00:00:09I’m Kaz Sato I’m a developer advocate
00:00:12for Google cloud platform and I’m also a
00:00:14tech lead for the data analytics team
00:00:16and I have been working at Google for
00:00:20over five years and for the last one and
00:00:23a half years I have been working as a
00:00:25developer advocate to having a
00:00:29presentation like this at many events so
00:00:34in these sessions I’d like to talk about
00:00:37these agendas but first I’d like to
00:00:40introduction introduce the concept of
00:00:42neural network and deep learning and how
00:00:44it works with very great demonstrations
00:00:48and also I’d like to introduce how
00:00:52Google has been deploying those neural
00:00:55network technologies to the Google crowd
00:00:57and then I’ll be covering the
00:01:00technologies and products that those are
00:01:04actually provided as a products from
00:01:06Google cloud platforms so what what is
00:01:11neural network so neural network is a
00:01:15function that can learn from training
00:01:17data set so if you want to have newer
00:01:20networks to do image recognition then
00:01:24you can put the for example cat image is
00:01:27converted into a load vector and then
00:01:31put that vector into the networks then
00:01:34eventually you’d have another output
00:01:36vector which represents the labels of
00:01:39the objects detect such as cat or human
00:01:42face so it’s designed to mimic the
00:01:47behavior of new ones inside human brains
00:01:50by using matrix operations so actually
00:01:54it’s really it has very basic matrix
00:01:57operations only there is no
00:01:59sophisticated or fancy mathematics going
00:02:01on everything you do with neural
00:02:04networks is you have if you matrix
00:02:06operations we have learned at high
00:02:07school for example the input vector may
00:02:12we present represent the image of cat
00:02:15then they will have a vector that would
00:02:17have the pixel data converting it to a
00:02:20vet
00:02:20and you would get another output vector
00:02:23as a result that represents the label of
00:02:27the detected images in in this case you
00:02:30will have the any number closer to the
00:02:331.0 that indicates the neural networks
00:02:37thinks the image must be a cat but you
00:02:41think’s going on inside in your networks
00:02:44is very simple all you are doing with
00:02:48neural network is the matrix operations
00:02:49like W X plus B equals wise where the WS
00:02:54are weights and B are biases and
00:02:58actually you don’t have to care about
00:03:00those WS and B’s at all you let
00:03:03computers find and calculate the WS and
00:03:07B’s so all you have to care is the yeah
00:03:10what kind of data you would want to put
00:03:13into the networks and what kind of the
00:03:15result you want to get so let’s take a
00:03:20look at the some interesting
00:03:22demonstration of neural networks for
00:03:24example if you have a problem like this
00:03:26the very simple classifications how you
00:03:29can use neural networks to do the
00:03:31classifications of these two different
00:03:33data sets I’m not sure what that those
00:03:37data point means but let’s take a lift
00:03:41imagine that those are the data of the
00:03:44weights and height of people so that if
00:03:48a person’s weight and height are both
00:03:51loud and you can think he or she must be
00:03:54a child or maybe he or she could be a
00:03:58adult how how can you classify them if
00:04:03you using you if you try to use neural
00:04:06networks to do this problem then you can
00:04:08just apply the same equation W X plus B
00:04:12equal Y to do you’ve to do two
00:04:14classifications so you are putting the
00:04:17weight number and height number here as
00:04:19a vector and you get an output a vector
00:04:21like this where if you if what single
00:04:25data points isn’t classified as an adult
00:04:28then you would have one here and if it’s
00:04:31a child then you would have a one here
00:04:34and thing is that computer tries to find
00:04:38the optimal combination of the
00:04:39parameters such as the weights and
00:04:41biases by itself and you don’t have to
00:04:44think about what kind of parameters you
00:04:47have to set to neural networks computer
00:04:49does it for you so let’s take a look at
00:04:52the actual demonstration did you see
00:04:56that so I can do that okay so instead of
00:05:00the humans letting computer instructing
00:05:04learning you know you don’t have to
00:05:08teach computers how to solve these
00:05:10problems but everything you have to do
00:05:12is provide the training data set so that
00:05:15computer thinks by itself to optimize
00:05:18the combination of the parameters such
00:05:21as weights and biases so you see that
00:05:25the computer is try to change the the
00:05:29weights to do the classifications at
00:05:32optimal success rate and now computer is
00:05:40using an algorithm is called gradient
00:05:43lee set that means it tries to increase
00:05:46or decrease each weights and biases to
00:05:50make the combinations closer through the
00:05:54higher accuracy or lower loss slate so
00:05:59it’s just like we’re learning things
00:06:01from the the parents or there may be a
00:06:05senior people in your company where you
00:06:07this junior people or these children are
00:06:11learning from many many mistakes so
00:06:14computer makes many many mistakes but in
00:06:16the initial stage but if you provide
00:06:19much much more training data set than
00:06:22computer using the gradient descent
00:06:25algorithms try to minimize the failures
00:06:28so that that’s how it works so let’s
00:06:31take a look at another interesting
00:06:34demonstrations of new or networks where
00:06:36you have the another training dataset
00:06:38like this I’m not sure what what does it
00:06:41mean but we think that we have a some
00:06:46data set which requires a complex
00:06:48classifications if you a program has
00:06:50maybe with this data setting you may
00:06:53want to user maybe equations for circle
00:06:55to classify the two datasets through
00:06:58datasets and arrange a disease but by
00:07:01using newer networks you can just let
00:07:04computer things how to solve it now you
00:07:09saw the computer was trying to create a
00:07:15pattern to classify those datasets by
00:07:19using so called neurons in the hidden
00:07:21layers with the first example we didn’t
00:07:27have any hidden layers but with this
00:07:29complex data that you would need to use
00:07:31the hidden layers that means between the
00:07:35input data and the output neurons you
00:07:38would have another layers between them
00:07:41there has much pool new ones and each
00:07:45new ones does very simple things these
00:07:48neurons only classifies whether the data
00:07:51points is in the in the bottom left red
00:07:55area or the upper right area or now
00:07:58these neurons only classifies whether
00:08:01the data points is in the left or right
00:08:03just like that but combining those
00:08:05outputs at the the last new one neural
00:08:09network then the neural networks can
00:08:11compose much more complex pattern like
00:08:14this and if you have more and more new
00:08:16neurons inside the hidden layers then
00:08:18the network specifications speculations
00:08:23can be much more accurate like this so
00:08:31here by adding more hidden layers with
00:08:33more new ones you have to spend much
00:08:36more computation power but at the same
00:08:38time the neural networks can compose
00:08:40much more complex patterns and extract
00:08:44the patterns from the large data set how
00:08:48about this let’s try it this is a data
00:08:53pattern called double spiral if you are
00:08:55a programmer and your director or cast
00:08:58asked you to cross fight this kind of
00:09:01data set what kind of your program
00:09:03called you drive do you want to write
00:09:05many issue statements or switch
00:09:07statements with many threshold try
00:09:09checking the x and y partitions no I
00:09:12don’t want to do that instead I would
00:09:14you be using neural networks so that
00:09:17neural networks try to looking at the
00:09:19data points in the data sets to find the
00:09:23optimal patterns hidden inside the
00:09:25training data set that’s right this so
00:09:28this is the where in your networks can
00:09:30exceed the human performance human
00:09:32programmers performance it can extract
00:09:35those hidden patterns inside the
00:09:38training data set and you don’t have to
00:09:42specify anything decorative features you
00:09:46don’t have to any actual features by
00:09:50human hand instead computers can find
00:09:52the patterns from the training data set
00:09:54so that why people are so excited with
00:09:57the neural networks and that deep
00:09:58learning about this so if you have the
00:10:04problems of the identifying handwriting
00:10:07text you can still using the very simple
00:10:10GW Express peep equals why kind of
00:10:13neural networks to cross find this
00:10:15handwriting takes the networks would
00:10:18come up with disease complex patterns
00:10:21stacking classify those images into the
00:10:24year vectors with the labels like an
00:10:28eight or seven or six if you want more
00:10:31accuracy than you would have more no
00:10:33more hidden dangers so that you you
00:10:35could get like a 85 or 95 or 98 like a
00:10:39machine how about this how can you
00:10:42classify these cat images by using
00:10:44neural networks you have to have many
00:10:47more layers of neural networks that is
00:10:52so-called deep neural networks
00:10:54this isn’t diagram is called inception
00:10:56model pattern 3 there has been published
00:11:00by Google last year where we have used
00:11:03240 hidden layers in a single neural
00:11:06network design so it takes much more
00:11:09computation power and time but still we
00:11:12can
00:11:12much more complex competitions like this
00:11:14you know the new ones closer to the
00:11:17input vector could learn very simple
00:11:20pattern rugby’s like you know vertical
00:11:23lines or the horizontal rise but the
00:11:26neurons closer to the output vector
00:11:30could learn much more complex patterns
00:11:33or compositions such as eyes nose or
00:11:35human face again we didn’t put any
00:11:40features of patterns embedded in the
00:11:44neural networks before training did so
00:11:46everything can be trained from the data
00:11:49set
00:11:51yeah by using computation power so
00:11:55that’s the how neural network and deep
00:11:57learning works but as I mentioned it
00:11:59takes so much computation time and
00:12:01training data sets to use the planning’s
00:12:05for the production projects so there are
00:12:07two big challenges right now for the
00:12:09users for the deep runnings and this is
00:12:11why keep running is has not been so
00:12:14popular for you guys once we have solved
00:12:18these problems raka will have a plenty
00:12:21of computation power with the printed
00:12:23training data set then you can easily
00:12:26apply neural network so deep learning to
00:12:28your existing problems if you a game
00:12:30programmer you may want to apply the
00:12:32deep learnings to analyzing gear you are
00:12:35your game log server logs to check
00:12:38whether a player could be an cheatin
00:12:40player or spammer office weather or if
00:12:43you web designer or web systems engineer
00:12:49for the ad system then you may want to
00:12:51apply the logs for the as conversion or
00:12:55quick-quick flow rate our logs to neural
00:12:58network so that you can get you can have
00:13:00computers to learn from the year as log
00:13:03to get more optimization but you have to
00:13:08have computation power and training data
00:13:11so that’s the reason why we have started
00:13:14using Google cloud to train in large
00:13:17scale neural network Google cloud has
00:13:21racket and
00:13:23hundreds of thousand machines in our
00:13:26data centers in global and we have been
00:13:28building those computers at the data
00:13:31center as a computer not just a bunch of
00:13:33the computers Switzer building we design
00:13:36each Craster which holds like a ten or
00:13:39twenty thousand servers working as a
00:13:42single computer with a multiple our CPUs
00:13:46so that’s the reason why we can it’s
00:13:50it’s not so hard for us to deploy try to
00:13:53scale neural networks or odds can be
00:13:54created Processing’s to our google cloud
00:13:57if there are two basic very fundamental
00:14:02technologies inside google’s that
00:14:03supports the data center the computer
00:14:05one is the network we have been building
00:14:08our own hardware support for the network
00:14:12switch fabric that is called jupiter
00:14:15networks so we union we are not using
00:14:19the commercial network switches for most
00:14:22cases such as the Cisco or juniper
00:14:24routers those are not mainstream of our
00:14:29new our network backbones we we have
00:14:33been building our own hardware that can
00:14:35hold like a hundred thousand pots of 10
00:14:39Gigabit Ethernet ports that can eat at
00:14:43one point to pet a bit per second per
00:14:45our datacenter so that is the networks
00:14:48we have at Google and also contain a
00:14:51technology called Borg bo is our
00:14:53proprietary container technologies we
00:14:56have been using over 10 years for for
00:14:58deploying or almost all Google services
00:15:00such as Google search or Gmail or Google
00:15:03Maps
00:15:03bulk containers account hold up to
00:15:0710,000 or 20,000 physical servers in a
00:15:10single cluster so that you can do the
00:15:13large scale job scheduling right the
00:15:15scheduling the CPU cycles or memory
00:15:17spaces or disk i/os with that scale so
00:15:21that reason why you can deploy your
00:15:23single applications like did the neural
00:15:28network training or the big data
00:15:30processing into maybe hundreds or
00:15:32thousands of machines with a single land
00:15:35of default command
00:15:37and Google brain is the project where we
00:15:42have started applying the Google cloud
00:15:44technology 2d to build a large-scale
00:15:47neural networks this project has started
00:15:49in 2011 and right now the Google brain
00:15:54has been used for the many many
00:15:55production project in Google and fast
00:15:58the scalability of Google Google brain
00:16:00project for example rankbrain rankbrain
00:16:03is our GU algorithms we are using for
00:16:05the ranking of Google search service
00:16:07right now since last year that has been
00:16:11using Google Google brain infrastructure
00:16:14and with five hundred nodes and that can
00:16:19perform at three hundred times faster
00:16:22than single node so that means if you
00:16:26are training your deep learning model
00:16:29with single servers then you would take
00:16:32300 times longer than Google engineers
00:16:36and inception is the model for the
00:16:40visual visual recognition we can use 50
00:16:43GPUs to accelerate the performance at 40
00:16:47times faster so that reason those are
00:16:50the reason why Google has been so strong
00:16:52on applying deep running for the
00:16:54production project such as the alphago
00:16:58Frady we have the series of the core
00:17:02matches with the core professional they
00:17:05have been using the Google brain
00:17:06infrastructure for the training as well
00:17:09as the prediction of the Google match
00:17:12Google search has been using deep with
00:17:14the brain of since last year and we have
00:17:17been using the machine learning
00:17:19technologies for the optimizing the data
00:17:21center operation and also or she our
00:17:24natural language processing and visual
00:17:26recognition of speech recognition such
00:17:28as the Google photos what he voiced the
00:17:30conventions of the androids we have over
00:17:3360 production projects that has been
00:17:36using Google brain and deep learnings
00:17:38for last a couple of years now we have
00:17:44started to externalizing this power of
00:17:47Google brain to external developers
00:17:51the first product is called crowd vision
00:17:54API and the second product is called
00:17:57crowd speech API crowd vision API is an
00:18:02image analysis IPA that provides the
00:18:05pre-trained model so you don’t have to
00:18:08train your own neural network and you
00:18:11also don’t have to have the any skill
00:18:13set for the machine learning so it’s
00:18:16just on REST API you can just upload
00:18:18your photo image to API then you repeat
00:18:22receiving JSON result in a few seconds
00:18:24there has the the analysis result and
00:18:27it’s free to start trying out up to
00:18:301,000 images per month and it’s general
00:18:34generally available right now so it’s
00:18:36ready to be used for the production
00:18:37project it has six different features to
00:18:42be detected labial detections means that
00:18:45you can put any labels or categories on
00:18:47any images you uploaded for example if
00:18:51you uploading the cat images then the
00:18:53API will be returning the Arabians such
00:18:56as a cat or pet face detections can
00:19:01detect the location of face in the image
00:19:03OCR I can convert the text on image to a
00:19:07string explicit content detection means
00:19:10that you can check whether the images
00:19:13can contain the images contain the the
00:19:16adult or violent images on our landmark
00:19:19detection can detect the location of the
00:19:21images or popular places and you can
00:19:23also detect the product or corporate
00:19:26role let’s take a look at the
00:19:28demonstration
00:19:35so I’d like to show a demonstration by
00:19:38video at first this is the
00:19:41demonstrations by using the Raspberry Pi
00:19:43robot that sends the image to division
00:19:47API cloud vision provides powerful image
00:19:53analytics capabilities as easy to use
00:19:55api’s it enables application developers
00:19:59to build the next generation of
00:20:00application that can see and understand
00:20:03the content within the images the
00:20:05service is built on powerful computer
00:20:07vision models that power several to firm
00:20:10Google services the service enables
00:20:13developers to detect a broad set of
00:20:15entities within an image from everyday
00:20:17objects to faces in product logos the
00:20:20service is so easy to use as one example
00:20:23of the use cases you can have any
00:20:26Raspberry Pi robot like gulp I go
00:20:28calling the cloud vision API directly so
00:20:32the broad can sum the images taken by
00:20:33its camera to the cloud and can get the
00:20:36analysis results in real time it detects
00:20:39faces in the image along with the
00:20:41associated emotions the cloud vision API
00:20:43is also able to detect entities within
00:20:46the image now let’s see how facial
00:20:49detection works cloud vision detect
00:20:52spaces on the picture and returns the
00:20:53positions of eyes nose and mouth so you
00:20:57can program the bot to follow the face
00:21:04it also detects emotions such as joy
00:21:07anger surprise and sorrow so the bottom
00:21:11moved toward smiling faces or avoid
00:21:13anger or surprise face one of the very
00:21:17interesting features of cloud vision API
00:21:19is the entity detection that means it
00:21:22detects any objects you like you see
00:21:42cloud visitors likes developers to take
00:21:45advantage of Google’s latest machine
00:21:47learning technologies quite easily
00:21:48please go to cloud.google.com slash
00:21:51vision to learn more and I have another
00:21:58interesting demonstrations that is made
00:22:01by using the vision API that if this is
00:22:05called vision Explorer demonstrations
00:22:08where we have imported 80,000 images
00:22:11downloaded from Wikimedia Commons and
00:22:14uploaded to the Google Cloud storage and
00:22:17applied the vision API analysis so here
00:22:21we have the cluster of the images it is
00:22:2480 thousand images and each cluster has
00:22:28the labels such as snow or transport
00:22:32residential area means that the the
00:22:36cluster of the similar images for
00:22:38example if you take a look at here let’s
00:22:40go to a plant so each single dot
00:22:44represents your thumbnail of the
00:22:47uploaded images so if you go to the
00:22:49plant
00:22:50Craster there must be some cluster of
00:22:54the oh it’s oh it’s not showing why let
00:23:01me redraw this maybe because I’m using
00:23:05tethering
00:23:09okay let’s go directly to the cat
00:23:11cluster so in this cluster we have many
00:23:20many cats and closer to the cat cora’s
00:23:22we have the crust for dogs let’s go back
00:23:27to the cat cluster and if you click to
00:23:31image thumbnail image and you’ll be
00:23:34seeing the analysis result from the API
00:23:37right this the API thinks this must be a
00:23:40mimic of cat and it’s a cat as a pet or
00:23:45it must it must be a British Shorthair
00:23:47so this is these some things you can do
00:23:50with the deep learning technology and
00:23:52those results are returned in a JSON
00:23:55format erectus and with statistical
00:23:57stretches we can show it in a GUI if
00:24:03image contains any text inside it then
00:24:05we can come back convert it into the
00:24:07extreme example with this you can have
00:24:13the string like these three Kangaroos
00:24:15crossing next to you images if images
00:24:19contains faces also this API doesn’t
00:24:23support any personal identification or
00:24:26the personal recognition but it can
00:24:29detect the location of the faces with
00:24:32landmark locations such as nose and
00:24:35mouth and also it can recognize the
00:24:39emotions such as joy sorrow and anger
00:24:42and surprising in this confidence level
00:24:47and if your picture contains any popular
00:24:53places such as this dinner API can
00:24:57return the name of the landmark the API
00:25:01thinks it must be an image of the Citi
00:25:03Field Stadium in New York City with the
00:25:06longitude and latitude so you can easily
00:25:10in a put a marker on the Google Maps
00:25:12it’s too slow so I’m cutting it off also
00:25:15you can detect the product and corporate
00:25:18robot
00:25:21like this Android so this was the this
00:25:30vision API so it’s ready to be used for
00:25:33any applications and another API is
00:25:37called speech API which also provides
00:25:40the pre-trained model for the voice
00:25:42recognition so you don’t have to have
00:25:44any skill set or experiments with the
00:25:47voice recognition or training neural
00:25:49networks for doing doing that it’s just
00:25:51on REST API and G RPC API so you can
00:25:55just upload your audio data to the API
00:25:58and you’ll be receiving a result in the
00:26:00few seconds it supports over 80
00:26:02languages and dialects it supports both
00:26:05real-time recognition and battery
00:26:07recognition the API is still in limited
00:26:10preview so if you go to the speech
00:26:13cloud.google.com speech then you have to
00:26:16sign up with the form for immediate
00:26:19limited preview access but we hope to
00:26:22make it public better maybe in a couple
00:26:25of weeks I suppose let’s show some
00:26:30demonstration I’m not sure if this works
00:26:35in the event or not because this is the
00:26:38first time to try this and and I have
00:26:41some accent problems so I’m not sure I
00:26:43really not sure if this works or not but
00:26:45best right hello this is a testing of I
00:26:51think it’s not working
00:26:55maybe the gathering is getting so slow
00:27:01hello this is a test of voice
00:27:03recognition bar by Google Cloud machine
00:27:06learning oh yeah
00:27:11so final result is you know not bad
00:27:15right and you also saw the fast response
00:27:19so you could get the recognition result
00:27:22in recent one second district Iraq a 0.5
00:27:25seconds in real time so those are the
00:27:32api’s and but those api’s here are
00:27:37pre-trained model so that means you
00:27:41cannot train your own model with those
00:27:43aps and well one would if we country
00:27:46asked questions for those api is that
00:27:49whether the google will be you know
00:27:52looking at the uploaded images or the
00:27:54audio data to train your own model or
00:27:57doing some more research and as know
00:27:59those api saudi all of our products are
00:28:03provided by the Google cloud platform is
00:28:08is under the terms and conditions of DCP
00:28:12that has here some sections for the
00:28:16customer customer data we don’t look at
00:28:18the customer data except for the very
00:28:21special cases for the troubleshooting or
00:28:22emergency situation so basically we
00:28:26don’t look at the gyro data uploaded to
00:28:28the cloud but at the same time so you
00:28:30can so the APS cannot train cannot do
00:28:34the trainings for your data or your
00:28:35applications so that’s the reason why we
00:28:39provide the other options for machine
00:28:41learning with a stencil or cloud machine
00:28:44learning the other d-dick frameworks and
00:28:46platforms that can used for train your
00:28:49own data set train your own machine
00:28:52learning and neural network what is 10
00:28:55so for tensile Pro is an open-source
00:28:57driver of your machine intelligence we
00:28:59have published the libraries last
00:29:02November and this is the the
00:29:05actual framework we are right now using
00:29:09us via Google research of Google brain
00:29:11team so it’s not something’s
00:29:13outdated or stay out since the latest
00:29:16machine learning framework we are using
00:29:18are right now at Google for example if
00:29:22you want to design this
00:29:24restaurant work swag DW Express Pico why
00:29:28you can use Python to write it in a
00:29:31single line of code Roxas you can put
00:29:34the image of cat here this vector and
00:29:38then you would have an output vector
00:29:40that represents the labels of the
00:29:43detected objects like a cat or human
00:29:45face and you can let computers to find
00:29:48the obits and biases so it’s so simple
00:29:53and also it’s really simple to train
00:29:57your networks because you can just write
00:30:00this single line to have your networks
00:30:04trained for your training data set by
00:30:08using by specifying the algorithm
00:30:10Stryker gradient is set you don’t have
00:30:13to implement your own the procedural
00:30:17code or called – implementing the each
00:30:20the optimization logic actually I’m not
00:30:23good at math or those machine learning
00:30:25algorithms but still I can just copy and
00:30:28paste the sample code to my laptop and
00:30:30I’m praying with my own data sets missed
00:30:34and so forth so you can just let the
00:30:37chancel for runtimes to do the
00:30:38optimization and also the tool provides
00:30:44you a very good visualization tool so
00:30:46one of the problems we had at Google for
00:30:50applying the neural networks to the
00:30:52production production problem is the
00:30:55debugging so if you have many more
00:30:58hidden layers inside the neural networks
00:31:00you have to check the all these stages
00:31:03of the parameters whether the parameters
00:31:06are converging in a right direction or
00:31:10the parameters could be you know going
00:31:13away and having a wrong number such as a
00:31:16na or 0 elsewhere so it’s really
00:31:21important to visualize what’s happening
00:31:23inside in your networks and tensorflow
00:31:25provides the tool and also the
00:31:29portability is another important aspect
00:31:32of the framework so once you have
00:31:34defined your neural networks with
00:31:38Python code of tensorflow then you can
00:31:41start running you and you ready to work
00:31:42training or prediction with your laptop
00:31:45like a Mac or Windows Raptor but you
00:31:48will find that your laptop is too slow
00:31:52to trying the trendy deep neural
00:31:56networks so maybe you may soon want to
00:31:59buy some GPU class and instead of maybe
00:32:052 or 3 GB because in a single box but
00:32:08still usually it takes like a few few
00:32:12days usually a few days or maybe some
00:32:16people spending a few weeks to to do the
00:32:19trainings on their neural networks so it
00:32:22takes so much computation time so in
00:32:25that case you can applaud your tensor
00:32:27flow graph to Google cloud so that you
00:32:30can utilize the power of the tents or
00:32:32maybe hundreds of GPU instances we have
00:32:35we’re running at Google cloud and also
00:32:39once you have finished your training
00:32:41then the size of the parameter sets
00:32:44could be fit into our hundreds of
00:32:46megabytes or tens of megabytes then you
00:32:48can easily copy that parameter sets into
00:32:51the smaller devices mobile devices or
00:32:54IOT devices such as Android iOS or maybe
00:32:58Raspberry Pi so that you know you can
00:33:01have those devices doing the prediction
00:33:04like image recognition or voice
00:33:06retention without using any internet
00:33:09connection
00:33:09everything could be implemented within
00:33:13the framework of tensorflow and with the
00:33:18at the last Google i/o it was about 1
00:33:22months ago we have announced a new
00:33:24technology called tensor processing unit
00:33:26this is a replacement not a replacement
00:33:29maybe a complementary technology for the
00:33:32GPU and CPU so so far or maybe right now
00:33:37the any deep neural networks researchers
00:33:40or developers outside Google is using
00:33:43GPUs mostly for training the neural
00:33:45networks because it’s a matrix
00:33:47operations and by using GPUs you can
00:33:50accelerate the matrix office
00:33:52ten times or maybe 40 times faster so
00:33:55that’s what typical neural networks
00:33:59users are doing right now but the
00:34:02devices problem for GPU is the power
00:34:05consumption each consumers record 100
00:34:08watts or 200 watts per GPU card and we
00:34:13are having we were using thousands of
00:34:16them in a Google Data Center and power
00:34:18Concepcion is becoming the Rogers
00:34:20problem so by by designing the Asics or
00:34:25the editorship specifically for the
00:34:28tensor flow or deep neural networks we
00:34:31were able to reduce the power
00:34:33consumption and gain the ten times
00:34:36better for our performance – powerful
00:34:42performance for what result and we also
00:34:46use the special techniques such as the
00:34:49bit quantization rather than using a
00:34:5132-bit or 16-bit to calculate everything
00:34:54every matrix operations we use the
00:34:57quantization strike a quantized into the
00:34:598-bit where there’s not so not so much
00:35:05loss of the accuracy so that you can fit
00:35:09much bigger parameters into a very small
00:35:12memory footprint and we have been using
00:35:16GTP’s for many production projects
00:35:19already run greying alphago and google
00:35:23photos speech recognitions these are all
00:35:26has been using tepees since a couple
00:35:29months actually we have been using GPS
00:35:31for less than one year and the if you
00:35:38want to yeah we have been I haven’t
00:35:41discussed describing about the power of
00:35:43the Google brain such as number of the
00:35:45CPUs GPUs and TP is and if you want to
00:35:49utilize the power of Google brand
00:35:51infrastructure here’s the product we
00:35:53provide which is called cloud machine
00:35:56learning
00:35:56crud machine learning is a fully managed
00:35:59distributed training environment for
00:36:01your tester for graph so once you have
00:36:03written
00:36:05wrote your tensor flow graph and run run
00:36:08it on the laptop then you can upload the
00:36:11same types of rock for graph to Google
00:36:13Cloud messing learning so that you can
00:36:15specify the number of the GPS you want
00:36:17to use with the service suggested you
00:36:19such as 20 nodes or 50 notes to do the
00:36:23acceleration only training when a crowd
00:36:26Mao is in the limited preview so you
00:36:28have to sign up to start trying out but
00:36:31maybe I we suppose that the for
00:36:33availability record public better will
00:36:37be sometimes later in this year if you
00:36:41go to the YouTube then you can take a
00:36:43look at the actual demonstration of
00:36:45kratom a budget of teens where they have
00:36:47where he has presented demonstrated the
00:36:51actual chance of grotessa fraud based
00:36:53neural networks that takes 8 hours with
00:36:56single node but if you upload the same
00:36:58tensor flow graph to the crowd email
00:37:00then you can accelerate the performance
00:37:04to up to 15 times faster that means you
00:37:07could get the result of the trainings
00:37:10within 30 minutes rather than waiting
00:37:13for the 8 hours that is the speed we are
00:37:16seeing inside Google for any deep
00:37:18learning deployment and we externalizing
00:37:21the dispo to to you guys to have you
00:37:25utilizing the power for 40 you are for
00:37:29solving your own problems and also
00:37:31claudemir can be used for the production
00:37:32as well not only for the training and
00:37:34indeed mo solutions he presented at he
00:37:37demonstrated that the crowd email could
00:37:39be used for the body predictions at 300
00:37:43meters per second so those are the the
00:37:49topics I have covered and now we have
00:37:53two different products one is the Amero
00:37:55api’s like a vision API or speech API
00:37:58where you can just upload your own data
00:38:00to cloud so that you will be getting the
00:38:02results in a few seconds and if you want
00:38:06to train your own neural networks
00:38:09then you can use a try using the
00:38:13intensive roll or crud machine learning
00:38:15so that you can accelerate training
00:38:17you’re on or you’re on your network so
00:38:22if you take a look at the links on the
00:38:25resources of this lesson that you can
00:38:27start trying out those products right
00:38:30now thank you so much yeah I got two
00:38:41questions Trulia
00:38:44first of all few steps back on the
00:38:46machine learning and neural networks
00:38:47mm-hmm
00:38:49you talked about more hidden layers to
00:38:51more complex algorithm yeah but Sturm
00:38:54x-fighters they’re like we were told in
00:38:57University to use yeah you’re feeling to
00:39:03see how many layers you have any tips on
00:39:06that or yeah yeah that’s actually a
00:39:09really good question
00:39:10so maybe question is is there any good
00:39:14practice on designing the Union neural
00:39:16networks right so as far as I know
00:39:18there’s no one theory to optimize your
00:39:23design of neural networks so everybody
00:39:26even in the fifth at Google you know
00:39:28when I asked the Google research team
00:39:30and people would say you know let’s
00:39:31start with the five 5 children
00:39:34let’s see how it works all right so
00:39:35strata and that’s the largest challenge
00:39:37we have right now for deploying the
00:39:40neural networks for your own data or
00:39:42applications so you have to do it in
00:39:44many many trials just right the the
00:39:48people in the pharmaceutical companies
00:39:50trying to create a new drug so you have
00:39:53to have a different combination of
00:39:55obably hyperparameters hyper parameters
00:39:57means the parameters such as the number
00:39:59of hidden layers or the or new ones or
00:40:02the way you can import the data or
00:40:05extracting features so you have to try
00:40:08out every different combinations that’s
00:40:10the problem
00:40:10and also it takes much much computation
00:40:12power yeah I think that’s what all do
00:40:15yeah it’s not a theory behind it and a
00:40:18little bit later in the presentation but
00:40:20use cases will be supported in the near
00:40:23future
00:40:24we’re going from post training to
00:40:25runtime can network we and B be exported
00:40:31in order to use a better processing for
00:40:34example what is second question cannon
00:40:39can the network we can be be exported in
00:40:42order to reuse in embedded processing
00:40:44for example by exporting okay yeah
00:40:47before the first question which is the
00:40:49the online training I think it’s on the
00:40:52world map or maybe your do list of the
00:40:56nucleus but it’s currently it’s not
00:40:58supported but it’s possible that we’ll
00:41:00be supporting the online training where
00:41:02your neural networks will be gradually
00:41:05joined by the online data and second
00:41:09question is exporting yes you can export
00:41:12the trained parameter sets so that you
00:41:14can use the parameter assists to your
00:41:18une use cases such as the importing the
00:41:21parameters into the IOT devices or maybe
00:41:23you can even copy that data sets into
00:41:26different cloud record AWS to learn your
00:41:29predictions on database ok those
00:41:33questions are cut do anyone else has any
00:41:36questions no and thank you very much
00:41:44thank you so much