Přepis řeči - Making Sense of a Zettabyte World

0:00:13	each G for the past fourteen years the ieee present of the ieee jack it cool be signal processing metal
0:00:20	the some portal or fun the refer to as the kill be metal
0:00:24	it's what i you refers to as a major metal
0:00:28	it was found about the ieee signal process society has been generously funded since its creation
0:00:34	but the texas instruments company
0:00:36	jack you'll be love to be in an engineer use was an inspiration to many do decided to pursue electrical
0:00:41	engineering as a career
0:00:43	well he is no longer an are males
0:00:45	his influence contain
0:00:48	as of already mentioned because it's a major metal
0:00:51	the ieee jack he'll be signal processing metal is presented by of the ieee
0:00:56	with the ieee medals and awards ceremony the will be held issue you're in august and san francisco california
0:01:03	however it is been the practise
0:01:05	of the metals down in society
0:01:07	that would be us
0:01:08	to present a special commemorative plaque
0:01:11	from the society
0:01:12	to the kill be metal B recipient
0:01:15	this year the recipient of the two thousand eleven ieee jackie a skill we signal processing metal
0:01:21	is in great do over shape
0:01:24	professor over a is receiving this honour or for pioneering contributions and you theory and applications of wavelets
0:01:32	and a filter
0:01:34	in could not be with us today we she will receive the metal at the ieee tripoli medal ceremony in
0:01:39	august in san francisco
0:01:42	but i do recommend we go for a round of applause in it since yeah
0:01:54	ladies and gentlemen that concludes today's ceremony i want to thank you all for attending the in the at triple
0:02:00	you signal processing society
0:02:02	order order B against than a that triple follows the recipient of the
0:02:07	ieee james flanagan at speech processing metal and the research you know of the ieee tripoli jet kill be metal
0:02:15	the so sorry hold as awards ceremony and really and i cast
0:02:19	a look forward to seeing you again next year
0:02:22	thanks again
0:02:36	i john for something thing all these works
0:02:39	and can grow to your workplace
0:02:41	so this is closing that
0:02:42	the official opening
0:02:44	and sort money
0:02:45	and that the and let me thank our musicians mister hour and mister of ski
0:02:50	for making this a a more you and
0:02:55	okay thank you
0:02:56	we are on time
0:02:57	it's miracle
0:02:59	the we are moving on to the technical program of for conference
0:03:02	the first one talk
0:03:04	i will come on the code um uh are or for a speaker
0:03:08	and a from a nokia
0:03:10	and professor before cory one and from helsinki university of technology
0:03:14	we will present or for a speaker
0:03:16	and should six
0:03:19	Q
0:03:21	where is send them and uh i have a great place or to introduce the an our to speaker
0:03:27	but that it really a in or was present or and that of market research and
0:03:32	nokia research center break was to enable a new business opportunities for not
0:03:38	yeah race the responsible for
0:03:40	you you know we're white that first
0:03:42	struck in the
0:03:43	and work closely we or not give a a in to promote open in the base
0:03:49	and working phone research is in collaboration with we we equal global research universities and
0:03:55	is
0:03:56	i believe that means that expense
0:03:58	what of time not only in is of this but also at the airports and a
0:04:05	and it well it's a T in computer science from the universal of think in one
0:04:09	and he joint here in
0:04:11	do got from for as the results go in software applications well
0:04:15	is previous positions include working at a at and T you go uh and uh us a visiting research scientist
0:04:21	at a so things
0:04:23	where can going to the wells role
0:04:25	called the also for email
0:04:28	a or and cool or able or for more than the or as would a i of
0:04:32	a a personal computer science social science statistics and holes like that
0:04:37	some of these papers are are are very can the school of single produce or so
0:04:41	including machine learning and and uh
0:04:44	uh
0:04:47	um
0:04:48	okay okay
0:04:49	is been also used to professor at to use C berkeley and stanford it works
0:04:54	is other interest include a a a a a long board surfing and scroll boarding
0:04:58	and i think that's why you like look in california would
0:05:03	i that the crib leads to work in a and risky more much of a is and i enjoy every
0:05:08	ladies and gentlemen of and it could be very oh do you know ahead
0:05:17	okay
0:05:19	good morning
0:05:20	um those of you who are right from uh uh U S i know
0:05:24	it's very early in the morning high i came from the your can the data from on but i'm
0:05:29	uh like uh
0:05:31	was mentioned i've spent a lot of times on the plane side actually not sure which time zone i'm
0:05:36	and in
0:05:39	um
0:05:40	what i was thinking about
0:05:41	what's the topic
0:05:43	that i would like to touch to for sets are how would i say
0:05:47	wide scope audience
0:05:49	i chose data
0:05:51	and they are multiple reasons of choosing data
0:05:54	one was that i spend
0:05:56	more than twenty five years some my life for searching for the medical data set
0:06:00	that could prove that my method some better than the other
0:06:05	um
0:06:06	a i for charlie that search um
0:06:08	of course
0:06:09	uh was a a a a difficult one not only because of the maybe the problems with my methods
0:06:15	but the access to data
0:06:18	and um
0:06:20	having worked long in the field
0:06:23	and and and looking at multiple datasets i know that
0:06:26	where we use of lee and
0:06:28	in that type of a situation ease the fitting
0:06:32	are methods to that particular datasets that we have a
0:06:36	or generating synthetic datasets which we know that will have their own
0:06:40	problems when when you use them some my feel was
0:06:43	a lot of what you would call machine learning or
0:06:45	information T legal learning or bayesian learning
0:06:49	one of the reasons is i i was so on the bayesian learning as we will hear um
0:06:54	later on during the conference was that as i had so little data at the bayesian
0:06:59	let's some what's uh easier methods to apply than the frequent just methods
0:07:05	so i decided to talk about data
0:07:07	today
0:07:08	just because i think it's a very common topic for
0:07:11	most of the things that you see the con for
0:07:14	uh
0:07:15	but i'm taking a very different approach spend again a a lot of time of doing um
0:07:22	the
0:07:23	i writing papers on on on on the different that a mathematical properties of of learning
0:07:29	and also implementing that different algorithms
0:07:32	and today i'm and going back to my roots of hacking
0:07:36	so i'm celebrating
0:07:37	this year forty years of hacking
0:07:39	which is a long time
0:07:41	and and taking a little bit just an perspective as people see
0:07:46	now so let's start
0:07:49	uh the the title to make sense of as that of by twelve well might wonder what those green things
0:07:54	that they are that the going it's not the warm gain the famous weren't gain that just to be in
0:07:58	in in the old days and on the mobile phones
0:08:01	so if you look at those uh green thing
0:08:05	a new reveal something out of eight actually it nice a description of representation all a mobile device is that
0:08:12	would be in a moving
0:08:14	the cool
0:08:15	going round that
0:08:17	white
0:08:17	yeah area are use the congress centre
0:08:20	so those would be the sensors that would be driving around of roads in in the device sees uh
0:08:27	that you carry with you
0:08:28	so
0:08:29	here really
0:08:30	uh typically nowadays which why call mobile computers because some of can be just and this so everything is a
0:08:35	computer to me
0:08:37	now this is
0:08:38	synthetic data
0:08:40	as we say
0:08:41	but that's
0:08:42	zoom a little bit
0:08:45	so this is real data
0:08:47	and this is real data
0:08:49	as collected by did not get devices
0:08:52	in prague
0:08:55	a by looking the navigation request
0:08:58	oh the different devices
0:09:00	in real time
0:09:01	you know
0:09:02	so pretty posing them on this plane with the court in eight
0:09:06	and the timing
0:09:08	i think it's see there is no matter
0:09:11	and the name
0:09:13	but after a while you see just thing developing nicely to that
0:09:17	map of prop
0:09:22	so
0:09:23	if we go for are
0:09:28	and assume i'll
0:09:30	this is a similar picture
0:09:33	of europe up
0:09:35	as seen
0:09:36	by billy in some cory is
0:09:38	all the all be
0:09:40	max
0:09:41	navigation
0:09:42	not just by the way there are some interesting cultural differences this speaker start mining this
0:09:47	as the topic of day
0:09:49	um
0:09:50	the green areas in fronds and also in russia
0:09:53	so that the uh the the those areas the people are more interested in navigation as opposed to seven your
0:10:00	of what the point of interest meeting
0:10:02	finding restaurants or
0:10:04	uh those are much more calm
0:10:08	so there's a multitude of day that that i've
0:10:10	face in my
0:10:11	current to life
0:10:13	it's also an a i the be do prove each to be an and here where this they are
0:10:19	and the C or or search of mine
0:10:21	is
0:10:22	starting to be fulfilled
0:10:24	we have access to data
0:10:27	in a a
0:10:29	totally
0:10:30	new way is
0:10:32	to to to develop of computers to to develop and of a multiple
0:10:36	different aspects that i that touched
0:10:40	so first i wanna uh
0:10:42	this costs
0:10:43	would you
0:10:45	some of the source is of this type of data that that is available to you
0:10:49	i my papers is to challenge you
0:10:51	later on to think about
0:10:54	what do you need to do
0:10:56	what do you need to think about when you in real life want to address
0:11:01	datasets like this
0:11:02	and you wise thing and and uh
0:11:05	so of making sense of the state
0:11:12	so first of all i showed you now
0:11:15	data that was based on the G P S
0:11:18	location information
0:11:20	in a time code
0:11:22	actually actually mobile device
0:11:25	already already like some multiple the for and
0:11:28	sensor data at the same time
0:11:30	we don't that they have a X are meters
0:11:33	yeah actually for me camera and audio is also a a a a a sensor
0:11:38	so each of these sensors that you at that is
0:11:41	this device
0:11:43	that that now so prevalent
0:11:45	adds a new layer
0:11:47	and there um more or more sensors that are coming a in the different radios that we are adding to
0:11:52	these devices
0:11:53	you're a more or more
0:11:54	a a sensor data that coming from the same source is
0:11:58	but at E
0:11:59	to do is uh uh base layer
0:12:02	of of location and time
0:12:04	why do why consider a location and time to be different than the ad
0:12:08	because
0:12:09	from my perspective
0:12:11	location and time can dish
0:12:14	the usefulness of a lot of the data
0:12:16	so if you don't know your location
0:12:18	many of the application for example like asking where the close
0:12:23	beer
0:12:24	part
0:12:25	don't make very much sense
0:12:27	so what some location and time of very fundamental
0:12:30	but beyond that
0:12:32	you can add more and more layers
0:12:34	and you in up in these piles of data
0:12:38	which of course
0:12:39	if you collapse them
0:12:41	shows how he man amount of data we have available
0:12:44	already from this type of the don't
0:12:52	that about it
0:12:55	the
0:12:57	well at least i it a part of my language but only row room full of geeks so we most
0:13:01	of us are
0:13:02	so excited about that the different numbers
0:13:05	i i was always possible
0:13:08	i i one studied mathematics that they are these different terms for the same numbers and and of four in
0:13:13	particular could be just as we are particularly excited of using
0:13:17	uh just power some two
0:13:19	so we talk about set up
0:13:20	now now set up what is an interesting number because it's the time um
0:13:25	currently approximately this year
0:13:29	although much of the data E is in fact copies of each or
0:13:33	we are producing a one point to set up of digital information
0:13:38	but it it this number also as any a large number
0:13:42	tends to be uh
0:13:43	difficult to to grasp
0:13:45	so um
0:13:47	it before sort of try do uh look at the different sources of this
0:13:51	at let's try to sort of reflect the little bit what it is so uh
0:13:55	okay
0:13:56	it's
0:13:57	approximately close enough to be six that we am
0:14:00	ten to the power of twenty one so we have twenty one zeros
0:14:03	but that doesn't tell us too much
0:14:05	so uh typically E um
0:14:07	if you want to just to get an idea of a big number or you should reflect it with uh
0:14:12	some major that you know
0:14:15	so let's
0:14:16	say that one but
0:14:18	of this would be one meter
0:14:19	and a good question to you is that now okay if
0:14:22	if we have
0:14:23	set meters
0:14:26	oh but this that's how long
0:14:29	what that this that's speech if we start from here
0:14:32	is it here from to moon
0:14:36	here to to peter
0:14:40	here to i'll for sent that already
0:14:44	okay so those have very clever and fast with the uh and know something about astronomy
0:14:49	would figure out that
0:14:50	this
0:14:50	is is that meters
0:14:52	is actually uh
0:14:55	the same as the diameter of the milky way
0:14:59	approximately
0:15:01	which is about hundred thousand like your
0:15:04	a lot of it's a big number
0:15:06	now i actually prefer somewhat the
0:15:09	the that bit more mundane uh reference sees that uh you can find on on on the net that when
0:15:16	we were a looking at the numbers
0:15:18	so one is that
0:15:20	set up by just amount of information if all the people on uh at
0:15:24	would be to twenty four seven four hundred year
0:15:30	or
0:15:31	it would be like seventy five really and
0:15:35	sixteen gigabytes
0:15:36	i had
0:15:38	for a of the data
0:15:40	which actually fills at four times
0:15:42	the more line time
0:15:46	uh but my favourite he's that if you like T V shows
0:15:49	it would be watching the the you know the he
0:15:52	C uh series or actual the first is sort of series
0:15:56	all the T V series twenty four
0:15:58	uh a four hundred and twenty five million years
0:16:01	continuous
0:16:04	talk about the sort of a
0:16:06	getting bored a bit probably
0:16:09	someone that
0:16:10	okay
0:16:11	now um
0:16:13	the standard answer when a
0:16:15	what we look at this type of a large dataset is that hey
0:16:20	we should be using a approximations we should be using sampling we should be doing you know
0:16:26	not exact things we we should be somehow
0:16:29	you know many lady
0:16:31	uh the the data set and then is that is correct
0:16:34	it's actually very old idea um in this scare out have lights this is a very old tab
0:16:41	eight thousand years old it
0:16:42	it's the babylonian tablet
0:16:44	that in fact
0:16:46	shows
0:16:47	a a approximation
0:16:49	all the uh the square root
0:16:51	a a a a a a unique
0:16:52	uh square as sorry at the diagonal don't our uh
0:16:56	although a unit square
0:16:58	uh
0:16:58	then
0:16:59	that allows us
0:17:01	in fact
0:17:01	to do
0:17:02	uh
0:17:03	a square with calculation
0:17:05	a for construction and
0:17:07	and and complex
0:17:09	so also to approximate measures sum of course of very all thing
0:17:14	now unfortunately
0:17:16	uh approach in a measures also as we know lose information
0:17:20	set K C that is perfectly fine
0:17:23	certain cases
0:17:24	uh it cost is by sees that we know what will be very hard
0:17:29	but the amount of data that we are talking about today
0:17:33	by four
0:17:34	requires as
0:17:35	to go to these approximate method of course
0:17:39	oh good have
0:17:40	like was pressing the wrong but
0:17:43	so let's look at a little bit about the source C east of the data now
0:17:47	at this is a
0:17:48	different picture than the previous ones
0:17:51	because this is not showing the absolute capacity
0:17:55	it is still the relative capacity of the type of the data that that is available for you go
0:18:01	so
0:18:04	in the old days when i remember one computer networks started
0:18:08	the remember forty years of hacking
0:18:10	one time
0:18:11	there's a lot of F T P traffic going on
0:18:15	E became very popular in the early eighties
0:18:19	uh there was something like telnet i don't know how many
0:18:22	remember or anything like that
0:18:24	and
0:18:25	but if T P was by far did don't mean a uh D that that was available meeting file transfer
0:18:31	from one place do not
0:18:33	that's to nineteen ninety
0:18:37	ninety only nineteen nineties
0:18:38	we all know um
0:18:40	the one of the
0:18:42	still annoying fact to the computer scientist that the physicist
0:18:45	so the introduced the the H U T P protocol and and the way but it was not the computer
0:18:50	this
0:18:51	should
0:18:52	and
0:18:53	the web was born
0:18:54	as only as you can see
0:18:58	the
0:18:59	if T P part
0:19:00	start
0:19:02	diminishing diminishing proportionally remember
0:19:04	this is a proportional of the act up some will not a you know a amounts are going up all
0:19:08	the time
0:19:09	and to way easy grabbing a more and more she
0:19:13	newsgroups are pretty happy
0:19:14	that tell let this sort of disappearing
0:19:17	email keeps it sort of a constant
0:19:19	and in that
0:19:22	if you go further
0:19:24	to dine T five
0:19:25	where has already captured half of the traffic
0:19:29	and you see in the upper corner or something interesting
0:19:32	like data
0:19:34	appearing from individual
0:19:37	peer-to-peer communication of the computer
0:19:40	which didn't used to be a case
0:19:42	in the past
0:19:43	because we only had this
0:19:45	few mainframes frames go
0:19:48	and if we go even further
0:19:53	to two thousand
0:19:55	we did see that the video
0:19:59	and video information
0:20:01	starts to grab a larger and larger here
0:20:05	oh the digital traffic
0:20:06	now way be still strong
0:20:09	but we D L the the purple part
0:20:12	uh ease
0:20:13	ease
0:20:14	just
0:20:15	morning
0:20:15	there at the corner
0:20:18	peer-to-peer
0:20:19	and web dominating
0:20:20	but
0:20:21	it's scroll very fast
0:20:23	so if we go no further
0:20:25	of course to two thousand five you see that the video is getting bigger
0:20:30	and um i don't know what a a because of the various different type of legally shoes another is used
0:20:36	the the uh the sort of percent each of a peer to peer
0:20:39	anyway not growing anymore to saying way and way is going down
0:20:43	and if you a ripe to two thousand ten
0:20:46	one could argue now that one the most important and interesting data sources that we and yeah
0:20:53	is the video
0:20:55	and nice there is no sign of it
0:20:58	going down at the moment
0:20:59	it actually
0:21:00	if you believe just go
0:21:02	which of course
0:21:03	uh can be a little biased you
0:21:05	the video will be so dominating at in the next couple of years on the network traffic that it will
0:21:10	be the majority of
0:21:12	and when
0:21:13	is in fact going down
0:21:14	proportion
0:21:16	which is quite obvious because you think about
0:21:19	the bits
0:21:20	required from of the normal web back
0:21:24	so when
0:21:25	the quality
0:21:27	sort that type of data
0:21:29	that that he's he's
0:21:31	moving around in these networks
0:21:33	is obviously you video data
0:21:35	so a lot of the things
0:21:37	that i used to be interested in which where relate with
0:21:40	sort of pattern matching in text
0:21:43	or some stuff in the files of music
0:21:46	is actually replace now in in stress to do this type of the mining
0:21:51	or a a processing of video
0:21:55	that that was the web
0:21:59	now
0:22:00	a king about where i just one of major something which it because it touched me so much in a
0:22:05	related to well sort talk about video
0:22:08	i wanna touch something um that a three D really captured my heart in in march in long beach ten
0:22:15	this was there raw is
0:22:17	from mit media lab
0:22:19	where on capturing ninety thousand hours of video all he's child growing up
0:22:26	and mining that we deal
0:22:27	uh in such a way that for example he could show
0:22:30	in the speed up manner
0:22:32	or the development of the work
0:22:35	water
0:22:37	in he a you know in the the the language development of the time
0:22:41	again a unique experiment
0:22:44	but related to topic
0:22:46	even more interestingly he's company as blue then
0:22:50	is working
0:22:51	and and delivering a
0:22:54	they are sort of a uh and now it takes or visualization
0:22:59	all uh both of the tv V broadcast real at with the social networking track
0:23:06	basically linking
0:23:08	something that is
0:23:10	shown on the T V on the discussion
0:23:13	that you have one on the net
0:23:21	okay
0:23:23	the little but different domain as a reference of this is the large had collider for those of you who
0:23:28	have not seen that
0:23:30	right picture
0:23:31	um
0:23:33	i do remember
0:23:34	that basically when the large had drunk lighter in and the data read was planned
0:23:41	there was a lot of talk about the capacity capacities now first of all a happened collider
0:23:46	has a hundred and fifty million sensors
0:23:49	so that's a lot of sensors an and we all know of course that these sensors are also proved using
0:23:53	then the data
0:23:54	with that in yeah
0:23:55	so of very rapid speech
0:23:58	so uh
0:24:00	the actual
0:24:01	approximate uh
0:24:03	so the bound of data with you about in this a structure used one paid the bite per second
0:24:09	i do remember that the original specs when we started doing the data rate i i i of was there
0:24:13	in the in this huge european union
0:24:16	consort you which has its benefits and a normal of your opinion and
0:24:20	"'cause" so a a is
0:24:22	uh we were talking about four point five that are bytes per second so
0:24:25	the the the bottom some got up
0:24:28	okay
0:24:29	hundred fifty million sensors
0:24:30	cool so
0:24:31	this is the physics six experiments
0:24:33	this is the science
0:24:35	big science you know
0:24:37	what does it do with you know the regular well or whatever it is this that very special device
0:24:42	expensive device put some
0:24:45	that that us come back this is now a picture of all of the whole well
0:24:50	related to the picture use of already and you know prod your up
0:24:55	um
0:24:55	the the different uh core
0:24:58	on
0:24:59	on the on the navigation this based on twenty billion court
0:25:03	now
0:25:04	remember what i showed you earlier
0:25:06	this a button one point two billion a devices currently on a that that of course and the number of
0:25:12	a mobile device is is a four point
0:25:15	or more than four billion
0:25:17	if each of these one point to build and device has ten sensor
0:25:23	it's ten
0:25:24	more than ten
0:25:25	bill
0:25:26	sensors
0:25:28	these ten billion sensors
0:25:30	although they don't feet
0:25:32	the sis that
0:25:33	with the same speed than a large how drum or would be
0:25:36	are still is
0:25:37	super substantial amount of
0:25:39	data that is available for
0:25:43	and this is
0:25:44	i'm not talking about the future somewhere
0:25:47	i'm talking about
0:25:49	the actual
0:25:50	today
0:25:52	not saying that all that sensor information is now at where collect the in one place
0:25:58	but it really really really
0:26:01	ease shown the potential and the different
0:26:04	uh
0:26:05	past that we have in the fit
0:26:08	i mean the different types of sensors in this mobile about computers
0:26:13	and i mentioned the sensors that are relay a with the a lot of the user in the phase or
0:26:18	or or very different types of a
0:26:20	uh uh uh uh a sort of uh positioning and so on
0:26:24	but this an interesting you
0:26:26	uh source
0:26:28	that at
0:26:29	to this sensor
0:26:31	wall
0:26:32	and that's the cognitive radio i so that the of some papers in a common give radio in this conference
0:26:36	as use one
0:26:38	to just want to point out that from this
0:26:39	that about uh perspective for those also you by the way to calm radio used in the
0:26:44	then and make a a location of the radio spectrum
0:26:48	uh in such a way that the device itself can actually choose
0:26:52	which part of the spectrum meet using
0:26:54	uh a it's signal
0:26:56	a transmission
0:26:57	uh actually can be used for a out of things do and and a for the sensor at of
0:27:01	so
0:27:02	that put detailed
0:27:04	introduction of county the radio will already bring that
0:27:08	a again at new very interesting source of since information which is in the infrastructure itself
0:27:16	so the traditional picture of having the device is talking to a
0:27:20	power
0:27:22	sell power and we'd already know something about the sting all strings that they can year
0:27:27	uh a how to sell power can recognise of nice that the device
0:27:30	is gonna change
0:27:32	to a picture which is a much more mesh
0:27:35	what a device are aware of each other's
0:27:39	press sense
0:27:39	or partially aware of each other's presence
0:27:42	in different type a radius spec
0:27:44	now these fingerprinting information ads and not the layer
0:27:48	again
0:27:49	which is inherent to the billy and
0:27:52	mobile about computer infrastructure that we have
0:28:00	well
0:28:02	and that the source of data
0:28:03	available for us all in that one is
0:28:06	ease the social media
0:28:08	there's currently about how nine hundred million social media users in the well
0:28:14	of course
0:28:14	if you look at that
0:28:16	um that
0:28:17	in principle means that there's is but something like one one point five billion of it's just social networks every
0:28:24	day
0:28:24	each of these base it's
0:28:26	leaves a trace or a is a a operation
0:28:30	and of course
0:28:32	if you want do uh divide this we know what that a majority of this is coming from a single
0:28:37	source
0:28:39	place book
0:28:40	seven hundred million currently
0:28:44	but the important part here is that this about thirteen billion
0:28:48	or more pieces of content axe
0:28:52	by these users
0:28:53	and this is the richest just context we have a one uh for for my because this is a uh
0:28:59	as you know in face book
0:29:01	or sort of information of all different types it is
0:29:04	it is both image we is it's is a low eighties textual data
0:29:09	E Ds a different lean C is sort of informative in a very different
0:29:15	additional things that you can uh a of course
0:29:18	C in the social space is that we have about sixty billion three
0:29:22	expected in two thousand eleven
0:29:26	and this sixty billion to
0:29:28	and uh
0:29:29	is
0:29:30	still a growing number because we have a four hundred sixty thousand you tweeter are guns at a daily which
0:29:36	by the is not to growth rate because
0:29:38	there are also people
0:29:39	that a drop that or accounts
0:29:41	but still shows that
0:29:43	the actual so the that population
0:29:46	is growing
0:29:50	and of course
0:29:51	back to our favourite
0:29:52	video
0:29:53	that is that a lot of the traffic
0:29:57	and in in the picture that showed you earlier
0:30:01	comes from you two
0:30:02	but uh of course in areas like in you dies states another it comes also from net flicks and so
0:30:08	so thirteen million hours of video on
0:30:12	in you to that doesn't look at very big number think about i was talking about the billions of there
0:30:17	but these to remember that these videos
0:30:19	are in fine it
0:30:21	snippets or so
0:30:22	so that two these thirteen billion our
0:30:25	is much more
0:30:27	uh in the number of videos that we have a below
0:30:30	um G
0:30:31	do
0:30:33	pushed on on but
0:30:36	so that million hours of video a it to you
0:30:41	thirty five hours a new video uploaded per
0:30:44	so those are you working on video mining
0:30:47	you have
0:30:47	great future
0:30:49	and no head of
0:30:50	now i'm i'm an optimistic and possibly person so i like this thing when all my life that things go
0:30:56	up and the are upper right corner
0:30:58	i like things growing i like things becoming more challenging i like things become fast there
0:31:04	small
0:31:05	bigger and so
0:31:08	what's the problem
0:31:10	well the problem is that
0:31:13	as opposed to you know
0:31:15	having this thing on the paper
0:31:19	what as to form a
0:31:21	or even ask calculation in your machine
0:31:24	we're talking about real systems you
0:31:28	and this data that exist somewhere
0:31:31	we need to access it we need do you hand the like and if you want to make use of
0:31:36	a
0:31:36	we need to be build systems that that that sort of a a a a a able to
0:31:41	do or what but
0:31:43	what happens if you are not careful of building this just a
0:31:48	that's a crack
0:31:50	this
0:31:52	actually uh the the the cover is from an older days but it to tall some ten we know that
0:31:57	that the one some point
0:32:01	decline in new york stock exchange in in a very short period of time
0:32:06	which actually result of a complex
0:32:08	a cohort it
0:32:10	yeah a computer software
0:32:12	that where of course doing uh uh uh what they are supposed to do they are competing on the market
0:32:18	in the super whom human human speech
0:32:21	the available data that they have
0:32:23	making it looking at weak signals and in sort of a a
0:32:26	uh a sort of a a or a the crash
0:32:29	so like i always point out
0:32:32	it's nice to write a paper
0:32:35	then have a good learning out
0:32:37	and a good predictive model
0:32:39	a bit to be
0:32:40	much more certain and when you start applying that
0:32:43	in the real well and you course
0:32:46	some interventions in the real world
0:32:48	the two very different
0:32:52	so what i want to talk about i was talking about the what now
0:32:56	what is the data available
0:32:58	i would like the little bit touch
0:33:00	how and why wide what want to do it for this audience i mean i'm not talking in operating system
0:33:04	conference i'm not talking in a networking conference on not talking about
0:33:08	uh the people even in my formal of those databases database
0:33:12	community
0:33:15	when i was a can be just signed this undergrad
0:33:18	i've as support about the memory computing trade
0:33:22	and a little bit later i was told about
0:33:24	the paging and you know virtual memory and and somewhat asked
0:33:29	my point is that
0:33:32	we are now
0:33:33	unfortunately in a different architecture
0:33:37	we are using a difference just a and when we are writing our algorithm
0:33:41	and when we are running them
0:33:43	we need to take into account
0:33:45	to a a to greedy D tells which are beyond the turing machine yeah even beyond the for neumann
0:33:51	sort of traditional computer model
0:33:54	we need to look at that aspect if we really wanna work in the real
0:33:58	with this data
0:34:01	uh this a be yeah
0:34:02	between
0:34:04	what is the practise shouldn't traits correct
0:34:07	in in in in in menu plating in of so called internet companies for example for the data
0:34:13	and the work we do and the very at once work we do
0:34:16	in a sophisticated methods of understanding that
0:34:20	sometimes these things
0:34:22	the gap is smaller
0:34:23	sometimes it's much lot
0:34:25	what a talk about
0:34:27	a a and you with
0:34:28	how a all the important aspects of this just then
0:34:33	that we should take into account
0:34:35	what we are writing our algorithm
0:34:36	when we are
0:34:37	you know building them for that the by well
0:34:40	not building them for the
0:34:42	uh
0:34:43	well like my favourite was the or dataset that a over fit it so badly
0:34:49	and list the city
0:34:50	one of the things that has changed
0:34:52	dramatic is that
0:34:54	when we you building things we don't need
0:34:56	bill them
0:34:57	a in such a way that we have to uh make sure that the maximum requirement is somehow con
0:35:05	because in the old days when you had a computer it had a certain amount of uh computing power
0:35:10	said amount of memory
0:35:12	and was a box some
0:35:14	no L this C is an should
0:35:16	of course that has sort of sneak in with clout
0:35:21	so L L this T allows you to use dynamic to computation power or
0:35:26	of a larger or a in today
0:35:28	is such a way that you don't have to
0:35:30	uh uh so the the
0:35:33	three D term mine you model the competing power do you use
0:35:37	no this is a to to the same
0:35:39	development of you had in the uh i guess a into that read data structure of course
0:35:43	where one has
0:35:45	six table set
0:35:48	and then the dynamic to
0:35:50	and you defined find a dynamic table and you don't have to care about the the that that you will
0:35:54	ever go wild
0:35:56	so this
0:35:58	feature where you need more data
0:36:01	and you just grad
0:36:02	more of the competing power
0:36:04	and then
0:36:05	at the same time of course this is meaningful for only because you have multiple users at the same time
0:36:11	sharing
0:36:12	this particular pull
0:36:14	you will goal uh to a much lower T
0:36:16	so in some sense this ls this city
0:36:19	has um
0:36:22	allowed us to do uh things which were not add all
0:36:26	uh feasible in in the past
0:36:29	it was not that long ago i think about
0:36:32	seven eight years ago
0:36:34	we were running multinomial pca out
0:36:38	a on on eight i if fixed cluster for a search and stuff actually doing sort of
0:36:44	uh search engine uh a probabilistic modeling all the the work of course is in in in a hundred million
0:36:52	documents or more
0:36:53	and a typical run
0:36:55	there was
0:36:56	deep limited by competing power that we have
0:36:59	so we we had to run like three weeks in a role
0:37:02	in a small cluster together get a model of like to fifteen million
0:37:05	or twenty male and documents
0:37:07	those you not be a a a no pca
0:37:10	or mode only piece you knows that it sort of a under you do something clever
0:37:13	it's actually quite computationally intensive thing
0:37:16	and and there was no way you doing any kind of dynamic things so all our experiments where
0:37:22	where sort of restrictive by to computing power at the university which was not that great
0:37:26	so we we had to work on
0:37:29	but the the L to sit B for as a historical remark it's to an to you very old knowledge
0:37:35	and
0:37:36	my
0:37:37	okay very uh quotes from the science fiction church or
0:37:41	uh as many of the computer size yeah ideas have a are very only in church or
0:37:45	is this
0:37:46	first
0:37:48	uh commercially sold
0:37:50	um
0:37:52	story by our to see clock
0:37:54	uh from this thing size fiction called
0:37:57	uh a a rescue party
0:38:00	and rescue party
0:38:02	most telling
0:38:03	about
0:38:05	at race
0:38:06	a call uh how do already a place called powered or with the race some power or
0:38:11	yeah had a collective mine
0:38:12	and depending on the problem in the universe
0:38:16	the that race
0:38:17	collected more minds
0:38:19	dynamically to solve the problem
0:38:21	so but in that today if you had to read a really big problem
0:38:24	telepathic connection
0:38:26	in the different places and the unit thus allowed to solve much harder problem
0:38:31	this is nine this was really really are
0:38:33	so fifty so the uh a sort of
0:38:35	or was a used how
0:38:37	well some of the site fix at uh and a or to reflects of a
0:38:41	of the future
0:38:43	a second aspect
0:38:45	well as a list this C you might think okay so it's the cloud stuff
0:38:49	is the robustness are
0:38:51	rocks this argument easy a very complex one uh because it's basically depend
0:38:58	on one um
0:38:59	what do you wanna do
0:39:01	so this is nice result by eric rule or from cow
0:39:05	and this that depending on
0:39:07	which time men she's you are interested in
0:39:09	meeting you're accessing some data
0:39:12	and you want to do a consistent access
0:39:14	or you should be able to access when partitions are allowed in the network
0:39:19	or yeah data should be always available if need
0:39:23	you concept set is five or all of these at the same time
0:39:28	of the
0:39:29	you can only go in in this
0:39:31	different
0:39:31	or warners or a different of borders of the trying is such a way that
0:39:36	if you wanna do for example search
0:39:39	uh you're actually very petition for and
0:39:42	at you have certain type of consistency but not all things are always a
0:39:48	or you bit or end where you don't care about consistency at all
0:39:52	you just
0:39:53	basically are doing
0:39:55	uh transfers but
0:39:56	you very tolerant for petitions as as those who want to legally
0:40:01	imposed against the this type of a part of C no
0:40:04	uh and things have very available
0:40:07	and on the other hand in this
0:40:08	just just do body data bases that many many years ago and is working on them the consistency and availability
0:40:14	a very important
0:40:16	but they were not very tolerant for partition
0:40:19	now why this is important this is important because we have talking about highly to but it just of remember
0:40:24	i was talking about one billion
0:40:26	devices
0:40:27	that does the sitting in different parts of the well
0:40:30	yeah can make it by different make words
0:40:32	a or actually allowing different types of but uh
0:40:36	uh mouth functions and there is them
0:40:39	there is no
0:40:40	consist then everything use up all the time
0:40:43	notion i all
0:40:44	so when you building you out rhythm
0:40:46	they cannot be based
0:40:49	bone
0:40:49	getting all the necessary information but
0:40:53	by necessity is you of course i all assume that you already figured out that in this is that the
0:40:57	by well
0:40:58	the outward have to be on line H
0:41:01	batch algorithms of taking
0:41:04	you know this twenty billion queries and running them and doing it
0:41:07	it usually not the way to go because the response times
0:41:11	uh for the problems that you want saul
0:41:13	i'm not
0:41:14	uh these
0:41:16	so we had a ct and robust
0:41:18	but the
0:41:19	the one thing that is so dear to my heart
0:41:22	these energy
0:41:23	so i
0:41:24	i preached just every place i'd bin
0:41:26	now for the last two years i'm preaching it here too
0:41:30	and i'm pointing out if five where a do that
0:41:33	i would go to this field of energy
0:41:36	efficient compute
0:41:37	so my argument is that the current architecture as we will see
0:41:42	these fun the mentally
0:41:44	imposing the similar even theoretical call uh sort of boundaries
0:41:49	and to to to computing ask we used to do with memory
0:41:54	and computing
0:41:55	energy
0:41:57	is this so
0:41:58	so let's look at the real life situation why of things become difficult it's it's stick exam O
0:42:04	or of a a a a processing that you
0:42:08	you know
0:42:08	normally you used to this type of about
0:42:10	processing where you have a single box now we are in a well what we a multiple devices that i
0:42:16	connected to each other
0:42:17	and you have your how or it your great algorithm and doing the video stuff
0:42:22	has a choice
0:42:24	where to go
0:42:25	where to execute
0:42:28	to let to look about this because it this problem
0:42:31	in particular because E a very import
0:42:36	if you to could the it look at the experience domains
0:42:39	it it's experience i ancient
0:42:41	you can of course running everything in in in the device itself
0:42:45	which is the you know the this type of that mobile computer you could do the video editing here
0:42:50	you get the video you at to here
0:42:53	uh you get and the display
0:42:55	well we know that uh in the case of a larger things
0:42:59	this will be very very very slow all the user experience will be pretty bad and then in addition there
0:43:05	might be some other
0:43:06	sort of a user experience issues that yeah but it's basically little
0:43:11	of course we can do peer to peer so we can steal somebody's else computing interesting idea
0:43:17	taking a little bit more computational power or from the neighbourhood
0:43:20	to do the video eating
0:43:23	fine
0:43:25	or
0:43:25	what we can of course to ease that we think all okay we have this
0:43:29	yeah last the city up there
0:43:31	we have the clout so we send the P deal
0:43:34	uh to uh the the actual data
0:43:36	up to the cloud to be processed there where you can the fast
0:43:40	process of course there are you add now talk
0:43:43	to transmit
0:43:45	so in the user experience die main should you one of the things that you want to use lee
0:43:50	from a user perspective
0:43:52	to uh
0:43:53	uh just optimized is the time
0:43:55	it and can't use it doesn't care where it happens you just
0:43:58	as soon that these meeting
0:44:00	we are living in a very be a
0:44:05	well
0:44:05	this is the pure experience per spec
0:44:07	if you look at the economics perspective
0:44:10	oh this one
0:44:12	again
0:44:13	if you to everything in the device itself not is that i now divided in that's a way that the
0:44:17	law were in use the device upper in these the cloud in in the middle this of transmission
0:44:22	somebody have to pay for the clout
0:44:24	so but in the the day it usually doesn't come from
0:44:27	from a from nowhere
0:44:29	so
0:44:29	you have to you have to pay for that economics of the cloud
0:44:32	but form a a a a again at a user perspective
0:44:37	you also paying on the transmission data com
0:44:40	in most places
0:44:42	though that will also be expressed
0:44:45	but then we come to the very fundamental question
0:44:49	if it do
0:44:51	the video at eating on the device
0:44:54	you're running out of energy in the car and
0:44:57	mode very fast and we all know already that that even taking pictures
0:45:01	not even at thing on the video well run out of our batteries very
0:45:05	so okay
0:45:06	not a very
0:45:07	not a very feasible thing
0:45:11	so what about if we just
0:45:12	put it on the clock
0:45:13	so i mean then we don't run out all sort
0:45:16	the next thing is do we can steal our neighbours energy
0:45:20	doesn't make you very popular
0:45:22	because the the person who want to make a call next time and doesn't have any energy anymore
0:45:27	probably doesn't like very much the idea that uh uh uh he or she has support your
0:45:31	video at T
0:45:34	if you put go to the cloud side
0:45:37	we all know the problems already that
0:45:39	this is like a
0:45:41	no free lunch situation
0:45:43	the cloud
0:45:45	server from use is also and it might be a again G
0:45:49	that is not immediately effect you
0:45:51	but as a total off
0:45:54	somebody will have a problem
0:45:55	with the in G and we're talking about a green data centres nowadays a lot
0:46:00	and
0:46:00	this also fundamental lame
0:46:03	which is an interesting question that how much can you concentrate
0:46:07	in in a one place because you see in there
0:46:10	bottom uh approach
0:46:12	you're introducing new energy source every time you're uh introducing a new device that does to computing
0:46:18	in the cloud
0:46:20	the low energy load grows linearly
0:46:23	at the and uh of the cloud side with respect to the customers it has to sit
0:46:28	but even more fundamentally
0:46:30	this trade
0:46:32	and for D in real life
0:46:34	tends to be controlled by the fact that sending bits
0:46:38	but much more energy than computing
0:46:41	part it is because of the the coding in that we have
0:46:45	you know we are so far from the
0:46:48	uh channel limit made that basically uh we are essentially
0:46:52	uh i have to put
0:46:53	much more power than we would need to
0:46:55	to send uh the but be to to correct the noise errors that we have
0:46:59	but basically
0:47:01	this balance is good to did date
0:47:04	almost all of the the the data manipulation i've been talking to do
0:47:08	the balance between what do you compute locally and you energy
0:47:13	where do you put where you have the energy of here
0:47:16	and then what's the transmission energy
0:47:19	did you are doing if you think about rendering for example watch screens directly from the class
0:47:26	and
0:47:27	i wanted to point out that um
0:47:29	academically
0:47:31	energy a a fundamental thing
0:47:33	in a G something you cannot chi
0:47:35	so it's very appealing theoretically
0:47:37	in real life systems
0:47:39	the the economics and experience
0:47:42	are the ones
0:47:44	that really also dictate what will be uh used in pratt
0:47:51	so my question
0:47:53	i sort of problems that i one lee would you
0:47:56	is that
0:47:57	basically
0:47:59	we know
0:48:00	and that shown that we have multiple sources of data available
0:48:05	the question is that
0:48:06	what's to a kid had talk so how do we capture parse and analyse on them on the fly meaning
0:48:11	that that one do things on wine
0:48:14	and radically different source sees this is um
0:48:17	uh you know the sense sorry fusion he's one term that people use
0:48:21	and different
0:48:22	commune his we use different terminology
0:48:24	to more me just the head of is attributes of source of that
0:48:28	the second question ease the architecture question
0:48:31	what what like take to build these socket to that actually a robust but
0:48:35	and have strong elastic properties
0:48:37	how do you right go out it means such a way that
0:48:39	and
0:48:40	even if you running in the signal much
0:48:42	i would write the running T with what one billion device
0:48:46	cross the uh
0:48:47	maybe with the cloud component
0:48:50	the third question
0:48:51	is that
0:48:52	how do we tackle this energy efficient computing
0:48:56	because in fact
0:48:58	and
0:48:59	like i say
0:49:00	and from the theoretical perspective almost like a the perspective that you had this after a four and model
0:49:06	you can look at energy as a
0:49:08	the different component there and do a lot of analysis
0:49:12	even a like we need to balance environment concerns but this is practicalities
0:49:17	and the in user experience
0:49:19	for this and just
0:49:20	now
0:49:22	why Y in or G so fundamental to me
0:49:25	it's fundamental for a reason that this first time
0:49:29	in in my lifetime
0:49:31	we are reaching the levels that then i can argue that i can foresee service sees
0:49:36	that had not be bill not because of the call reasons but because of the reason that we don't not
0:49:41	have enough energy on yeah
0:49:43	the runs at to
0:49:44	six billion user service
0:49:47	what a run on the device and what is run on on the on the back
0:49:52	so this is what i wanted to leave you a today
0:49:55	i just one to remind you that my
0:49:58	medic search
0:49:59	for the medical data sets
0:50:02	that quest actually has been filled
0:50:04	i have always think than things and and a well
0:50:08	exciting me there
0:50:10	lying there
0:50:10	and now i see that about twenty five years of my life
0:50:15	i haven't solved very many things related but this
0:50:18	this is
0:50:19	i've sold a lot of things would related but i don't think
0:50:22	would you guys in a community with the great P H student is that i've have the privilege to work
0:50:26	with
0:50:27	but basically
0:50:29	i guess said
0:50:30	i now know
0:50:32	that we are facing an air a
0:50:34	so there are almost need
0:50:37	to our exist then is that we as a research communities
0:50:40	need to address in a different way
0:50:43	okay
0:50:44	so
0:50:44	that's what i want to say to stop thank you very much for your attention
0:50:55	i Q for very exciting presentation
0:50:58	we we have a very
0:51:00	challenging research problems to work on with a whole community here
0:51:05	so problem
0:51:07	in in them
0:51:09	scientific uh uh uh four
0:51:11	as
0:51:13	for
0:51:14	global
0:51:14	a local uh
0:51:18	energy consumption and problem of the whole or
0:51:21	and uh we have a time for couples or questions so please
0:51:30	okay
0:51:31	as mike so
0:51:32	excellent
0:51:33	i
0:51:33	very much
0:51:34	oh
0:51:35	two
0:51:36	two we
0:51:37	a
0:51:37	so i know it's your
0:51:38	or good thing of be important in this environment
0:51:41	our privacy and security
0:51:43	and so i
0:51:44	i so the like
0:51:46	my own close
0:51:48	excess
0:51:48	there
0:51:50	have
0:51:51	yeah
0:51:51	for all network
0:51:53	we work
0:51:54	yes
0:51:55	and
0:51:56	i deliberately chose
0:51:57	not the say the work privacy and security
0:52:00	as somebody characterised rice me a long time ago that henry
0:52:04	if we reset aspect you always doing is violating
0:52:08	i
0:52:08	C really people's privacy and that's of them
0:52:11	and no i think it very seriously yeah i just decided to leave it for the question because i knew
0:52:16	that the question what car
0:52:17	and
0:52:21	of course
0:52:22	first of all
0:52:23	at so the that they can very sign this perspective
0:52:26	the more we get or information
0:52:28	the more i can reverse engineer
0:52:30	that's a fundamental that and i could even do it in the ways that
0:52:34	people think it uh you know you you and and mice thing sell you at run
0:52:38	noise but if you can predict the noise model you can reverse engineer a lot of stuff and you could
0:52:42	do very complex things
0:52:44	now
0:52:45	for me first of what privacy is always a trade off it strike a reliability really
0:52:50	this is certain pay all you get from something
0:52:53	and certain cost that you have
0:52:54	if the cost
0:52:55	is higher than the pay off you should not do it so the cost for your privacy
0:52:59	and and the aspect you should be able to first of all use and always be able to opt out
0:53:04	that that's that's a the the first
0:53:07	but the second i of point out that that learned that and never thought so much was that the or
0:53:11	a lot of things where you could do a a or trained
0:53:15	how to get and alice
0:53:16	without violating any kind of privacy ask
0:53:19	you basically
0:53:20	at hating like this traffic
0:53:22	sure sure the ear
0:53:23	that does that just totally and no anonymous
0:53:25	no idea who is their don't will the individual points are followed
0:53:29	in a sequence
0:53:30	yeah just point in the time in the data say you don't know that they coming from the same source
0:53:35	you can still do a meaningful alice
0:53:37	that's the first good
0:53:39	second good nice for research community that the are privacy preserving make is
0:53:44	that you can build
0:53:45	in this
0:53:46	yeah sort of put the traffic
0:53:48	ways of handling this thing so it's a future research brought
0:53:52	and
0:53:52	a third question is that
0:53:56	i usually do the channel see how many you guys in this room has actually to of your cookies
0:54:02	i mean just to be very popular
0:54:05	just to be very popular but also this some if fit of not turning of them more relays in this
0:54:09	and that again
0:54:10	so i E each very cultural location only what happens but i hope we you
0:54:15	and and my main point always is
0:54:17	uh and it is nice
0:54:19	you should know
0:54:21	and use should be able to opt out
0:54:24	but if this it twice
0:54:25	if you want to uh a a a a get some benefit out of that information that is available
0:54:29	security ease of different ish
0:54:32	i i think everybody shares the the the question of the security problems uh the concept of the security problems
0:54:38	that we have
0:54:40	and and D sub very complex is used again called zero even regular to issues
0:54:45	that we face a different ways in your a
0:54:47	a in us any nation
0:54:49	my lattes are working and i'm working on growth economies to
0:54:53	was than well and i can tell you that these issues a very different in different
0:54:56	mark
0:54:57	security uh protocols are a good research topic do i think the the privacy consent tends to be the higher
0:55:04	one among the people a we know all the problems that currently are
0:55:08	face books and google another side facing it is but don't
0:55:12	i decided not to it over in for size this because i was talking from a size perspective
0:55:18	and thus a recess perspective but we have to be of course it it just about that a a work
0:55:23	we do to
0:55:26	right right no short question
0:55:28	and hopefully a ask

Making Sense of a Zettabyte World

Plenary Talks

Přednášející: Henry Tirri (Nokia Research Center, Palo Alto, CA, USA), Autoři: Henry Tirri (Nokia Research Center, Palo Alto, CA, USA)