yelpio.hongo.wide.ad.jp/
- 28. 28 In Medical Imaging Segmentation Offline due to image’s copyright
- 30. In Speech Interfaces 99% is much different with 95% (Andrew Ng. - Baidu) https://medium.com/s-c-a-l-e/how-baidu-mastered-mandarin-with-deep- learning-and-lots-of-data-1d94032564a5 30
- 31. In Drug Discovery Active Active Inactive 1 0 1 0 0 1 10010000110011 Chemical compound Assay data Finger print + Activity Deep Neural Net PubChem Database Prediction of Drug Activity multiple targets http://on-demand.gputechconf.com/gtc/2015/presentation/S5813-Nobuyuki-Ota.pdf 31
- 32. Applications in IoT(Internet of Things) More difficult tasks in AI, robotics, information processing Huge amount of time series data and states of sensor & devices data RNN (recurrent neural network) Difficult to get supervisor data VAE (variational auto encoder) Take action in conditions, environments Deep Reinforcement Learning Image source: http://on-demand.gputechconf.com/gtc/2015/presentation/S5813- Nobuyuki-Ota.pdf 32
- 33. RNN (1/3) Recurrent Neural Network: loop inside neural network Represent for time series data, sequence of inputs (speech model, natural language model,…) . . . . . . . . . x(t) y(t) z(t) = f( z(t-1), x(t) ) 33
- 34. RNN (2/3) Time series extension of RNN … … t = 1 t = 2 t = 3 t = T Back propagation through time (BPTT) + Long Short-Term Memory How to learn parameters? 34 … … … … … … … … … … …
- 35. RNN (3/3) Neural Turing Machine [A. Graves, 2014] Neural network which has the capability of coupling the external memories Applications: COPY, PRIORITY SORT 35 NN with parameters to coupling to external memories
- 36. 36 VAE (variational auto encoder) Learn a mapping from some latent variable z to a complicated distribution on x p(x) = ∫p(x,z)dz where p(x, z) = p(x|z)p(z) p(z) = something simple and p(x|z) = f(z) = neural network Encode
- 37. VAE Example: generate face image from 29 hidden variables http://vdumoulin.github.io/morphing_faces/online_demo.html 37
- 38. VAE Example: learning from number and writing style Experiment (PFN): given a number x, produce outputs as remained numbers with writing style of x VAE: useful in half-supervised learning process (especially when training data are not enough) Image from ディープラーニングが活かすIoT, Preferred Networks, Inc. 2015/06/09 Interop 2015 seminar
- 39. 39 Reinforcement Learning (RL) Environment (street, factory,…) Agent (car, robot,…) Action a(t) Reward r(t) State s(t) Total rewards in the future R = r(t) + µr(t+1) + µ2r(t+2) + … (µ < 1) Design next action
- 40. 40 Deep Reinforcement Learning (RL) Q - learning: need to know future expectation reward Q(s, a) of action a at state s (Bellman update equation) Q(s, a) <— Q(s, a) + ß ( r + µ maxa’ Q(s’, a’) - Q(s, a) ) Deep Q-Learning Network [V. Mnih, 2015] Get Q(s, a) by a deep neural network Q(s, a, w) Maximize: L(w) = E[ (r + µ max Q(s’, a’, w) - Q(s, a, w))2 ] Useful when there are many states
- 41. 41 PFN - demo https://www.youtube.com/watch?v=RH2TmreYkdA
- 42. Deep Learning and Business Models –Albert Einstein “If we knew what it was we were doing, it would not be called research, would it?”
- 43. The future of the field Better hardware and bigger machine cluster Better implementations and optimizations Understand video, text and signals Develop in business models Deep learning is just the tip of the iceberg 43
- 44. Model 1 - Framework development Bridge the gap between algorithms and implementation Provide common interfaces and understandable API for users Pylearn2 44
- 45. Model 2 - Building a deep-able hardware New hardware (chipset) architecture for Deep Learning Synopsys http://www.bdti.com/InsideDSP/2015/04/21/Synopsys Processors’ convolutional network capabilities Qualcomm Zeroth Platform Cognitive computing and custom CPU drive Snapdragon 820 processors https://www.qualcomm.com/news/snapdragon/2015/03/02/cognitive- computing-and-custom-cpu-drive-next-gen-snapdragon-processors NVIDIA DIGITS DevBox https://developer.nvidia.com/devbox 45
- 46. Model 3 - Deep Intelligences for IoT Sensing Action Feedback Analysis Union Image from ディープラーニングが活かすIoT, Preferred Networks, Inc. 2015/06/09 Interop 2015 seminar
- 47. Model 4 - Host API and deep learning service Data in Model out https://www.metamind.io/vision/train https://www.metamind.io/language/train Usability + Tuning + Scale out 47
- 48. Model 5 - Personal Deep Learning Data in Label out DeepFace (Facebook) http://www.adweek.com/socialtimes/deepface/433401 Facial verification & tagging Applied previous solutions Almost free 48
- 49. Summary Deep Learning has the capability to find patterns among data by enabling wide range of abstraction. Deep Learning shown significant results in voice and image recognition compared with conventional machine learning methods. Deep Learning has potential applications and business models in some important key sectors. The challenges are your ideas ! 49
- 50. “Thank you for listening.” • Some good materials for learning • Papers, code, blog, seminars, online courses • For beginner: 深層学習 (機械学習プロフェッショナルシリーズ) • Deep Learning - an MIT Press book in preparation • http://www.iro.umontreal.ca/~bengioy/dlbook/ 50