Обсуждение:Использование технологий NVIDIA для решения задач глубокого обучения

Материал из MachineLearning.

Версия от 00:41, 29 мая 2015; Strijov (Обсуждение | вклад)
(разн.) ← Предыдущая | Текущая версия (разн.) | Следующая → (разн.)
Перейти к: навигация, поиск

Hsung Chan Huang

Shild smartTV device 18000 tryals deployments cars on the road Tesla as a secret computer platform Tesla is supercomputer for AI platform Visual recognition as a superhuman

GTC starts 2008

150000 \ 3 mln Cuds doenloads 27 \ 319 Cuda apps 60 \ 800 Univ teaching 4000 \ 60000 academic papers 6000 \ 450 000 Tesla GPUs

54000 supercomp teraflops

Titan X 8 ln transistors 3072 Cuda cores &Tflops sp / 0&2 Tflops dp 12 GB memory

100 square miles of 3d graphics

50 million plans

physical-looking rendering


MmASXTUIsr7o


Chapert: what tools we can use

What was the movie? Real-time rendering?

Trnaining AlexNet (what is it?) Xeon CPU TITAN TITAN black cuDNN TITAN X cuDNN


See the papers Krizhevsky, Hinton et al 2012 ImageNet classification with NVIDIA GPUs University Toronto Jafferson? lab Alex-net —what is it? Crashed the conuter science in one day Alex Krizhevsky - introduced Alex-Net


LeCun, Bottou, Benigo, Haffer 1998 Convolutional networks for handwritten digital recognition ———————————————— The big bang of computer perception The digits are imprecise 55mln images 27 thousands if categories ——————————

"All we need now is supercomputer" Democtatization Large data Contribution of Alex net

————————Net Output result is histogram ——— Монополия на знания. Тот у кого будет сеть супервосприятия, выиграет рынок. 19 layer-DNN Alex-net

Deep learning visualised (Tuning) back-propagation (good for parallel computing)

VS An idea: repeat Alex-net for time series of human behaviour VS Growth of good models and terminated projects could be used for further training (meta-training)

Andrey Karpathy Fei-Fei-Li CNN - what size? Supervised problem image to its description 100 000 images abd 100 000 sentences 1 week on the GPU

structure earning

Digits DEVBOX MAxGPU out of a plug Multi-GPU training and inference

stochastic gradient descent

Pascal 10X Maxwell Train problem is one of the most challenging problems today and we make it ten times faster.

SENCE FPA CV ASIC PLAN GPU ACT warn brake steer accelerate

we will learn the behaviour drinking all the time

Free space detection for the car problem

how to lean and get the supervision labels as the man acts? The answer is in the example how to teach a baby play pingpong Right and wrong behaviour

225 K images for training data DAVE-DARPA project Inputs are images the outputs are drive commands

Alexnet on Drive PX See silde What can you teach Drive PX todo?


Nvidia Drive PX Self-Drining Car compuer Available May 2015 $10K

Ultrasonic sensors

Government policies without braking the law The car is definitely safer the person 3 years for law regulations

mechanical failure fundamental logic failure multi-car hack additional level security able to penetrate the car…

Qs

Qs on cuDNN, TITAN X, Digits

How to assemble a get-started case for total x What will be the architecture? Will Matlab work? Shall we wait? Show the case. How to assemble a classic recognition model? Will Matlab operate memory normally? Will Python? What are the actual difference between cuDNN, convent 1,2, CNN, BLAS What the actual difference between Teano/Torch/Caffe.

Why do we need cuDNN? What art the capabilities?

Can I read the parameters on the network off? Digits - all stuff preinstalled? What is the configurable digits (build your own DevBox). Why do I need register as cuDNN developer to download cuDNN. Open CL, how Xilinx works with Titan Are there Titan / DevBox solutions on the clouds?

Theano Universite de Montreal

wvzF87DPpDh5 Baidu USA is located at 1050 Enterprise Way, Suite #230, Sunnyvale, CA 94089. https://itunes.apple.com/fr/app/pdf-editor-pro/id422542706?l=en&mt=12


SEE!! Multi-task learning

Aide deep speech splits the time series


Talks

Talks Really inspiring

S5581 - Visual Object Recognition Using Deep Convolutional Neural Networks Rob - super adequate, write him

Use this link for the paper http://devblogs.nvidia.com/parallelforall/accelerate-machine-learning-cudnn-deep-neural-network-library/#more-3632

5 billion connections and four days to lean the model of speech.

Deep speech SWB data set

The next step is how to analyse the paper to make an automatic model construction tool.

Include also Getting started with Torch

Silicon Valley Artificial Intelligence Laboratory

Личные инструменты