Introduction in Deep Learning

Video in TIB AV-Portal: Introduction in Deep Learning

Formal Metadata

Title
Introduction in Deep Learning
Author
License
No Open Access License:
German copyright law applies. This film may be used for your own use but it may not be distributed via the internet or passed on to external parties.
Identifiers
Publisher
Release Date
2019
Language
English
Production Year
2019
Production Place
Dubrovnik, Croatia
Complex (psychology) Pixel Execution unit Workstation <Musikinstrument> Medical imaging Estimator Roundness (object) Different (Kate Ryan album) Core dump Kerr-Lösung Social class Intelligent Network 12 (number) Software developer Binary code Gradient Electronic mailing list Sampling (statistics) Sound effect Wave Phase transition Reading (process) Arc (geometry) Spacetime Point (geometry) Twin prime Momentum Ripping Online help Diallyl disulfide Raw image format Computer Number Goodness of fit Propagator Term (mathematics) Authorization Lie group Data structure Computer-assisted translation Computer architecture MIDI Host Identity Protocol Information Key (cryptography) Weight Schmelze <Betrieb> Physical law Line (geometry) System call Software Personal digital assistant Network topology Information retrieval Greatest element Interior (topology) Differential (mechanical device) Multiplication sign Decision theory Set (mathematics) Insertion loss Function (mathematics) Mereology Optical disc drive Mathematics Bit rate Matrix (mathematics) Electronic visual display Endliche Modelltheorie Extension (kinesiology) Area Email Discrete element method Rational number File format Closed set Fitness function Determinism Knot Oscillation Connected space Type theory Linearization Right angle PRINCE2 Freeware Metric system Resultant Thomas Bayes Row (database) Metre Filter <Stochastik> Dataflow Functional (mathematics) Mobile app Link (knot theory) Real number Image resolution Information systems Ultraviolet photoelectron spectroscopy Repetition Menu (computing) Airy function Wave packet Causality Well-formed formula Operator (mathematics) Software testing Proxy server Mathematical optimization Task (computing) Forcing (mathematics) Horizon Planning Icosahedron Denial-of-service attack Object (grammar) Routing
and good yeah aiw a gives d. introduction in deep learning a ev herded many times to pour i'm not sure or have everybody knows the phone daish an the off it i've of quickly goal oval then my don't pont to gold to deep into the mehta met takes there because i think it's ev airy boring actually i want to point ault some opt tick ops tickles which you can phase during the training and how you can.
it's often an souls egging already show ailed you this graphic ear and as he also mentioned in two thousand twelve a to calm the lucian new annette rukh spur intrade used and dissed was this in the chun at challenge bay you owe rig given in image and you should distinguish between a one thousand op trick class s. so they arcs aus us like cats talks tay ins everything you can.
and of think abboud and you should distinguish between them and of be forder the uh oh mean the engine year dh or hen craft to teatro as how we are was that that the whole the called and the ed if it did you pp learning hits and synth then there were many improvement to yak even low awed this just put ought o'day a or eight and improve to perform unt it of with this reason.
and into those and fifteen v e t of or even able to ought perform human kep abilities on the status that fs and.
so why is the pl owning so popular are compared to to derision the computer was an approach as i also rep with him before you have to think a bald which peach ers you need and you have to many lee yi designed and and this is very it time consuming of chorus and sometimes not the pick uppal tree or earth's an aerials and there are some mentioned here which were are very popular are like sift.
andel ca binary www header gns and what deep learning actually it dusts now is have completely it takes over do a off he tray kes trick shull pot so you don't take care of a jiu just put data end and you get they to out and to pee tray kes tech sz as theyve as completely dumb by the tea per earning approach so and your lily heirs of peep learning a feer talking of auld kundu lucian yoon are in it rir accts so i have a bake on than computer risch an.
and that's why am always talking hauled the images and the law a lever the features are some bay seek a a etch peach earth was rhyl have already used for example of four as as sobel thill dall something like that when you would goal for the had croft it keach ers but it autumn a to give though on ce a those mid that the features and hide of the features us well and if you've aunt to do dem menu lee you will it.
yet yuval not end up of it there he quit model think sold and it's completely takes the over dis part saul the question is also ll why is that so importing to have also ther qc with now bickle zz actually it they pte spec to the yeah fifty in ninety fifty's of where the you i'd ye of was already intrade use and the a model upped per sep for ofs.
used but the main reason this that there was not nef data a way lable to train to his knit rick spe cause menu of are millions of perry meters unit millions of training data uf rob a plea and also ll it's on now this of yeah cheap you was our have a will and quite affordable bay you can train i'm on because d. are high read charolais is a bill and just very important to train them and.
the eye you see here in the type line as ever a a actually his son n you're net rigs already in ninety ninety five for the to trickle geen or shun but and it basically stopped and ick your diggin into tolls and twell ff.
r.k. soul potential f.c. cations sold the mork less eke part and here in the uppal left you have to have basic him ish that's if you cation they ate they ca on mich net also though look at he sation may you also try tool find a boning vox for these objects i a thor you can also ll callan tea in since as and also's and then teak war thank mentation or infant sake.
in taishan which was very un corton for example an oughton aum striving and also at plight there then dh awful in some other airy us ill you can pined kep shun for image as autumn eh tickle ie or even yet this are thought goal which a yeah a calm peta the against some champions and in goal when you was evel to a ought perform them.
and also ll of yeah at in there and prince taishan ful thell the plan sation of takes i use with quite often saw he don't know the eva al use rir cheikh it allitt the soul it's very it popular are everywhere on thou oh.
and yeah at this come to do a a ix the hell they weyrich and they aim to imitate the human on not on the human the prey nw nd and the smallest a other mm unt as the per sup tron all za called neurone end the in core this usually ist i'm a leyshon that could be a visual or the us tin leyshon or even yell could off another yn you're on all.
was something out and they are connect that by ce nets as which are called weights and at the end this new on makes a binary decision it i dough fire spall though on fires and of this is how the modeled it than x. tedia was an extension hears alls will biased care mm and this the term unt this of the he cw asian pour that to basically have all the in puts and a moggi plied.
an by the weight summed an all up and then you use a ek you ation function which should in on than yar we will come to debt and you get an out put its somewhere the someone has an idea what the problem woodley eve with this ect you ation function it so if you have a step function lay ship to cause you want to make binary decisions one euro.
r.k. saul actually your kind use it so it you want to use heels in on than your ff hunton but you kind use of because you would use beck propagation for training and baker were gauging uses dig radian zz and if you use us dept function ike that the want be able to dyffryn she of the tool it make to their id a the route to thell it so.
well of do re such us came up with some other function that base you key it those simulates the same stuff fray from of the thi cme want or the hype uppal he kanga end for which is very he use flows directly platten your you needs a nz a ce soul easy to calculate and why do we neat that soul if you just use linney our function and you have of yeah a.
just like that to the sag a short use i make some holt off that the net rivm he aunt the only lure on a linear ck us if cation no metre whole many lay us you will at of bilal not p able to solve stalls in in that non than your problems for you sought bay thi qt the the to something like that and yeah you can trite korea's served it think you have pts ix as to the slights this actually in ice.
them way you can trial different curl knowles and and you use linear ect of ation functions bill not ver a qc for the bait up points like that.
we saul at if you use a xiu are able tool a proxy made it this non than jaret he is and can mull the very complex factions.
seoul what isn't you a net rick and right now we have ges us to a king about one your own new net drug as nothing more than a kumpel cision off more to than your owns.
and what you the what you know is the in pl pts ent d. out colds and every thing in good tween are called hidden their zz for reason because it they i put often kunda that that's the plec wealth base he you don't the what hep unst their its meche ache and this is also sometimes of problem bike elver rica's as are trying to solved that now but that's thumb.
mean i quit talked another hour of ought thill hip on go into t. to dh their ity you don't know what happened ste there.
andrs deep new a network's there's nothing what an ending nj more hidden layers to it and to usually yeah a the the the vin ers now rza live first a and that wreck alix nut the had i think eight hidden units and ands or even more andrs yet another yard at one hundred fifty layers already or hit new units andrs yeah it's.
wevill have a of flight put that a look thought to p.d. to biechele at the is a deep new an at ricks contain per you gns of weights.
it in seoul it's gone the very ste maalik thump of first nw if you have ft this yeah kind of problem let past of class and you have your ex teachers would just a numb bows of lectures you attend and x. to the ehlers bent on that crowe take and you have this training data us oh he have to train annette wreck you have to lee with training data end now you want to predict the.
of coca lls point dare you.
so what you do isse you court give bint in sold in numbers and then the model the gives you in some estimation and you also have to real auld port and if you want to train it as a cuil s. if haye xii's guess if you cation a proto put the usually do is few k.k. late the binary clauson sz appeals for ache slob ll which is kate lead it cake allay tidd and that i qt indus formula and then you try.
right to opti my ste as problem and hickam up of the best weights as off this or problem.
you could even the who would have ful muleta dh to the the difference lee ear you can also say ok i want to find odd which great evel get and then you have a rick resch an problem than then you would use another a lost functions a basically the law stunk shun defines look you want to do with today tale you hat from.
and had sole.
how all odd a tray nj that this is extra the of very difficult i don't wave aunt go into the tales there but nd u z pod as the de aim oth that you find to net rick weights ba ch are most so ubel to sought yuri a of problem and ick some tear is law spawn shun o or the show did them folk that's icky cation unt aggression and then you can i'll up to my the fray khouzam pp.
of of the of grady inti cent neath it and what it dusts it me it first picks a random point in this so this these are the weights and this is the corps pa nding lots so you want to find the globe an many muddy air so you randomly paik dubber you one and got the u.c. row a and then you cake ulay pts the gradient thes.
seoul there's as the gradient now and now you have find elop the how it change n w wand with the ff ect the loss and also in w.c. role in this the sed a called pick couple gauging to piggly of may ins of ways and you want to find auld which weights have the highest him picked on your loss.
and then you would change to last the the east weight the cording the sold his display pts our app dated and dare huss learning weight in it was just actually the very important for training the with net ricks.
nw and yet then you just through pete the step and thay you gets to some kind off many nguema if it's to cuil obel many nguema or low commenee much look the of this mordant.
seoul what dust this than he weight and how can he set it so if you'll help that any meat it yet it is you just said it's a small you training pts on let thence i do training the ailed paik the rare he long time and the if you are in low can many nguema and can be that it's stuck there and it clan got out soul it even vz i'm kind could even cause on the fitting in this case.
base if you pick it right it's very ff ah stare and who probably leave the find the chlo glum ect the koba many mirth well and if you pick it too high because of want to have the solution very fost i with from the advise against that the call it the the jus a that leave a come on stable during training and the lists oles and also there's a very nice of the ink.
the air of which you can try frail self how this of in rate a sects though a on optimization pl them thes.
seoul rere of urate the hour of all ce just showing you to weights and rere rut re so just were trying to find ault the lense cape of a lost function with different weight and it could look that ike like that for ache celt that and its you can see its veh every des fickle to find of a cuil obel i'm in he might the eye and also this prady in dissent in this method there's a reedy a approach so it's not guerre in.
he'd it you find as clove an inning.
saul dare are some solutions i just call them now on for xom but depp the floor ning of d r intrade fuse momentum but they sick the a dozen ce just a using the a grady in from deter ation before and keeping this information to get a yeah some flow in tudor low up to my say shull poppen a not cake a dating for each step by itself and also.
all an eight i'm this a very poppy low a pro trait no and watt also zeid also slickly done is that you don't just use mourn point in your one feigning point in cake a later lost what it you should use mohd of the wants because it get in better s.t. made a off way you are actually and its heidi her lies of the slits much fost on much more like ol and the had a.
you glue as used the high is that's sys ooh can thea a a fort the ff your ought bay you it you het.
yan the are also ll many more but just that dales aum eve ease a most important one.
andrs yes and three on his would limit at on time i put the link there you can even tried this is some kind off x. or problem you could player round can't at some hidden lay earth a nuke graphical int of face you can at them muir un's and find auld what happens you can even shange the ick cuve ation panchen also to linney on yuval see of with moat rick even if you at eight hidden mean layers that.
the north.
ok solor under on a law fitting of think this should be quite thier to every won soul if you have thes training data and to train pl too long time where you mall those bay torre complex or the problem it could ease of the lawn every f'ing from the dane eke training data eye and if you ups i why if you want to applied on testing date i'd quick note where.
they're close yoav earth it oil your mada this to out for he summed if you'd use lynn the arc that's ify of the slow evo zoo esta made to a couric this avic ation bone the rees and yuval under fit sold this is also ll the have every buik problem of on the earned over fitting and the f. war nw you have to doom improve to generis ation ih dzus is usually done by.
i some regular he sation zz and their auld to solutions a king if you hear but their many more frache sump of prob welt what this space he'd us as a just sets some muir owns and their respective wh weights the siro and this is just the up the them bui course you want to ek sz lee that in it rug does knots was i on ste the fifth take pasta get this pacific auld coats so.
oh you sets some waun see row and to train and then in the next ration you said others on see row and the sed for generalization and also and desist maybe the most important pots or if you face deep learning most important part as to have of addi daish an set so you sher teh free sets training that he daish an and testing and your ing training you shook up surf holl it p hey a ho you in that rupe.
hates on this buddy daish un data so n't you if you train of course the last on the training date uf the goal low although and lower abut what could actually happened is on the ville he daish un's that bill increase again or id stops and when it's dopp here and the the if the crease us year at means that its long ied a training data not the testing data anymore so you dis stopping us very importance a.
he and his as of a's should stop to training than a soul what we a thi the on ce a fought have has thought tra nz us a a some tulle mulls of force hoople west learning that activation factions a very important to ing nw that intrade use non the neri t. that you're in net ricks a nothing more the mnd comes as the sz un's often your un's and be have learned some techniques.
it's for are regular he station and prop to my station.
zsolt know i will cyst talking a ball it x. one and ix to but usually your data this way more dench ist two features so if he have an image liked et what could be due in el holl do we train it forrey midges so the first allusion and was also ll done by dinette which were with in nineteen ninety six they ge us used all the.
pixels and decked them he sold but maybe you fee where the problems are because actually you have stole many wait it's the year a mohd of new rinse here times dem aunt of you in see a so intrade use many weights there and also you are completely in a collecting does patients and formation.
seoul the drewe objects odd that you'll have no own look l. it t. pers their way shun to basic he don't use any spatial connections and you have many weights so how can be yeah www maintain this beiges structure and air come the the at the are dick condi lucian the ll layers of which were intrade used or a plight and into toll xna n't well.
nw and basic see what they do with peed don't just lawn one weight they lit on a matrix of wh weights and this matrix of weights does or shea out all over the image because actually if you have a etch in the a pall left he want tool des take the same edge and the bottom right frache sump as olds not just for one region it's applied on the hold image likely can t. hears all.
but this far vz ole where every pot and what is actually dust was through bill get some kind of ect key whe ation att for the specific keach or so let's say a year or in this x. nom of this is a soul of think the third or eh f'ing wh you will detect all the edges with this kind of peach or and tend actually you don't nobile knee wants that the you lawn marta bill oft i'm and what it dust for each of foot.
have you put each have to furred or you'll though on you get one ek t. whe ation meh pp a nd the output.
so if you want to or a nd sixty four fert those servile get that depth osce tick sti for their.
it in so on now we come to actually do winner of to image nets would culpa gnc wealth with just addict net and id looked liked eth soul it first the af you're in put him it's which is in this case dh plundered tent four times twin to than fenty for pics of and its ob g b's wet hess pts a free channels for den you have plight and a evan times as leaven kerr on them so the sussed a sys off your.
on will who shun the footer and you lawn the weights would at and day a even up lightest trite a four and what the thigh us you can thieu tier basically is skips one him put part and the bill rid use d. there's aleutian off the old put what that lee's saudi a playas trite or force that that's why the naik she's ert has on the a they con you viewed i think if.
he five hence the fee fife thing an the resolution and it has a death of.
they ninety six oh did it the at huss lunt ninety six different for thurs and added goes on like that but they don't the ply ce try ting any mm on the use nicked waling in book makes putting dusts the tes of yet decreases to spatial resum utian by effect or of to you call the just looked at for ache sump of do was four picks the use the mex and they you unst aust at the way.
yet and then day did it fall of yup and to pour six come lucian a layers and then dh a basic the you can fated this is de fee tricks plec shun part because dis is we out it's still some kind of spatial information he cause us a metrics and then this tech to dick ends of eeg are coming back to the densely are sfi its team before the.
u f just phone decked or and you trained excessive quite of this is actually also rare he him corton because you can train annette rip quake sample four pp tech that's if you cations and then you can use doze p chills which you have flaw nd also fought of us tough for retrieval ta spin just ix trek tippi chizen make rich we ll or clustering ove whatever you will in frain you all of that's ify author don't need to rely on those thaw if you're.
make sounded just want to have acct that's he five four through hundred up chick classes the as no need to train it all again you can just use to fee ch of and try to solve it like that.
it.
yeah andrs now we are i was beginning with dat in emit own awful anning the that you can see here de results again and there is some kind of development in the numbers offs layer ste the intrade use a at first ever aid layers and thick sti won the in para meters then seiler and for www as came d. at more filters and then two dozen potting never already.
nineteen layers and to told courtine there will twenty two layers and then and to those of fifteen that's where day and exceeded to human performance theyve or even bone hundred in fifty two layers so days this nice meme and lit richest dec more layer arse the few want to improve to results an but i will hood also strongly at vice against that your should always try to fine to best.
but model or the best a architecture for your problem and also if you use very complex model its you i guen maybe intrade use over fitting into unit rick so always think about what's the best i keita take triggers a i get yeah it.
you pay the what it a yes yoest tayde used to same said of in their ges and that in of humans and with tate to set and their well i i don't i don't have tea tits all many a humans were put its a pay ting but to do riz eigg where the ick five point two percent ever or and i also had to enter tate at some point myself and a.
was even marty claw sore you at one image and you have a find old all elim andrs which were in their andrs you all are sees all menu things as a human and net rex debase e t say tse some things and the becker on vey age don't pay attention for the as a human soul and garre the ll it direct a he so yeah but.
c he aaj a but i thing this is of predict ok they av wa of one thousand i don't i and i actually knots with the used every op trick class for the menu and a patient thirty bicol of.
keeping in mind vong tells an clough's as to n o taters of costa fickle old but on any wace they are also other airy us mia other a humans were auld perform the lot a even had ness that's of paper owned hugh look ation his team ation and humans a hat the stent the which a nz to a phase ragle can isn't as also very good for ford keep unit rugs so them many airy.
the us me i human so but this is a webby all it all see a star that think of you start of the dead as some on better who with knowledge in many airy us war in vaughan kotick you law that's where the un your net rex lech yadier a very good at one task but fao nz all of other ste an m soul but sometimes the des shun of information can be very useful to.
all for taf but are.
it.
you was on aunt me on he yet you you hugh yeah bei by it a add heene it yet it ease the these these whine but all owe a in the a you a yet as a good and they and harp a you cheech monies this more the now aunt.
ns off what though own an yell ff course owais a yeah an the o. about have it a will there but nor op of whom he or za there.
it.
i.
and a yes saw it you want to reap them thing at put some reference of a yes.
Loading...
Feedback

Timings

  178 ms - page object

Version

AV-Portal 3.19.2 (70adb5fbc8bbcafb435210ef7d62ffee973cf172)
hidden