Přepis řeči - Trends in Signal Processing Theory and Methods

0:00:13	so the person methods
0:00:15	um this paper
0:00:19	okay and
0:00:20	assume the presence of so
0:00:22	when
0:00:23	for
0:00:24	and the lab as comedies and
0:00:27	uh_huh of a
0:00:29	speech two systems where
0:00:32	two
0:00:34	a this just that the sequence of presence an to assume a present and number some the and so on
0:00:40	so should just a
0:00:41	but most work two
0:00:43	a is a web i'm um uh uh if
0:00:47	become a member of members are bones were exposed was in like to as or have
0:00:53	has been to do so
0:00:55	um uh if you want to become a member can the to and of the members of of a spectrum
0:00:59	to some
0:01:00	uh
0:01:01	and two
0:01:04	um
0:01:05	we have a
0:01:06	to
0:01:08	as shows but that's prince
0:01:10	and uh
0:01:11	the fruit uh well as a
0:01:13	perhaps or i'm so it from us and uh
0:01:17	is
0:01:18	but a man to talk about the
0:01:21	process the mission
0:01:24	so
0:01:25	this session
0:01:26	as the presentations of them after each presentation of the few is for questions
0:01:32	and and the um
0:01:33	more to have come to ask questions to a sparse image
0:01:39	okay thank you very much on the hawk and that would like to thank of few for coming
0:01:43	have have coming here
0:01:44	unfortunately i have one of them right of the this so i apologise i that in that's if i'm of
0:01:49	a little bit
0:01:50	fast than usual
0:01:51	no as you can imagine a a a a speech M is very but are we cover lots of a
0:01:55	has lots of interesting at as
0:01:57	uh uh in the general of statistical signal processing adaptation and learn in compressive sensing and so forth
0:02:04	and and and be a question was
0:02:06	you know i
0:02:07	should be but a but but are or should be a big some topic what want topic or two topics
0:02:12	and but that can be deeper
0:02:14	and um
0:02:15	for consultations with of a common with the of we decided let's pick a topic "'em" at least
0:02:20	expose at the m-th at least to what's going on in that at and many fat interested of course you
0:02:24	can go
0:02:25	uh i i i and check the reference and a to us and would be not to be you more
0:02:28	and more
0:02:30	yeah
0:02:30	input put into that
0:02:32	no i would be talking and or so of a come would be thought of it becomes a possible compliment
0:02:36	my
0:02:37	essentially that are two uh a general approach is that and women each other in this area yeah and very
0:02:42	excited about the same as that going to see he takes adaptation and learning to new levels it adds a
0:02:47	i have to this
0:02:48	a to this very exciting a year i usually like to to start my by showing this fit you're
0:02:54	a piece i have a video here
0:02:57	i a look about to i one to see how used to something
0:03:01	i not sing anything okay
0:03:03	so maybe somebody can help because this was playing a or level
0:03:11	see something
0:03:17	much showing anything
0:03:20	to just so i
0:03:22	in
0:03:24	that see if we have a thousand about
0:03:26	okay okay
0:03:26	uh
0:03:28	that meant to see here is
0:03:30	for a lot of them but it's a i'm down to the left them to the right now you tell
0:03:34	me
0:03:35	that is not a but tell them what to do that is not but i think these that the words
0:03:39	and what direction we should go
0:03:40	we have a thousand people in a state will they hit each other or not do you tell be find
0:03:45	a lot the people that want to on or but each other
0:03:48	how but does do that
0:03:49	that's incredible behave i will go two algorithms they you were processing that you like learning that you was how
0:03:56	they had job
0:03:57	okay a to a two that was i um them at this is a highly sophisticated be here but if
0:04:02	you if you look at the dock in system each but
0:04:05	is a the processing information just a it to make decisions about how to go on to to go
0:04:10	and you have seen patterns of intelligent behaviour like this H nature would probably just
0:04:15	go to
0:04:16	but the me when it just a look at the just start time to understand but yeah going to see
0:04:20	how D this behave at is how difficult actually is to produce that kind of behaviour okay
0:04:26	sampling to this is a you to video about the up my video
0:04:29	you
0:04:29	so
0:04:31	and
0:04:36	okay
0:04:37	now
0:04:38	but some the example of intelligent him and that you see that
0:04:42	we have a very very smart man and here you have an example of this is a picture
0:04:46	a a and joe graphic just look are they
0:04:49	a a a a all that have the colour last same most and parallel some
0:04:54	such they see an obstacle that not
0:04:57	and a nice i that the shape and then they can call to them should now you tell me how
0:05:02	to achieve that kind of behaviour
0:05:04	is that in the fish that and key a position yourself at this location and you position yourself at that
0:05:10	location
0:05:10	that's not what's happening and this kind of smile in go be behave that that just to that is the
0:05:16	that of highly localized information process
0:05:20	i come information processing can do you know that to and that kind of sophisticated behave but this out of
0:05:26	N one questions and a very interesting question
0:05:29	i and if you look at the literature of course people have we study this or many many using the
0:05:34	fun disciplines them by would you complete the science
0:05:37	and uh in a colour G
0:05:39	and and i think now
0:05:41	a seven is an hour in our field uh uh a include this because we had also time to to
0:05:46	to understand as best as we can that kind of behaviour and to to produce it not that we want
0:05:51	to
0:05:51	and and make and my behaviour but i think that is something come from them to back to make a
0:05:56	algorithm as the smart a robust morning just a
0:06:00	a of and so for okay
0:06:01	yeah a just but its of information
0:06:04	but also by the a from that the for that i had but also a fly in your formation
0:06:08	the have sense for that okay
0:06:10	and for the call reasons for that but see it
0:06:13	before the
0:06:14	ration to this kind of that this i that time to be the first one and then a end and
0:06:19	the for to the the bound to flight that a and like this is that is out of highly localised
0:06:26	in information processing that is and this kind of
0:06:30	i
0:06:31	a rational behave and how do they do it that interest be have yeah right
0:06:35	so
0:06:35	a more to look into it get
0:06:37	i you get excited at that this is another example of some which are P about a to use that
0:06:43	well if some some press i have a it video to show you here
0:06:46	but that like to see
0:06:49	so if you were
0:06:50	this time
0:06:52	okay
0:06:52	but that somebody to i mean lots of that you have a this sub time
0:06:56	and and it was that i
0:06:57	in in a dish
0:06:59	where is i believe was some oxygen and like where the made and then the but this time
0:07:03	and there is at least i was size devices you can hardly see them
0:07:08	well
0:07:09	that's that that's you to left on the item and them on a motion
0:07:13	now when you all the of the and that i score my micro quite this is what to see this
0:07:17	kind of but
0:07:18	you see these time of devices is or and thus the in
0:07:22	a the set that action
0:07:23	so that's a here a K i like T V "'cause" i don't know i'm just asking questions that a
0:07:29	back to you
0:07:31	oh this kind of high
0:07:33	gram may behave at a notation in a single direction that that's an interesting kind of behaviour
0:07:38	so that's happening at me
0:07:40	a and we had a eyes level of so this seven
0:07:43	we this thing questions that come up when you stop observing
0:07:47	this kind of behaviour in nature
0:07:50	okay
0:07:51	so if we look at these examples
0:07:53	case
0:07:54	was
0:07:54	but to be okay
0:07:56	a much of to see is that what each individual H and in a i is not sophisticated
0:08:02	but them together
0:08:03	that
0:08:04	a the interaction action at that highly localised level needs to
0:08:10	oh patterns of behaviour or people have of this for years of course
0:08:14	look at the question now for us
0:08:17	am and just uh a show it on the next the slide but this is the convolution of this examples
0:08:20	that to see here examples are
0:08:23	actions and
0:08:24	the agents
0:08:25	i to sophisticated complex behaviour of course on S
0:08:30	as i like good as that can do nothing i'd these it to have some
0:08:34	some cognition in them okay know
0:08:37	okay but when they put them to to have that it leads to that kind of interesting behavior so can
0:08:42	and our a little is
0:08:43	but i in statistic signal processing an adaptation and learning in cognition in machine learning can i
0:08:49	and make that kind of intelligence
0:08:52	that lots from that the
0:08:53	but but that's in a question that will like
0:08:56	for example an important questions that to make a so that
0:08:59	i'm i i behave i i i is is
0:09:02	i i love for a kind of that action is you to have to generate a data set kind of
0:09:05	sophisticated be here the might be interested in a big a level
0:09:09	a of G
0:09:10	should exist between a and so that they show an information in such a way that that is and be
0:09:15	kind of carbonated behave but that you would like to result
0:09:18	okay
0:09:19	how much should information be quantized
0:09:21	should be used at and share information at the very high presentation level is
0:09:26	low quantization level enough i'm sure that the fish number see a shop but they can shared information like this
0:09:32	P of the shop but the colour of the shot that action but just a segment and that's that's that
0:09:38	just a shot
0:09:40	i i and everybody gets i about the shark
0:09:43	but this is one example of quantization how much information relation position
0:09:48	such a that that kind of behaviour
0:09:50	yeah appears and the last question is then a of them you know if you think about it
0:09:55	at least example that i sure that have a static networks
0:09:59	and is at a time
0:10:01	maybe
0:10:02	and may was five minutes from now the connections that change of the time
0:10:06	this is that high networks and a T had
0:10:09	and for is learning that inference is mission okay and the other way and i also be a a a
0:10:14	a a to five and that may also in france to an environment "'cause" of this call this questions are
0:10:20	tied together and i decide
0:10:23	address
0:10:24	a from the perspective of a a a a two a from the perspective of signal processing in general are
0:10:28	there are many many techniques you come up i
0:10:31	that this problem like this to the calm is going to but also about some additional techniques which are should
0:10:36	have listed yeah
0:10:37	so a you have this to the process can of course is at the heart of a lot of this
0:10:41	for the estimation adaptation and and the assume game phonetic a and methods
0:10:47	uh i i a statistical as
0:10:50	that you can use to try to but that kind of behaviour
0:10:53	i like i said
0:10:54	people problem uh that uh that's have this these behaviours as and they have exploited that
0:10:59	and a different ways you even see and nations of naked behaviours in movies
0:11:03	mag
0:11:04	but i have a computer that science have generated an issues
0:11:08	that you see are going to a uh uh i
0:11:11	a to produce this kind of behaviour yeah about insist of course
0:11:14	and the the different level of processing the different kinds of questions how the whole algorithm
0:11:19	that can generate
0:11:21	that kind of high level intelligence
0:11:23	right from high localised it's set
0:11:26	actions at the local level look at this is a high level question we have mine
0:11:30	no of course i don't have time have to gossip algorithms that
0:11:34	sure
0:11:35	i as we have a we have a a a a a large family of
0:11:38	i a
0:11:40	can can and i guess
0:11:42	adaptation i'm of such conditions you have more networks
0:11:46	and that that actions that local i with them up to discuss the algorithm
0:11:51	this the of everything but i would like to show some simulations examples to example
0:11:56	are the things that are able to do that
0:11:58	oh
0:11:59	and interesting thing this of a lot of them to come up i to my be B here for a
0:12:04	in the back to the that behaviour fish a a
0:12:06	and have that as so that that is something that are okay for example a to to see in the
0:12:10	first example
0:12:11	a a fish
0:12:13	i don't know
0:12:14	okay
0:12:15	we to find a different is that want to to gather have to make to have a fine but the
0:12:19	for this and in that direction
0:12:22	and that's not a shock peers
0:12:24	so they have to be a bit of that environment they have to also to that where the shock
0:12:28	so for that is the first
0:12:30	object of shop as a moving object that they have to track it and this problem
0:12:35	so i went to see the simulation this is not a kind of application i would like to
0:12:40	sort of consensus
0:12:42	because if fish that's scores
0:12:44	as to behave in a different manner than a fish that's found away from the shock and the before
0:12:49	this is a kind of example that the as we to have something of age
0:12:53	i have like to to add that are but in addition to apply to them from the but it's not
0:12:58	enough to for a to teach my senses would you but you have to be a a a a a
0:13:03	a a and and that's that's the situation
0:13:05	but on your on nice estimate
0:13:06	so he had to have
0:13:08	and and this is just
0:13:10	the of an adaptive a cognitive networks are i i-th that to get that in the and to which are
0:13:15	with each other and learning from each other and but skip that
0:13:18	and this is an example you see that for a is that that i
0:13:21	on that i i a that you have before and the fish like to have as before
0:13:25	no this is a a and that if network okay it's
0:13:30	went to see it from the stop down here
0:13:32	that that was the for for a i think that actions and then that but that is present in the
0:13:37	environment at two
0:13:39	and then go
0:13:40	i
0:13:41	this is using a a an adaptive algorithm on the kind that should be for nothing or
0:13:45	yeah a highly localized that action
0:13:48	and is out and to make the kind of used in H okay
0:13:53	this is what zap of the next example of to see two networks working dance each other and also X
0:13:59	the the kind of be to find in the true that that's the group of shots
0:14:02	time time back a of fish
0:14:06	in of the so of the fish
0:14:08	a so you have a network of shocks score than a with each other
0:14:12	so that they
0:14:13	the fish and the network of fish score than a with each other that so that they know if it
0:14:17	is where the shots i don't not they should do or that's what are able to see next example
0:14:21	okay
0:14:22	so he it is
0:14:23	a function of mm the shots B
0:14:27	also also that have the kind of a but this is a adaptation and
0:14:31	oh
0:14:32	and this is a kind of thing that i like i said
0:14:34	you can expect this fish to check check limit of everybody else that i
0:14:38	i
0:14:39	to say of themselves are K they do that best like an a show then neighbours but not it comes
0:14:44	time to decide for your life then you have to take some decisions on your
0:14:48	or or or okay so we start to examples that i not to show a about how patient and let
0:14:54	as in high and i make a bad environments of every multiple agents so give some going to stop you
0:14:59	a from this gives an overview
0:15:00	okay i five
0:15:02	yeah yeah is okay
0:15:09	uh okay so if there questions of you have any questions as a consequence
0:15:14	oh
0:15:17	and
0:15:18	a
0:15:19	and and uh i i thank you thank you about sally things
0:15:23	okay can i like to uh
0:15:25	i ask become "'cause" number of T are also
0:15:29	can has i a um about
0:15:33	oh
0:15:34	signal process and thank you very much
0:15:36	it is going to talk about uh
0:15:39	if
0:15:39	as a
0:15:41	a like to to that this that's per session is not a summary of what's happening at time "'cause" mom
0:15:48	i'm sure
0:15:49	or we find the uh that is that the that uh put to but we were try to to focus
0:15:56	here
0:15:57	a a a a a as and we can't cover everything so opens the pressure
0:16:03	we can uh
0:16:04	oh have an appreciation of what the
0:16:06	where the we thank you and thank you are so uh uh a a a i only have two slides
0:16:10	here
0:16:11	a in the first i'd i just want to give a incomplete list of several
0:16:15	oh important areas
0:16:16	based on submissions and i guess this year and i guess a extrapolation of what might be interesting
0:16:21	so this is an incomplete list is always these for like popping to some few leagues and i apologise to
0:16:27	any of that
0:16:29	so
0:16:30	these are six
0:16:32	here is we've listed here
0:16:33	the first two signal
0:16:35	processing a signal dynamic rows and as many of you know that is actually a fairly important here at the
0:16:40	moment
0:16:40	we but looking at consensus formation information flows games
0:16:44	so that
0:16:46	and a
0:16:46	quite multi disciplinary as well you you think about it
0:16:49	not only does it affect signal processing but i guess
0:16:52	if you looking at things like social networks you point to model house a siding functions in terms of some
0:16:56	of these
0:16:57	a dallas sings out with
0:16:59	then another area which is also
0:17:02	uh of
0:17:03	in not most that our T in in
0:17:04	take are are uh S P D and society is
0:17:07	a dimensional signal processing get sparsity
0:17:10	mean uh if you look at this year's i guess is been
0:17:12	uh enormous number of people in that area
0:17:14	oh from a mathematical point if few many of these results to you with
0:17:18	this very uh seminal idea of concentration of measure your room
0:17:22	and and
0:17:23	compressed sensing is one such example
0:17:25	oh many of you would known that there are
0:17:27	special sessions uh in i cast on that
0:17:30	the transaction a signal processing which is you must papers in that area
0:17:33	i so the to special use of journal selected topic
0:17:36	the processing
0:17:37	again it's multi disciplinary
0:17:39	in people and computer science machine learning and also by a medical imaging
0:17:43	the uses that
0:17:45	another area which is also a
0:17:47	significant importance
0:17:48	coming from signal processing this financial time-series
0:17:52	traditionally a people in mathematical finance do things it continuous time
0:17:56	the use things like stochastic was and so on but
0:17:58	there is also a because we did you do things a discrete time and and those was are
0:18:02	people do financial time to
0:18:04	and of course is a smart grid and and most of us know that that's something with a lot of
0:18:09	a a potential
0:18:10	as distributed sensing decision making a control and that's really what i'll was talking about it in this part of
0:18:15	the top my next slide i wanna give a few more ideas behind that
0:18:19	so you can think of these as called a systems the uh ability of a system to use feedback
0:18:24	to we can figure its behaviour and on top of bit more about that later
0:18:28	and and think this is very multi disciplinary
0:18:31	or people in economics of done some very serious work in this area or uh and and also with other
0:18:37	is
0:18:37	and finding the signal processing in the life sciences
0:18:40	so we
0:18:41	comes a a laid down from
0:18:44	oh molecules how you can simulate the behaviour of small molecules
0:18:48	and how drop spine to small molecules
0:18:51	that really up to a large scale by medical imaging and instrumentation
0:18:55	i just found as little uh snippet but from looking at the nature but side nature or column nature one
0:19:00	of the highest in that turtles and by
0:19:02	uh and as a top here
0:19:04	so that actually companies like pfizer or and no part is which are drug companies
0:19:08	uh investing
0:19:09	billions of dollars
0:19:11	in high speed stochastic simulation
0:19:14	at the molecular level to see how drugs
0:19:16	buying buying to Q
0:19:19	and we use
0:19:20	mark of T the column
0:19:21	so so that's
0:19:22	something which i guess people in signal processing to quite a bit of
0:19:24	and uh you might be interested in the things that but that map lab is also developing a
0:19:29	so i all the G two which allows
0:19:31	no no uh people to
0:19:33	and i
0:19:34	chemical part ways
0:19:36	and also estimate parameters and sensitivity of these part we so so it's be a very active area
0:19:41	as the normal significance to signal processing
0:19:44	that so that's
0:19:44	a but incomplete list of some areas of importance
0:19:48	i and i one a in my last slide simply focus on the area of distributed sensing which
0:19:53	which i guess uh what
0:19:54	or something which are ads on to what all these said in his still
0:19:59	so that's just kind of a a look at the big picture your once again
0:20:02	so it's standard signal processing we do things open loop we have a sensor measuring
0:20:07	a a signal and noise
0:20:09	and then we want a process that noisy measurement to estimate the the like signal
0:20:14	no you can think of a philosophical that bounce in that were use some form of feedback where
0:20:18	the estimate
0:20:19	a if signal
0:20:20	this fed back to the sensor
0:20:22	so that the sensor can reconfigure its peak a pure in time
0:20:26	one such example could be
0:20:28	a distributed system where
0:20:30	the the picture of the right hand side
0:20:32	but you have multiple agents
0:20:34	sensors or agents
0:20:36	and each should are one individual control unit
0:20:39	i each guy wants to we figure its behaviour to do something
0:20:43	so i mean one can think of a that experiment where perhaps
0:20:46	you sensor so is more drink some sort of target
0:20:49	too many sensors to and all the the wasting battery life
0:20:52	so they don't all all to at all
0:20:54	to a few sensors to not if too few measurements
0:20:57	and then if you do what day of data fusion and you have a very high variance
0:21:01	so
0:21:02	all the the sense decide
0:21:04	without a thing to all the sensor should turn on a roll
0:21:08	keeping in mind that
0:21:09	it's not completely capitalistic
0:21:11	was company capitalistic it would say to have with everybody else i what sleep but not to at because that
0:21:15	present about battery life
0:21:17	so a company socialistic goes that it would simply be to all the time but out to die
0:21:21	and uh wouldn't be very useful
0:21:23	so without talking to other people
0:21:25	by using rationality the ability to predict what other senses of there to do
0:21:29	how does the decide when it which is on a is it's which off
0:21:32	now what economists do uh and really a talking in this slide from economics point
0:21:38	an economist would do to add a lie such a problem is they would introduce the assumption that each agent
0:21:43	is rational
0:21:45	a we we say to devour rationality that of actually a very similar in some sense
0:21:49	so by shall only needs the following that if i get a measurement
0:21:53	i i can put big from that
0:21:55	right
0:21:56	are the agents a gonna do because all the had estimation
0:21:59	so i are use by measurement to predict what other people are gonna do and what i a
0:22:03	based on that
0:22:04	and i don't people know that i'm gonna react to a measure it in the way i'm gonna do
0:22:08	so that you know what i'm doing
0:22:10	but you know that i i know that the you know what i'm doing and so on ad infinitum that
0:22:14	really results is something the called equilibrium
0:22:17	a types of put the breed a because if you're
0:22:19	nash equilibrium correlated equilibrium so on
0:22:22	but the main idea being that here as a system on the right hand side where eventually if things were
0:22:27	done correctly
0:22:28	individual agents like fig
0:22:30	or shots or whatever
0:22:32	can we can figure the because you would minimal communication
0:22:35	to achieve a common goal
0:22:38	so
0:22:40	one can play was your at all three are really
0:22:42	which people and micro economics of studied in great detail of the last ten fifteen years
0:22:46	uh
0:22:47	the first question can be
0:22:49	how can agents autonomously
0:22:51	man the behavior
0:22:53	and that's really the i guess the question i poles
0:22:55	now
0:22:56	you can think of it off sure of that as
0:22:58	if each image and that something simple
0:23:01	how can you ensure that it a little your is complex
0:23:05	or ration
0:23:07	and again a uh i i just what a point out to one it'll paper we didn't by so your
0:23:11	heart
0:23:12	use who's a really famous economist
0:23:14	i to look at it can a metric or two thousand and five
0:23:16	yeah the fifty page people which gives you an amazing perspective of how
0:23:21	individual a in a can we do something very simple
0:23:24	like grand a simple adaptive filter
0:23:28	the performance is remarkable
0:23:30	but eventually every agent at a user consensus
0:23:34	not just good as an estimation
0:23:36	but can as an action that is for all is eventually
0:23:40	this same decision policy C
0:23:42	so they have some remarkable to that in micro economics how
0:23:45	to come up with some
0:23:46	and the to the user is really game theory as an analysis but that all
0:23:51	also so into reading the learning out with them
0:23:54	which further refine how agents can act
0:23:57	a second question one could pose
0:24:00	the standard power in signal processing
0:24:02	typically typical one is looking at agents and sensors
0:24:05	each sensor has measurements
0:24:08	and it up estimates based of those measurements
0:24:12	now you can think of a further the refinement of that where agents not only get measurements
0:24:17	but look at
0:24:18	actions
0:24:19	other each
0:24:20	so you look and see what of the people done
0:24:22	how could i learn from what other people have done
0:24:25	if you think about that
0:24:27	and i really is some sense
0:24:29	quantization an observation because the agent gets and observation
0:24:32	the agent processes that maybe no rational way by bayes rule and then picks an action
0:24:37	if you simply look at the action
0:24:38	it's in some sense a plant ties version of the observation
0:24:42	you but some information
0:24:43	so how can be agents learned from other agents that if you do this
0:24:48	do things actually work out
0:24:49	well there is this really classical
0:24:52	uh for example which has been studied in great detail in economics at i just one quickly mention that
0:24:57	results in something called a rational for
0:24:59	and
0:25:00	so just to give you the example of that
0:25:02	in nineteen ninety five
0:25:03	to manage but gurus
0:25:06	C and mum wrote a book
0:25:08	the market
0:25:10	for reviews so what they did was
0:25:12	we secretly for fifty thousand copies of the old ball
0:25:15	and i got the best seller a new york times
0:25:18	people so this is a best or it must be good T
0:25:20	everybody started by it
0:25:22	so it rational people
0:25:24	i know
0:25:25	because close following the actions of other regions
0:25:29	see if of the observation as opposed to the action
0:25:32	the observation would be who bought the book
0:25:35	and then you what was the all as buying the book they want to port so that do you act
0:25:39	purely only based
0:25:40	all all the actions of other and
0:25:43	you can show in many cases eventually or agents
0:25:46	and up sampling something called the rational for
0:25:49	the old eventually make everybody else
0:25:51	and they throw away the actual observations
0:25:53	so that means a very interesting work done in this area in economics and if you look at
0:25:58	for example the book by we
0:25:59	published by cambridge university press
0:26:01	and rational words
0:26:02	to to you a very interesting perspective on how you can model
0:26:06	complicated system
0:26:08	and how they before
0:26:09	now this also that obviously applications in sensors where you might put them but the she's agent and
0:26:14	which to look at the at a particular way
0:26:16	i it makes every the each would eventually act in that way
0:26:19	so you can think of a applications in in several areas
0:26:23	but one question which one it again can pose again it's and by people that micro economics as
0:26:28	how to local decisions
0:26:31	a
0:26:31	it will bill decision making
0:26:34	so uh in in a little picture or the better of the slide have uh a on the left hand
0:26:38	side a bunch of red and green dots that it's actually just a a a a a very straight for
0:26:42	example of a sequential change detection problem
0:26:44	so that as you me have a bunch of agents
0:26:47	each agent gets an observation of a particular thing
0:26:50	uh i and
0:26:51	we can't really the observation to the next guy
0:26:54	a can relate is you do they detect a change so it's change or it's not change to maybe green
0:26:59	means it hasn't changed
0:27:00	but means it's changed thing so
0:27:02	the first i it's maybe it's change a second that is not change and so on
0:27:05	each each make making a local decision
0:27:08	if you could or take these local decisions
0:27:10	how how a global decision maker decide
0:27:13	to say that there was a change in stop that's a standard quick as time sequential detection problem
0:27:19	turns out that the more you raise your
0:27:22	global decision on local actions
0:27:24	take scroll catastrophic leave wrong
0:27:27	so you can show that the optimal policy in this case
0:27:30	as a double threshold and to look at it's
0:27:32	hi country to it so that's of the right hand side of the ball
0:27:35	so on the X axis
0:27:37	as a probability
0:27:39	posterior probabilities a patient detection problem of change
0:27:42	and the S
0:27:43	S uh and the vertical axis is the decision each agent takes
0:27:47	so you can see that i as the belief
0:27:49	as a posterior distribution that the changes happened goes
0:27:52	no
0:27:53	you declare a change
0:27:54	amazingly but it gets even larger you signal does not change
0:27:58	the reason is you have a discontinue reading a bayesian update because you basing it based on actions of the
0:28:02	previous agents
0:28:03	and that of course but the around you say there is a change so
0:28:06	these multiple threshold policies are
0:28:08	cut very contrary to because why would you
0:28:11	she and decision if you get more sure about what you measure
0:28:14	and it's it's something again which is big study
0:28:16	uh in great detail uh by people in economics who really try to
0:28:20	more complex
0:28:22	phenomena
0:28:23	in terms of local and global decision
0:28:25	and how
0:28:26	decisions affect each of
0:28:27	so those are three things are just a i'd mention in terms of
0:28:30	applications of some of the methodologies all these side pose
0:28:34	a unifying theme behind all of this is how does not behavior affect global will behaviour
0:28:40	a if you like the by just you can view to say
0:28:42	if i have the structure
0:28:44	which is little cool
0:28:45	how could predict the function which is global
0:28:47	that's or by all the just do they have
0:28:49	the structural of different small molecules
0:28:52	a wanna see if these modules
0:28:54	function in the cell
0:28:55	how to predict what the actually do
0:28:57	given the structure and it's really really
0:29:00	that's all i wanted say thank you
0:29:06	X and you much read from so
0:29:09	can can see to it that to this you two has become a no research to i
0:29:14	and the
0:29:15	this is a uh one came out of these two do so
0:29:19	a the comments or uh
0:29:21	questions from the is i'm so this session is very short but uh
0:29:25	with the
0:29:26	a lot some questions is then
0:29:31	i as one when with had an example of to the case of the model of my malicious uh a
0:29:36	up lot of i yeah
0:29:39	so uh so i i i could do another example was i do quite a bit of mathematical finance and
0:29:43	many if you would do to do that all forty percent of all treating is done by computers
0:29:48	oh
0:29:49	where street runs now add second time scale
0:29:52	and and wants all buy and sell it a millisecond time see little quick enough also
0:29:56	a lot of the buying and selling is done by computer programs
0:29:59	no many people know that so if you want to design a computer program which deliberately for example started buying
0:30:05	your role
0:30:06	a the the the value for you would go well
0:30:09	if if you are a computer program was running from a reputable agency like woman sex a hope will be
0:30:14	go but set as you but anyway you would make
0:30:17	but we because of a reputation that the your is really going up
0:30:20	are these are the computer programs a custom trained to for a will once as doing they was start by
0:30:26	and then you would suddenly cell it's called a speculative can see attack
0:30:30	a big liens
0:30:31	and the for the other that's can react
0:30:33	that's a you're all of the market then you made a lot of money so this of things are widely
0:30:36	use an economics the call speculative currency C tax
0:30:39	and so if you look at for example how paper
0:30:42	she talks about those in great detail
0:30:45	to about but you shouldn't and not be should versions of such games
0:30:48	for
0:30:48	four
0:30:51	these are all of dynamical systems so not not none of this is static we talking but dynamical systems that
0:30:56	how you can control K
0:30:59	be another question
0:31:00	we four set session goes only till twelve fifteen
0:31:04	so round question or comment
0:31:08	if this is not the case are like to thank you again
0:31:11	a a to have a with the session and the C or not

Trends in Signal Processing Theory and Methods

Expert Sessions

Přednášející: Abdelhak Zoubir, Ali Sayed, Vikram Krishnamurthy, Autoři: Abdelhak Zoubir, Ali Sayed, Vikram Krishnamurthy