_KrlGF발표자료_AI

Size: px

Start display at page:

Download "_KrlGF발표자료_AI"

인석 즙
5 years ago
Views:

1 AI 의과거와현재그리고내일 AI is the New Electricity

2 AI! 2

3 Near Future of Super Intelligence? *source l 3

4 4

5 I think that we live in a world where just as electricity transformed almost everything almost 100 years ago, Today I actually have a hard time thinking of an industry that I don't think AI will transform in the next several years 100 AI. 5

6 Bill Gates to college grads I expect AI to create breakthroughs that makes people better learners. Get a job in AI, but don t forget inequity around you. AI. AI,. 6

7 AI is a technology that gets so close to everything we care about. It s going to carry the values that matter to our lives, be it the ethics, the bias, the justice, or the access. AI.,,,,. 7

8 ( Positive & Negative ) vs, vs vs Fake/Garbage vs 24hr Everything 8

9 1. AI 2. AI? - 3. State - of - Arts 4.

10 AI Breakthrough *source l 10

11 AI *source l 11

12 Level - 1. If ~ then ~ 12

13 Level - 2. "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E. by Tom M. Mitchell Classification Regression Clustering *source l 13

14 Level - 3. Deep Neural Networks (DNN) DNN, AI *source l 14

15 AI Next 15

16 1. AI 2. AI? - 3. State - of - Arts 4.

17 AI AI Answer : it's a girl is brushing her hair. *source l (MS ) 17

18 ?? 18

19 ? 19

20 AI Option Pricing. AI. 20

21 ? 21

22 ? Darpa Robot Challenge

23 AI AI 23

24 AI? AI 24

25 AI = ( )?? [ ] [ ] [ ]. 25

26 F 26

27 (Machine Learning)? (T, Task) (P, Performance measure) (E, Experience). 27

28 ? AI,! 28

29 ? ( ) (Supervised Learning) >> >> SL >> ( Google -> Kaggle, DR ) (Unsupervised Learning) >> >> ( : ) >> ) Clustering (Reinforcement Learning) >> >> >> ) 29

30 -? *Video l 1m40s 30

31 (Machine Learning) (Artificial Neural Network) (Deep Learning) 31

32 AI Neural Information Processing Systems International Conference on Machine Learning 2012 AI breakthrough 2012 ImageNet 32

33 AI - Y. LeCun IBM, Google, MS G. Hinton Google Y. Bengio Facebook Andrew Ng Baidu (ex) " (Neural Network)... by Yoshua Bengio 33

34 AI - (ICML, NIPS) 34

35 AI? - 35

36 (ANN) (Neuron) (Synapse) 36

ANN 28 x 28 = 784 0~9 = 10 Labels Output 0 0.1 1 0.05 2 0.3 3 0.

37 ANN 28 x 28 = 784 0~9 = 10 Labels Output x

ANN Output Labels Input 0.1 0 0 0.05 1 0 0.3 2 0 28 x 28 = 784 0.2 0.41 3 0 4 0 0.

38 ANN Output Labels Input x 28 = RGB 28 x 28 x 3 = 2, Label Output Input Weight 38

39 ANN

40 Train : 94.6%, Test : 87.7% 40

41 ANN ANN? 28 x 28 = 784? Input Data Parameter Linear Combination Output Parameter 41

42 DNN (Deep Neural Network)? Hidden Layer 2 Deep Layer 2012 SuperVision ,

43 DNN? Linear Equation Combination Output Parameter 43

44 Deep Learning (1) - Overfitting Data Over Fitting Data Fitting 44

45 Deep Learning (2) - Vanishing Gradient Signal 45

46 Deep Learning (3) - Local Minimum 46

47 Deep? Not Deep ImageNet Label Layer Deep,! 47

48 Deep? Performance Amount of data *source l Andrew Ng Deep Learning Lecture Note 48

49 DNN (Deep Neural Network)? 2012? Nvidia GPU Tesla (miles) Computer Processing Power GPU *source l (Tesla ) 49

50 Nvidia GPU Computing Power Nvidia

51 Google Cloud Machine Learning 1,000 GPU 25 51

52 Google I/O 17 Chip ( TPU ) 52

VOC 2K 30K objects 2010-2015 Image Net 10M

53 Data? Critical Mass Deep Learning 2004 Caltech K Images Pascal VOC 2K 30K objects Image Net 10M 15M images Image source: Image source:

54 Data? A B C Deep Learning Neural Networks Log. Regression Deep learning K-Nearest neighbors Support vector machines Boosting Artificial neural networks Bayesian networks Sparse dictionary learning Regression forest 54

55 1. AI 2. AI? - 3. State - of - Arts 4.

56 Quality Data Labeled Data ( Supervised Learning ) 56

57 Input Output Application Home Features Ad, User Info Image Audio English Image, Radar, Sign Price Click Ad? (0/1) Object( cat, dog) Text trasnscipt Chiness People, Car Real Estate Online Ad Photo Tagging (name) Speech Recognition Translation Autonomous Car 57

58 TTS AI *source l 58

59 DeepMind WaveNet Google DeepMind TTS 1/1000 TTS WaveNet *source l 59

60 Music Generation Daddy s Car, a pop song in the style of The Beatles. *source l *Video l 3m00s 60

61 Recognizing Pain ARTIFICIAL INTELLIGENCE COULD END ANIMAL SUFFERING BY RECOGNIZING PAIN *source l 61

62 사진으로 나이 알아 내기 *source l (by MS제공 ) 62

63 *source l 63

64 이미지 설명하기 *source l 64

65 Breast Cancer Detection Google uses machine learning to detect breast cancer better than pathologists *source l 65

66 Image Generation (1) [ ] AI [ ] [ ] AI [ ]? 66

67 Image Generation (2) - GAN GAN(Generative Adversarial Network) g-net : d-net : *source l NIPS

68 Image Generation - 16 x x 64 (x4) Augment *source l 68

69 1. AI 2. AI? - 3. State - of - Arts 4.

70 GAFA (Google, Amazon, Facebook, Apple) AI 70

71 AI 71

72 AI in 2017 Tech giants acquired 34 AI startups in Q AI 17 1 Amazon $ 19M buy harvest.ai (AI-security) 17 2 Ford $1B buy Argo ( ) 17 5 Apple $200M buy Lattice Data ( ) 17 5 Cisco $125M buy Mindmeld ( ) > AI *source l 72

73 AI 73

74 AI? 1 AI AI 3 Domain API, 2 Data Garbage In -> Garbage Out Big Data 4 Computing Power Trial & Error

75 AI 1. ( SL ) 2. ( ) AI. 75

76 Tesla Real Time - Data Processing 76

77 AI KAKAO AI REPORT import kakao.ai.dataset.daisy import kakao.ai.image import kakao.ai.classifier import mxnet as mx def Conv(data, num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0), name=none, suffix=''): conv = mx.sym.convolution(data=data, num_filter=num_filter, kernel=kernel, stride=stride, pad=pad, no_bias=true, name='%s%s_conv2d' %(name, suffix)) bn = mx.sym.batchnorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=true) act = mx.sym.activation(data=bn, act_type='relu', name='%s%s_relu' %(name, suffix)) return act def Inception7A(data, num_1x1, num_3x3_red, num_3x3_1, num_3x3_2, num_5x5_red, num_5x5, pool, proj, name): tower_1x1 = Conv(data, num_1x1, name=('%s_conv' % name)) tower_5x5 = Conv(data, num_5x5_red, name=('%s_tower' % name), suffix='_conv') tower_5x5 = Conv(tower_5x5, num_5x5, kernel=(5, 5), pad=(2, 2), name=('%s_tower' % name), suffix='_conv_1') tower_3x3 = Conv(data, num_3x3_red, name=('%s_tower_1' % name), suffix='_conv') tower_3x3 = Conv(tower_3x3, num_3x3_1, kernel=(3, 3), pad=(1, 1), name=('%s_tower_1' % name), suffix='_conv_1') tower_3x3 = Conv(tower_3x3, num_3x3_2, kernel=(3, 3), pad=(1, 1), name=('%s_tower_1' % name), suffix='_conv_2') pooling = mx.sym.pooling(data=data, kernel=(3, 3), stride=(1, 1), pad=(1, 1), pool_type=pool, name=('%s_pool_%s_pool' % (pool, name))) cproj = Conv(pooling, proj, name=('%s_tower_2' % name), suffix='_conv') concat = mx.sym.concat(*[tower_1x1, tower_5x5, tower_3x3, cproj], name='ch_concat_%s_chconcat' % name) return concat def Inception7B(data, num_3x3, num_d3x3_red, num_d3x3_1, num_d3x3_2, pool, name): tower_3x3 = Conv(data, num_3x3, kernel=(3, 3), pad=(0, 0), stride=(2, 2), name=('%s_conv' % name)) tower_d3x3 = Conv(data, num_d3x3_red, name=('%s_tower' % name), suffix='_conv') tower_d3x3 = Conv(tower_d3x3, num_d3x3_1, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name=('%s_tower' % name), suffix='_conv_1') Vol AI CODE int FaceAlignment::track(unsigned char *src, int width, int height, int rotation) { m_img.img.wrap(height, width, src); m_img.gt_enabled = false; for (int i = 0; i < m_face_cnt; i++) { } if (m_data[i].enabled == false) // pass if it is not valid continue; // connect image to data DataInfo *data = &m_data[i]; data->img_ptr = &m_img; // compute bounding box of current face MyRect<float> bbox = FAUtil::compute_bbox(data->current_pts); data->fd = bbox; // put mean shape to the face in image data->current_norm_pts = mean_shape; data->q = FAUtil::compute_similarity(data->current_norm_pts, data->current_pts); data->inv_q = FAUtil::compute_inverse_similarity(data->q); Vol data->current_pts = FAUtil::apply_similarity_transform(data->current_norm_pts, data->q); data->initial_pts = data->current_pts; // DO align align(data); AI CODE vector<pairif_t> c_dmf_topk( ) { Map<const VectorXf> vv(v, M), AA(A, N); Map<const MatrixXf> XX(X, N, M); vector<pair<int, float>> result; RowVectorXf dot = hidden_layers[0].transpose() * vv; for(int j=1; j < (int)hidden_layers.size(); ++j) dot = dot * hidden_layers[j]; dot = dot.array().max(0.0f); VectorXf ex = fast_softmax(xx, dot, AA); get_fast_knn(ex, k, result); return result; } vector<pairsf_t> prediction_by_vector(const float* v, int k) { auto ret = pool_.enqueue([](const float* _v, vector<factortype>& hidden_layers, const float* sfxb, const float* sfx, int N, int M, int _k){ return c_dmf_topk(_v, hidden_layers, sfxb, sfx, N, M, _k); }, v, hidden_layers_, (const float*)softmax.data(), k); auto val = move(ret.get()); vector<pairsf_t> result = as_result(val); return result; } Vol tower_d3x3 = Conv(tower_d3x3, num_d3x3_2, kernel=(3, 3), pad=(0, 0), stride=(2, 2), name=('%s_tower' % name), suffix='_conv_2') pooling = mx.symbol.pooling(data=data, kernel=(3, 3), stride=(2, 2), pad=(0,0), pool_type="max", name=('max_pool_%s_pool' % name)) concat = mx.sym.concat(*[tower_3x3, tower_d3x3, pooling], name='ch_concat_%s_chconcat' % name) return concat } // validate aligned face return validate(src, width, height, rotation); 77

78 :, Question

79 End of Document

KAKAO AI REPORT Vol.01

KAKAO AI REPORT Vol.01 2017.03 import kakao.ai.dataset.daisy import kakao.ai.image import kakao.ai.classifier import mxnet as mx def Conv(data, num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0), name=none,