情報テクノロジーの進展がもたらす近未来社会の姿を考える AI AI AI: artificial intelligence 1983 8 nontenure-track AI AI AI 1983 AI A* A star Dijkistra Prolog (deductive method) 3 3 A B B C A C 2 3 Prolog 2 2 man(socrates). mortal(x) :- man(x). X X X ":-" " " B:-A A B A B man(socrates). mortal(x) :- man(x). true. mortal(socrates). true. prolog web https://swish.swi-prolog.org/example/
examples.swinb Run! mortal(x). X=socrates mortal socrates. Run! true 1983 Prolog otter) otter Argonne National Laboratory 3 otter 1983 otter 1985 otter otter automated reasoning 3 google otter otter otter 2018 otter otter otter 3 A,B,C) 3 3 A,B,C) 1 "C " A 2 "A " B 3 C "" 3A, B, C otter P(N(B)). P(L(C)). P(T(A)). otter B C
A set(hyper_res). list(usable). -P(T(x)) -P(Says(x,y)) P(y). -P(L(x)) -P(Says(x,y)) -P(y). P(T(x)) P(L(x)) P(N(x)). -P(T(x)) P(L(x)). -P(L(x)) P(T(x)). -P(T(A)) -P(N(B)) P(L(C)). -P(L(A)) -P(N(B)) P(T(C)). -P(L(A)) -P(T(B)) P(N(C)). -P(T(A)) -P(L(B)) P(N(C)). -P(N(A)) -P(L(B)) P(T(C)). -P(N(A)) -P(T(B)) P(L(C)). end_of_list. list(sos). P(Says(A,L(C))). P(Says(B,T(A))). P(Says(C,N(C))). end_of_list. : normal.in 3 A,B,C otter Ubuntu (sudo apt install otter) Ubuntu Ubuntu IoT $ sudo apt install otter 1 $ wget http://web.sfc.keio.ac.jp/~takefuji/ normal.in $ otter <normal.in... ** KEPT (pick-wt=6): 51 [hyper,49,1,13] P(N(B)) P(T(A)). given clause #17: (wt=6) 50 [hyper,48,4] P(N(B)) P(L(C)). ** KEPT (pick-wt=3): 52 [hyper,50,2,14,22] P(N(B)). inductive method) (ensemble machine learning) deep learning 100 100
CPU GPU Adaboost RandomForest ExtraTree ExtraTrees GradientBoosting Bagging VotingClassifier artificial neural network ANN functional-link net 1992 IEEE Computer, 25, 5, pp.76-79 pdf http://neuro.sfc.keio.ac.jp/publications/ pdf/paocomputer.pdf GPU GPU CPU CPU Central Processing Unit 2 4 GPU GPU Graphics Processing
Unit GTX1080 Ti GPU CUDA 3584 GXT1080 Ti 1 10 2018 GTX 2080 Ti GPU 5376 CUDA CUDA Compute Unified Device Architecture 3584 3600 3600 60 GPU NVIDIA AI Python 1 Google Tensorflow 2 Keras 3 PyTorch Chainer 5 Caffe darknet YOLO You only look once YOLO C YOLO 1 NVIDIAAI GPU AI NVIDIA GPU NVIDIA CUDA toolkit NVIDIA CUDA toolkit https://www.zoo.org.au/melbourne/ wild-encounters/close-up-encounters
YOLO person 3 giraffe) bottle bottle: 25% 174, 270, 198, 297 person: 77% 34, 77, 131, 230 person: 35% 124, 96, 217, 303 person: 84% 50, 130, 194, 378 giraffe: 88% 174, 0, 619, 264 GPU YOLO x1, y1, x2, y2 car truck left top x1, y1 right bottom x2, y2 x, y 0,0 C http://www.c-kumiai.net/service/ warehouse.html car: 31% 513, 235, 556, 266 truck: 65% 154, 207, 308, 364 YOLO Wikipedia : Simulacra 3 YOLO
simulacra door dog 3 google YOLO YOLO 32 sigmoid YOLO :conv :max :route reorg :reorg https://soranews24.com/2014/12/07/dog-seems-toappear-from-plank-of-wood-in-either-coolest-pettrick-or-silliest-astral-projection/ YOLO dog GPU 20 30 GPU 1 YOLO YOLO YOLO YOLO layer filters size input output 0conv 32 3x3/1 608x608x 3 -> 608x608x 32 1max 2x2/2 608x608x 32 -> 304x304x 32 2conv 64 3x3/1 304x304x 32 -> 304x304x 64 3max 2x2/2 304x304x 64 -> 152x152x 64 4conv 128 3x3/1 152x152x 64 -> 152x152x 128 5conv 64 1x1/1 152x152x 128 -> 152x152x 64 6conv 128 3x3/1 152x152x 64 -> 152x152x 128 7max 2x2/2 152x152x 128 -> 76x76x 128 8conv 256 3x3/1 76x76x 128 -> 76x76x 256 9conv 128 1x1/1 76x76x 256 -> 76x76x 128 10conv 256 3x3/1 76x76x 128 -> 76x76x 256 11max 2x2/2 76x76x 256 -> 38x38x 256 12conv 512 3x3/1 38x38x 256 -> 38x38x 512 13conv 256 1x1/1 38x38x 512 -> 38x38x 256 14conv 512 3x3/1 38x38x 256 -> 38x38x 512 15conv 256 1x1/1 38x38x 512 -> 38x38x 256 16conv 512 3x3/1 38x38x 256 -> 38x38x 512 17max 2x2/2 38x38x 512 -> 19x19x 512 18conv 1024 3x3/1 19x19x 512 -> 19x19x 1024 19conv 512 1x1/1 19x19x 1024 -> 19x19x 512 20conv 1024 3x3/1 19x19x 512 -> 19x19x 1024 21conv 512 1x1/1 19x19x 1024 -> 19x19x 512 22conv 1024 3x3/1 19x19x 512 -> 19x19x 1024 23conv 1024 3x3/1 19x19x 1024 -> 19x19x 1024
24conv 1024 3x3/1 19x19x 1024 -> 19x19x 1024 25route 16 26conv 64 1x1/1 38x38x 512 -> 38x38x 64 27reorg /2 38x38x 64 -> 19x19x 256 28route 27 24 29conv 1024 3x3/1 19x19x 1280 -> 19x19x 1024 30conv 425 1x1/1 19x19x 1024 -> 19x19x 425 31detection mask_scale: Using default '1.000000' Loading weights from yolo.weights...done! dogt1.jpg: Predicted in 0.017147 seconds. dog: 58% 71, 12, 339, 318 amazon.co.jp kindle GTX 1080 Ti GPU 25 30 GPU " " GPU X y y=f(x) f f x1 x2 http://www.ohmidog.com/tag/simulacra/ GPU GPU parallel computing for machine learning in Python 2017 temp, street, c
y=temp*x1+street*x2+c n) Y Y=f(x1, x2,..., xn) Y=y1, y2,..., ym " GPU CPU scikit-learn scikit-learn 8 1 2 SPT 3 4 5 ABC 6 7 8 185 Python11
Python 148 149 Python http://web.sfc.keio.ac.jp/~takefuji/face.py count.jpg from skimage import io, filters from scipy import ndimage import matplotlib.pyplot as plt from skimage import measure im = io.imread('count.jpg', as_grey=true) val = filters.threshold_otsu(im) drops = ndimage.binary_fill_holes(im < val) labels = measure.label(drops) plt.imshow(drops, cmap='gray') print(labels.max()) plt.show() 1989
"Science Nature Science 19 20 Science 1989 9 15 http://neuro.sfc.keio.ac.jp/publications/ pdf/science.pdf MRI Science MRI IEEE Trans. on Medical Imaging, 11, 2, 215-220, 1992 http://neuro.sfc.keio.ac.jp/
publications/pdf/sundar.pdf BV-6000) EURO2002 1 1 EURO 201210 12000 2003 http://neuro.sfc.keio.ac.jp/publications/ pdf/banknote.pdf EURO AI AI GPU AI MIT Stanford CMU BV-6000 EURO EURO EURO Google, IBM AI AI NIPS 2017) 6 6 140 AI
365 AI 6 http://neuro.sfc.keio.ac.jp/publications/ pdf/quiz.pdf https://github.com/mk-takizawa/elmo chess https://arxiv.org/pdf/1712.01815.pdf 2 agent environment environment state reward agentagent actionaction environment state 6 vs AI AI reinforcement learning AI 2017 AI elmo 2500 USI reward action agent environment AI" AI AI 1
AI AI 2 AI agent environment AI 3 AI AIAI AI 4 AI AI 5 AI AI 6 AI AI AI AI AI AI 1) tenure tenure
15 36 tenure tenure 3 2) A* gif https://en.wikipedia.org/wiki/a*_search_ algorithm 3) Dijkistragif https://en.wikipedia.org/wiki/dijkstra%27s_ algorithm 1955 1983 1991 36tenured 1997!! NanoPi NEO, Arduino IoT 2017 2016 2013