Přepis řeči - IMPROVING KERNEL-ENERGY TRADE-OFFS FOR MACHINE LEARNING IN IMPLANTABLE AND WEARABLE BIOMEDICAL APPLICATIONS

0:00:12	but the introduction
0:00:14	what i the like to put in C is variable and a the for applying were some many common
0:00:20	this is motivated by the need to money to increasingly complex
0:00:23	system using more process and uh
0:00:26	and the application we will consider of is it is by men a patient monitoring device
0:00:34	what i mean we that's say input again
0:00:36	or hmmm
0:00:37	what i mean when i say intelligence by method devices is
0:00:41	device is that are capable of of
0:00:43	deciphering
0:00:44	specifically a logical states you know patience
0:00:48	if we can do this
0:00:49	then the D this can actually
0:00:51	there are six really that as like
0:00:53	this big guys an every way to you know directly at closed loop weight
0:00:58	or
0:00:58	perform from corning king and monitoring
0:01:01	know directly way
0:01:03	or
0:01:04	you reliable for this
0:01:07	but the challenges that
0:01:08	the signals
0:01:09	there we can get from the body
0:01:11	or physiological a logical the complex
0:01:13	and difficult to interpret
0:01:17	to illustrate the complexity is
0:01:19	i have shown the example all
0:01:20	see section
0:01:22	based on a like trying to follow grass
0:01:24	or a easy
0:01:26	E easy the signal all you we can measure on the scale
0:01:30	a good thing is
0:01:31	this signal is available non invasive lee
0:01:35	but the problem is it does not represent a caesar activity with high C
0:01:41	for example
0:01:42	this strange and the brewers
0:01:45	a is G signal drinks lee
0:01:48	we need to discriminate this
0:01:51	with
0:01:52	the is
0:01:53	which corresponds to the onset of the season
0:01:57	the second challenge is that
0:01:59	the characteristics
0:02:00	of house teachers or at eight
0:02:03	or different from patient to patient
0:02:07	the excitation of read don't impatient a
0:02:10	is different from patient be
0:02:13	to overcome bossed is
0:02:16	on extremely the powerful technique
0:02:18	is
0:02:19	data driven
0:02:20	which the less us to construct a a or there
0:02:23	hi space be city model
0:02:25	of the signal
0:02:28	now that challenge is to apply these models with a low energy
0:02:33	a a line
0:02:35	our errors
0:02:36	discuss the end it's off
0:02:38	in data-driven modeling
0:02:40	in the framework of supervised of learning
0:02:43	then
0:02:44	our proposed a at the
0:02:46	for all were coming
0:02:48	and just scaling with more the complex
0:02:51	then our present our experiments
0:02:53	and we don't
0:02:57	we can
0:02:58	a a data driven model using supervised learning
0:03:02	the reason that the data driven thing is powerful is that
0:03:07	recently centrally data bases have emerged it in clinical the domain
0:03:11	where if is a lot of seeing a lost
0:03:13	recording the in hospitals
0:03:15	are
0:03:17	uh are sort of along with the clean call annotations
0:03:21	but
0:03:22	what makes it important for low power devices
0:03:25	it's that
0:03:26	the same signal also available in these devices
0:03:30	do and but tori recording technology
0:03:33	so
0:03:34	but low power devices
0:03:36	can directly
0:03:37	take advantage of the data driven models
0:03:40	construct the from hospital they are based
0:03:46	the typical framework of the data driven detectors is shown here
0:03:50	there are two phases
0:03:52	the first is it's training
0:03:54	and the second phase is able power real-time detection
0:04:00	training involves constructing a a a high or high specificity the model
0:04:05	from previous observations
0:04:07	that have been assigned with clinical labels
0:04:10	but this phase is assumed to one offline and you frequently
0:04:17	detection occurs continuously and in real time on advice
0:04:21	so energy is concern
0:04:24	with detection
0:04:26	there are two components
0:04:28	so the competition
0:04:29	first
0:04:30	feature extraction
0:04:31	and second
0:04:33	classification of
0:04:34	features by applying be high of the remote
0:04:39	feature are extraction
0:04:40	does not involve mobile mean
0:04:43	here we simply major D of the use of
0:04:46	wonders
0:04:47	that we believed to be correlated
0:04:49	with the states of interest
0:04:53	and record them by markers
0:04:56	it's a job of the classification to G screen eight is correlation as
0:05:00	using the model the was construct
0:05:07	two and a nice the energy of detection
0:05:09	we have considered
0:05:10	to by medical applications
0:05:13	in detection
0:05:15	by are are spectre energy
0:05:17	extracted in eight
0:05:19	different frequency bins
0:05:21	from each each E the channel
0:05:23	or what three E pop
0:05:25	up to eighteen he the channels
0:05:27	this gives the face every them is not like key of force that thirty true
0:05:33	you know a real a detection
0:05:36	morphology of of the T C you form
0:05:38	sample of around the to as complex
0:05:41	is used
0:05:42	which are least two twenty one piece of a them is not weekly
0:05:47	the feature vector them is not a cure for thirty two
0:05:50	and twenty one
0:05:51	real fact
0:05:52	the energy
0:05:53	um
0:05:54	the energy just scaling of the classification
0:06:00	the next that is
0:06:01	cost buying features
0:06:03	to detect caesar's in um them yeah
0:06:07	we use a a popular machine learning classifier
0:06:10	chord a support vector machine
0:06:13	a conceptually and svm examines
0:06:17	um
0:06:18	"'cause" of vectors you know high the all that is space
0:06:22	you use this training data from
0:06:25	positive if
0:06:26	and they are two classes
0:06:29	and it's samples vectors
0:06:30	at the edge of
0:06:32	these
0:06:33	distributions
0:06:34	to represent a decision boundary
0:06:38	the set of selected vectors
0:06:40	a long boundary
0:06:42	it's called the support vectors
0:06:44	and these are used
0:06:45	in this
0:06:47	color computation
0:06:49	then to classify the incoming test
0:06:52	feature vectors based on the resulting sign
0:06:55	it can have function
0:06:57	okay
0:06:58	it's commonly used
0:07:00	to transform the feature vectors into a higher dimensional dimensionality space
0:07:05	which effectively allows the system some boundary
0:07:08	to be much more flexible
0:07:13	so that a a number of
0:07:15	a support vectors and the feature vector dimensionality thus
0:07:19	you to mean complex at of the color competition
0:07:25	in this slide
0:07:26	i the energy analysis of
0:07:28	features extraction and and classification
0:07:31	most like the egg create instructions assimilate
0:07:35	classification energy
0:07:37	but over the feature is there's an energy
0:07:39	by a factor of thirty and almost twenty
0:07:43	in these applications
0:07:45	so
0:07:46	the classification energy is what would like to focus on
0:07:51	and this
0:07:52	so if a part
0:07:53	has to keep are amateurs
0:07:55	which represent
0:07:57	uh more the complexity
0:08:00	a first
0:08:01	the number of support vectors
0:08:03	and second
0:08:04	a recognition a like
0:08:06	and the application we choose
0:08:09	stress each of these respectively
0:08:13	or the much requires
0:08:15	fifteen a house and a support vectors
0:08:18	right yeah six C S section we parse as many yes six hundred
0:08:23	but the feature vector dimensionality can be at
0:08:26	can be as high as for the two
0:08:28	which a least to the high classification energy
0:08:34	that's slide
0:08:34	all we discuss the importance of con non linearity
0:08:38	which the "'cause" is energy scaling with the number of support vectors
0:08:44	if the can have "'em" can a were of the in a function
0:08:48	then the colour the computation is that dot product
0:08:51	between the support vector
0:08:53	and that test spectra et
0:08:56	that as the linear
0:08:58	re can pull of X from the summation
0:09:01	no bowing the summation to be pretty
0:09:04	well what all the support vectors
0:09:06	so that we can all work on the in just scaling
0:09:10	the problem is that
0:09:12	the linear kernel this not provide sufficient
0:09:15	but but lady you know this some bound
0:09:18	in this example
0:09:20	a gonna be seen to prince
0:09:22	many of the non see the point
0:09:25	right to radial basis function kernel
0:09:27	provide a much higher flexibility in addition boundary
0:09:32	base a result
0:09:34	this has been widely used in by mental applications
0:09:39	but
0:09:40	when energy is concern
0:09:42	then we need to worry about the energy scaling
0:09:45	of the all of carnal
0:09:46	because the exponential function in the log of call
0:09:50	precludes spree competition
0:09:53	so
0:09:54	the con all nonlinearities importance for accuracy
0:09:58	what we need to worry about the energy scaling
0:10:04	in this work
0:10:05	um
0:10:07	we have
0:10:08	turned off all "'cause"
0:10:09	to not type of carnal
0:10:11	called the polynomial kernel
0:10:14	the problem on all
0:10:16	well for an intermediate level of flexibility
0:10:19	compared to the art of kernel
0:10:21	and because of that
0:10:23	it has not been widely explored you met are with them
0:10:28	but what is important here is that we propose a way to
0:10:33	dropped
0:10:34	wait reese truck this
0:10:35	all
0:10:37	um so that we can
0:10:39	sorry um
0:10:41	so is important here is that
0:10:43	we propose a way to structure the on kernel
0:10:46	that
0:10:47	permits spree computation
0:10:49	all well all the support vectors
0:10:51	so that we can all work on the in just scaling
0:10:55	the dot product and securing in the protocol call can be rewritten can in the back a modification for
0:11:03	that
0:11:04	you can pull out
0:11:05	the vector X
0:11:06	out of the summation
0:11:10	and then the computation becomes a a vector
0:11:15	matrix
0:11:16	vector
0:11:17	multiplication
0:11:19	between the test vector
0:11:20	in the new D gen
0:11:22	matrix
0:11:24	and because of that
0:11:25	we can all work on the just scaling with the number of
0:11:29	support vector
0:11:34	actually uh what our proposed restructuring
0:11:38	does this that
0:11:39	it alters energy trade
0:11:42	to illustrate the new energy is space
0:11:45	i have a and the lies
0:11:46	the energy profiling result
0:11:51	the first figure shows the energy with respect to the number of support vectors
0:11:57	in the in a kernel
0:11:59	the energy is constant and vol
0:12:01	but the actors is also low
0:12:04	which will see in the next slide
0:12:06	the all we have and the can should be a scaling
0:12:11	but we can all work can this in the plot the carnal
0:12:13	but using the composition restructuring that we propose
0:12:19	the accuracy of the plot a is a concern
0:12:22	and i we so the results in the next slide
0:12:25	the second figure shows how the energy scales with D feature vector dimensionality
0:12:33	because all the computation restructuring
0:12:36	that transforms
0:12:37	support
0:12:38	vectors into a decision matrix
0:12:41	now we have
0:12:43	for the energy scaling that than the here
0:12:47	but what is important here is that
0:12:49	there are several to every
0:12:51	and several application specific techniques
0:12:54	they we'd used the feature recognition T
0:12:56	as always so
0:12:58	and because of that
0:13:00	this
0:13:00	energy trade
0:13:02	i give us a value the over option
0:13:07	in this slide
0:13:08	i showed the performance results of already with and text in all but for six patients
0:13:13	from a I T B I it's already made database
0:13:18	sis this are with them requires a large number of support vectors
0:13:22	and just scaling is a key sing here
0:13:26	but first
0:13:27	as shown in the table
0:13:29	the big a all shows poor performance
0:13:32	right to plot the kernel it's very close performance
0:13:36	to the or of colour
0:13:41	because all of that competition with structuring oh sorry um because of at the large number of support vectors in
0:13:47	this application
0:13:50	the energy just scaling the energy saving
0:13:53	i thanks to the plot can of restructuring are of substantial
0:13:58	there is first a moderate rate energy savings of approximately two point three X
0:14:05	simply going from one all of kernel
0:14:07	to the simpler for the carnal
0:14:10	but then
0:14:11	computational restructuring
0:14:14	gives additional
0:14:15	eleven a hundred X energy savings
0:14:18	by all were coming
0:14:19	this support vectors skimming
0:14:25	use are the performance reach of easy base
0:14:28	C just action
0:14:32	since this is patient specific out with them
0:14:36	we construct a a classifier model
0:14:38	and present
0:14:39	results for each of twenty two patients
0:14:43	i pose i that these numbers or impossible to read
0:14:46	but i would read to you to uh the overall result in the bottom
0:14:51	and the of in is also shown "'em" break is for compare
0:14:57	for individual patients
0:15:00	oh we have found that the or with colours were required
0:15:03	for few cases
0:15:06	but for the majority
0:15:08	the paul the connors or is effective
0:15:10	and for some cases
0:15:12	even the a con as work fine
0:15:16	well but try to patients
0:15:18	the every the performance of using the pour the all for the most cases
0:15:23	is close to the performance of the or we have colour
0:15:26	as shown in the bottom people
0:15:32	in the new energy trade-offs space
0:15:34	introduced by computation restructuring
0:15:38	best feature vector dimensionality is as bandages
0:15:41	to maximise energy saving
0:15:45	but see section has a a number of features
0:15:50	in addition to a generic techniques that have been reported
0:15:54	it has also shown for see that action
0:15:57	that the feature vector dimensionality
0:15:59	can be be used
0:16:00	by channel selection
0:16:03	to so that channels
0:16:05	we incrementally add channels one-by-one one
0:16:08	until we get you close performance to the full channel
0:16:14	as shown in the figure with only two easy channels
0:16:18	the performance is close to the for eight channel
0:16:23	and this is the number of of features
0:16:25	to forty eight
0:16:29	we applied a similar techniques for other patients
0:16:31	and the results are shown here
0:16:40	after to this channel selection
0:16:41	competition restructuring can be exploited
0:16:44	for you can can further their energy saving
0:16:49	as an example
0:16:50	i i i have shown energy uh and you say being impatient
0:16:53	number seven
0:16:56	um
0:16:57	of going from an out of can the plot the kernel
0:17:00	save energy by
0:17:02	eleven one point
0:17:03	to and
0:17:06	you to you quadratic energy scaling
0:17:08	of computation structuring
0:17:10	with a with tech to the feature vector dimensionality
0:17:14	it's not the each as
0:17:16	top light competition restructuring direct
0:17:20	but
0:17:21	after
0:17:23	after the channel selection
0:17:25	computational restructuring
0:17:26	save energy by
0:17:28	we point
0:17:29	to at
0:17:33	and uh what you if the total of thirty six
0:17:35	X energy saving
0:17:37	by combining the use of a the all and computation we structure
0:17:43	the energy saving for other patients are also shown in the table
0:17:52	if you're a summary and conclusions
0:17:54	in is yeah and for by couple complications
0:17:58	the polynomial corners are on the you to light
0:18:01	even though they all for some flexibility in the decision boundary
0:18:06	that a to use of the poly they cannot is that
0:18:09	it gives an opportunity for computation with structure
0:18:15	i'm petition restructuring trays of energy in the space defined by the number of support vectors
0:18:21	and the feature vector dimensionality
0:18:25	this energy trade-off is favourable probable
0:18:28	when the feature vector dimensionality east
0:18:30	well
0:18:32	which is the case in by meant that vacations
0:18:35	and this leads to
0:18:36	C than it can energy savings
0:18:39	thank you
0:18:44	okay i think you mister T any questions
0:18:51	okay means to T you have i have one question
0:18:54	and uh
0:18:55	is is the current no you use is signal not dependent no single dependent
0:19:01	i mean
0:19:01	and
0:19:03	oh can is kernel no applied to
0:19:06	a the above such as based detection or speech recognition
0:19:11	uh other indications side yeah i i i are not out a and recognition applications an speech recognition on the
0:19:18	at the
0:19:19	and of
0:19:20	people and other applications yes yeah yeah
0:19:22	while
0:19:23	a we didn't export that area yeah but i believe
0:19:26	we can explore we can we can apply the same techniques for other um application since this it's of technique
0:19:33	is um
0:19:34	for each an every
0:19:35	propose classifier
0:19:37	um
0:19:38	but
0:19:39	i'm not quite sure about the energy savings that we get from the are with them
0:19:44	or because the energy savings of key um
0:19:47	meant to you know or am from uh you know a um are was a foundation
0:19:51	uh because um we have nice for the uh
0:19:54	okay okay of course we can we can like this that same taken other
0:19:57	application of okay be nice is the energy comes trend
0:20:01	a current size is a yes
0:20:04	of course
0:20:08	so uh this if he's using as to put a like the motion
0:20:12	which is a quite a generic the tool for
0:20:14	in in uh
0:20:16	pattern recognition clinicians see
0:20:18	lead to the patent to can be presented by vectors
0:20:22	the the performance
0:20:24	and and it is a different story
0:20:26	but in terms as support of that the missions the yet
0:20:29	to to parameters when sick
0:20:32	one is to the pin effect of the
0:20:34	yes a a a addition
0:20:36	martin in the beat that you to do we
0:20:38	a fine to for different applications
0:20:41	so that is that
0:20:42	that's typical in
0:20:43	okay yeah okay
0:20:44	thank you
0:20:46	why
0:20:46	in our questions
0:20:48	pretty
0:20:52	can you also to use the number of where like to of that you use in the at least mean
0:20:56	a a a a a cases um and i get a good performance and that at the savings
0:21:01	so with the you in the number of
0:21:04	i like terms of to using the uh at mean the kind of the uh complicated a number of electron
0:21:09	um
0:21:10	yep or what when in in a easy example what we did was actually exact like that
0:21:15	we have a eighteen for um we have a eighteen channel
0:21:19	for easy um
0:21:21	so my a petition
0:21:22	but use only need to three or four
0:21:25	oh you G channels easy rose
0:21:27	a a four our um implementation
0:21:30	and um for easy G
0:21:33	where
0:21:33	for all the station are with some of the use one um easy you be so um
0:21:39	the think the C
0:21:40	uh
0:21:41	takes a maximum that you can
0:21:43	we do
0:21:44	i think
0:21:52	sure am wondering does the importance of this kind of job because you know
0:21:57	um you are doing for reducing computations since saving yeah the right so
0:22:03	but this kind of a especially especially you focus on cedar detections
0:22:07	um
0:22:08	this kind of job could be though you know points these right still
0:22:12	uh a to the high this time is important for the
0:22:17	or mine and the same thing
0:22:19	implementation
0:22:21	um is presently for online because
0:22:24	by to in or mine we can um
0:22:26	we can enable the close loop
0:22:29	operation of these devices
0:22:31	as an example
0:22:32	i have this slide
0:22:37	here
0:22:38	if we're a like um
0:22:42	for example in this application
0:22:44	we we have some caesars
0:22:46	detected online
0:22:48	using our devices
0:22:49	and this this device work
0:22:51	uh can actually it
0:22:52	some like
0:22:54	um simulators or
0:22:56	or it can actually rate
0:22:57	some drug delivery system
0:23:00	on mine so
0:23:01	okay well i i i i one more social questions used any possibility to redo number of support vectors
0:23:07	people put this kind of cell
0:23:09	that that's
0:23:10	so you "'cause" it is it is it possible to to what isn't it despite so a we that how
0:23:15	what possibility to reduce
0:23:17	hmmm the you here we in we to as the number of sub to us
0:23:20	yup so uh we can we do
0:23:22	a was we be
0:23:24	to read as the out of you were probably of course
0:23:26	we can do this that we can we we do see um
0:23:30	the energy of this um classifier
0:23:33	but
0:23:34	it
0:23:35	yeah actually depends on the signal corps of the correlations in the signal
0:23:40	so
0:23:41	well i think
0:23:42	there's not many things that we can do for the number of support vectors i guess
0:23:46	um
0:23:47	yeah we we we explore and that the possibility of pop push the to use a circle
0:23:52	uh
0:23:52	and
0:23:53	the
0:23:54	special top to on as in uses such D
0:23:57	and it you know send is you can
0:23:59	but at the expense of performance
0:24:02	so if you lose two percent cent you don't want to it
0:24:05	three pins
0:24:06	a to the seas
0:24:07	so
0:24:09	to to use one thousand times
0:24:12	in that uh and the keep out that in each serving
0:24:16	i don't think you can quite do to maybe due can to put the vectors
0:24:19	without the need to company
0:24:21	the expense
0:24:22	uh you need to a T could be can be performance
0:24:25	so
0:24:26	so this is it may be that kid trade off
0:24:32	and yeah questions
0:24:35	okay
0:24:36	thank you very much for your panties participation
0:24:38	i Q

IMPROVING KERNEL-ENERGY TRADE-OFFS FOR MACHINE LEARNING IN IMPLANTABLE AND WEARABLE BIOMEDICAL APPLICATIONS

DSP Algorithm and Architecture Optimization for Hardware Implementation

Přednášející: Kyong Ho Lee, Autoři: Kyong Ho Lee, Sun-Yuan Kung, Naveen Verma, Princeton University, United States