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