Merken

# An Introduction to Deep Learning

#### Automatisierte Medienanalyse

## Diese automatischen Videoanalysen setzt das TIB|AV-Portal ein:

**Szenenerkennung**—

**Shot Boundary Detection**segmentiert das Video anhand von Bildmerkmalen. Ein daraus erzeugtes visuelles Inhaltsverzeichnis gibt einen schnellen Überblick über den Inhalt des Videos und bietet einen zielgenauen Zugriff.

**Texterkennung**–

**Intelligent Character Recognition**erfasst, indexiert und macht geschriebene Sprache (zum Beispiel Text auf Folien) durchsuchbar.

**Spracherkennung**–

**Speech to Text**notiert die gesprochene Sprache im Video in Form eines Transkripts, das durchsuchbar ist.

**Bilderkennung**–

**Visual Concept Detection**indexiert das Bewegtbild mit fachspezifischen und fächerübergreifenden visuellen Konzepten (zum Beispiel Landschaft, Fassadendetail, technische Zeichnung, Computeranimation oder Vorlesung).

**Verschlagwortung**–

**Named Entity Recognition**beschreibt die einzelnen Videosegmente mit semantisch verknüpften Sachbegriffen. Synonyme oder Unterbegriffe von eingegebenen Suchbegriffen können dadurch automatisch mitgesucht werden, was die Treffermenge erweitert.

Erkannte Entitäten

Sprachtranskript

00:00

I could there's an introduction to deep learning so bispectrum jet it commodity than succumbing

00:21

Canada let's start by thanking Interreg Rashid the going is excellent gentle introduction to no networks on and build upon that and hopefully to hopefully here show you how to um develop some of the networks have been used to get to the really good computer vision results that we've seen recently so of focuses mainly gonna be on

00:41

image processing this morning and this talk is come although principles and mass behind it in the code and the reason is is quite a big topic is colored go through and NGOs screws and an outside the littlest code that had a this useful so as a quick overview organ goes through we are going discusses the honor library which is the 1 I personally use all of also love that tends to flow and we're gonna cover the basic model of what

01:11

is in our network just building on reeks told they're are going to have a car go through convolutional networks and these similar networks that get in a really really good results that we've seen recently the model briefly lasagna which is another Python library built

01:28

into the honor to make it easier to build new on x will discuss why is there and what it does and then I'll give you a few hints about how to actually build on your network you now have she structure what led us to choose um just to say you have a rough idea how to train them the of just a few contends and tips to practically get going and finally uh time permitting i'll go through through the Oxford net feature gene

01:53

network which is how to use a pre-trained that was the you can download on the Creative Commons from node to the University of use that yourself because all go through why it's useful sometimes users and there whether somebody else's trained for you and tweak if own purposes now the

02:12

nice thing is there's some 12 materials this is the based

02:15

of the tutorial I get up I data London in May and if you check out the gate there right fury deep learning uh tutorial data 2016 you'll find that there's they get high rep either all my works of you belong get help so you should go to see you everything there in your browser I would ask the the please please please do not try and run

02:38

this code during the talk and the reason is is because when you run the stuff that uses the VGA Oxford models that will need to download a 500 makeweights file and you will kill Wi-Fi fuel start doing so please it on your own time such big X and also uh yeah

02:58

this if you want to get more in depth about the Amazon here up up some slides we

03:03

check that 1 speaker dep profile will be the to the B. This talk slides and will also be intuitively on a lasagna as well so that will give you a break down of Python code using the animals idea what it does and how to use it um and furthermore if you don't have a machine available you don't wanna set of itself there are set up and uh MI for you so if you wanna go use 1 of their

03:28

GP use you can go and grab a hold of that and the Ronald code there everything's all set up and I hope it's relatively easy to get into alright now to get into the

03:39

meat of the talk and what better place to start than image imagine that there is an academic image classification datasets we got round million images I think it might be even more now the divide into a thousand different classes so you got various different types of Belarus different poses of at flowers buckets whatever else swallowing came up with rocks males and the

04:04

way the ground truths as in you got a bunch of images that means greater flicker and you got of what a ground truth for each image is and while was

04:12

prepared as they went and got some people to do over terms in Mechanical Turk now the top 5

04:18

challenge we've got do is you go to produce a classifier that when a given an image will produce a a probability score

04:26

what it thinks it is and you score heads if the ground truth class therefore true class is someone within your within your molecule or um in the know network or whatever is used its top 5 choices for what it thinks the images and in

04:41

2012 the best approaches of the time used a lot of hand-crafted features for those if you move computer vision is quite SIFT hogs Fisher vectors on my

04:51

stick into classifier maybe even a classifier on the top-5 error rate was around 25 % and then the game changed because of the

05:02

system and hence and in that paper envision that classification with deep convolutional neural networks that a mouthful they manage to get error down to 15 % and the last 3 years more modern network architectures the goddamn further now a dance about 5 to 7 % and I think people a Microsoft even got down to 3 or 4 I hope that this talk is going to give an idea of how it's done

05:27

OK let's have a quick clicker run over the unknown no networks offer

05:33

comes in 2 flavors was kind on a spectrum really you got the kind your never toolkits the quot high level 1 and the other and you go expression compilers and and you nematode please specify the new network in terms of layers with expression compilers this somewhat lower level and can describe the mathematical expressions that tree recovered that affect behind the layers that effectively describe the network and some more powerful and flexible approach the annotation expression compiler you can run NumPy style expressions I going to compile it to be the same

06:12

to on your CPU all CUDA 2009 NVIDIA GPU if you have 1 of those available and I once again

06:20

if you wanna go to have again and again to that there must slides so that I mentioned earlier there's a lot more there's a lot more to the other so go check out the deep learning . net website to learn more about it and find out about that like is a full description the API and everything will

06:35

do uh some of which you may want to use there are of course others there's

06:40

tends flow developed by Google Maps gaining popularity really fast these days of them I will be the future will see OK what is a

06:48

neural network well we're gonna cover a fair bit about 3 covered in the previous talk but it's got multiple layers and the data propagates through each layer and is transformed as it goes through so we must all travel the image of a bunch of bananas is going to go through the 1st hidden layer and get transformed into a different representation only a transformed against the next hidden layer and finally we end up with a similar we're doing image classifier we end up with a probability vector effectively all values in that middle

07:21

son to 1 and all predicted so uh out pretty into class is the corresponding row in the probability that and the probability that Robert has the

07:32

highest probability OK and this is 1 that a kind of looks like we see there are weights that you saw in the previous talk uh that connect all of potential use between the layers and see data being put in on the input propagating through an arriving in the output breaking down a single layer of a neural network we've got our input which is basically it is a vector of an array of numbers multiplied by a weight matrix which is the crazy lines and then we have a bias term which is simply an offset use out of vectors and then we have our

08:08

activation function or nonlinearity those terms are roughly interchangeable and that's that's the upper layer

08:15

activations will then goes into the next slide all the output of this the last letter in the network mathematically speaking x sample vector what is the output we represent all weights by about by their weights matrix that's 1 of the parameters of our network all parameters the bias got a molarity function and normally these days that's really rectified linear units it's about as simple as they come it simply maps of x is 0 that is the most that's the part that's the activation function has become the most popular recently In a Nutshell y equals f w x + b repeated for each layer as a goes through and that's basically in your network just that same formula repeated over and over once for each letter and so would make a misclassified would take the pixels from our image when sperm onto a vector stretch them out row-by-row bonds the network and get a result so in summary no no work is built from layers each of which is and matrix multiplication then and bias the plan nonlinearity OK and how to train a neural network we've got to learn values for all parameters the weights and the biases for every letter and for

09:48

that we use backpropagation we're gonna initialize always

09:52

randomly that with more on this later we're going initialize the biases alters 0 and then for each example our training set you want to value has to make said we got to evaluate on on average prediction seawater reckons the output is it compared to the other the actual training the actual training out what it should produce given the input we got to measure of cost function which

10:17

is roughly speaking the error that's the difference between water network is predicting what should

10:24

predict the ground-truth output no the cost function is kind of important so we'll just go and discuss how little bit the classification where the idea is given an input and a bunch of categories which category best describes this input of final we use a function called softmax as or nonlinearity are nonlinear linearity or activation function and outputs a vector of

10:53

class probabilities the best way of thinking about it is that say I've got a bunch of numbers must some mall up and I divide each element by the sun value roughly the proportion or probability assuming all of our numbers to start with a positive but they

11:09

can also go negative your network so what we do with the softmax is at 1 little wrinkle the full summing up of what we do is we take our input numbers we compute the exponent of the mall namely someone not we divide the exponent by and the

11:22

exponents that's softmax and cost function function is negative log likelihood also categorical cross-entropy 0 and to do that you gotta take the log of u if you to say you have an image of a dog you

11:40

take you around the image to the network you see what the predictive probability is the dog you take the log of the probability which is going to be negative if it's break the probabilities was then a lot of that's going 0 if it's like 0 . 1 is can be quite strongly negative indicate that seem again that's log and so the idea is that if if it's supposed to output dog it should give a probability of 1 if it's giving a probability of less than that the log the negative log of the quite positive which indicates high error so that's your cost the

12:12

regression is is different but rather than classifying an input and saying which category closely matches this try quantify the measuring strength of something will strength some response typically

12:25

without a final and doesn't have an activation function is just the identity linear annual cost is going to be sum of squared differences then what we're

12:37

gotta do with unknown I would could reduce the cost reduced the error using gradient descent more we have to do as we have to compute the derivative of the gradient of the cost with respect to all parameters which is all always smaller biases within our letters the cool thing about it is the iid does the symbolic differentiation for you I can tell you right now that you don't wanna be in a situation where you have this massive expression for you on your network and you've got to go and compute the derivative of the cost to the spent some parameter by hand because you will make a mistake you will flip a minus sign somewhere and then you'll never want learn and debugging will be a goddamn nightmare because we really help to figure out where it's gone wrong yeah I didn't I would recommend getting any m a symbolic and symbolic mathematical package for you will use some of the on is handles at all and actually write that code there

13:34

because by the way this is the undergrad course weight and mother told his do this as well just to save you time and sanity

13:43

and many of their your parameters you take your weights and you subtracts the learning rates which is lambda and multiply by the gradient and and i'd Stanley learning rate should be somewhere in the region of 1 1 times 10 to the minus 4 1 times 10 to the minus 2 something in that region you also going you typically don't train 1 example that once you got tables where mini-batch about 100 samples through dataset you're going to compute the costs the

14:18

each of those samples Abergel cost together and then compute the derivative average cost with respect to all of your parameters um and the ideas you know with average in that so the idea is that that means that you get about 100 samples

14:34

persistent parallel on them and we run on a GPU that tends to speed things up a lot because it uses all of the the parallel processing power of GPU training on all the examples in your entire training set is called an epoch and of want multiple epochs of training never some not or 300 so In summary take a mini-batch training samples evaluate running through the network measure the average error error called cost across the mini-batch and use gradient descent to modify the parameters to reduce cost and

15:11

the baby above and all right

15:16

multilayer perceptron this is the simplest neuron network architecture there's nothing we haven't seen so far he is only 1 and is fully connected all dense layers in a densely each unit is connected to every single units in the previous layer and to carry on from To can pick up from 3 to talk and then missed the analyst handwritten digits dataset is a rose is a good good place to start and on their work with 2 hidden layers but the 256 units after 300 iterations gets about 1 . 8 3 % validation Sarah so a is about 98 . 1 7 % accuracy which is pretty good

16:08

however these hundred digits acquires special case all

16:14

digits nicely center when the image there or for the same position scale to about the same size and you can see that the examples that and are fully connected networks have 1 weakness has no translational invariance if margin you wanna like take emission detectors go detector bowl somewhere in the image uh um what effectively means is you know it's let me learn to pick it up pick up the ball in the position where it's been seen so far it won't learn to generalize across all positions in the image and 1 of the cool things we can do is if we take the the weights that we learn and we say take we take and which 1 of the new 1 of the units in the 1st hidden layer

16:57

and take the strengths of the weights that link to all pixels in the input layer and this that's what you end up with so you see that your 1st hidden layer weights offered to for bundle feature detectors that detect pick up the various strokes that make up the digits and so is kind of cool too busy lies above that shows you how the dense layers are translationally dependent and and so for

17:22

general imagery like say contact cats dogs various spiders and everything that makes up versatile creatures mortgage things you gotta have a train set large enough to have every single possible feature in every single location of all the images and you will have a network that's got enough units to represent all this variation um the Casey again to have a training set and the trillions and your network with billions and billions of nodes and you don't have enough again the enough about all the computers in the world and the death of the universe more training so material convolutional networks is how we address that convolution it's a fairly common operation in computer vision and signal processing because a slut a common convolutional kernel

18:11

over the image and what you do is imagine say the image pixels on 1 they don't tell your kernel which is got a bunch of a bunch of little weights sponsible values multiply the value in the kernel by the pixel underneath it the Frawley for all the for all values in the kernel you know take those those but those um parts and some all up in the slide the kernel 1 position to the same slide this do you want to do the

18:40

same and what you end up with is and outputs and

18:46

while often used for feature detection so brief dates a ball filters if we produce these bunch of the these these filters which product over the sine wave and the uh gas function in that what is the loss of usable way visible soft circular way things and if you do convolutional see that they act as a feature detector that detects certain features in the image so you can see are roughly corresponds you can see the 1 of the vertical bars there roughly pick out the vertical lines in the image of the anonymous the horizontal bars pick out the horizontal lines and you can see how convolution convolutions act as a feature detector and that was used pretty much they is quite a lot for that too bad contract to convolutional networks bank public and forget recapped that's a fully connected that looks like with all of our input connected to all of our outputs In a convolutional layer you'll notice of a note on the right is only connected to a small neighborhood of nodes on the left and the next node down as catch to small corresponding neighborhood the weights also shared so that means use the same value for all the red weights and for greens and for all the others and the values of these weights for that kernel that features of a practical computer vision worthy of these kernels manly learning like in a

20:16

convolutional network more than 1 kernel has to be used because you got extract a variety of features is not office not sufficient just to build a gentle of just the horizontal edges you want at vertical ones and all the other various orientations and size as well so you have a range of kernels so you're going have

20:36

different weight kernels and the output the ideas he got image that 1 channel only input and about 3 channels on the output or you might find in the tube you convolutional memory manager of 48 channels all 256 face examples later some architectures in the inner part of the scene for a high dimensionality in this sort of the channels up of and OK stage calculates all pixels image or text to the pixels and all channels in the previous lesson draws in the doors dates from all channels in the previous layer however the maths still the same and the reason is because a convolution can be expressed as a multiplication by weight matrix is just the way major square sparse but we know the

21:21

mass doesn't really change as far as uh conceptually and I was fortunate for us because it means that the gradients and having done so far this still works and as for how how you go about figuring that out I just recommend lengthy energy for you that wouldn't be

21:39

myself when recommended there's 1 more thing we need downsampling so typically if you don't if you've worked in further Balkan point

21:48

is an image-editing factors on shrink an image down by certain amount say about 50 %

21:54

um you watching resolution and for that we use to operation the the max pooling of striving max

22:01

pooling what you can do is you can see G. level image up there is divided into 4 color blocks say the blue block has 4 pixels what we do is we take this for pixels we picked on the maximum value we use that's Robin averaging just take the maximum and that's max pooling and the downsamples the image by by the fact that if p is thus of size of the pooling um that operates on the channel independently the other option is striding and what we do there is effectively pick a sample skip a few pick sample skip a few it's it's even simpler this is

22:40

often quite a lot faster because what you can do so the the convolution operations support Strider convolutions were rather know rather than

22:48

taking sort of them in the output inferring some way that is to effectively jump over by the pixel each time so that's faster and you get similar results so moving on yeah the

23:02

current used convolutional networks to currently used to solve the image the and this status in 1995 and this is a simplified version of its architecture so we've got is you've got to 20 kernels you got was 20 between input image 1 channel this is monochrome we got 20 kernels 5 by 5 so they would use the Mr. 24 with 24 that's not 20 channels max-pool shrink by half have 50

23:33

kernels 5 by 5 now

23:36

go 50 channel image a bait med school shrink it by half we found mentioned uh do a fully connected densely 2 2 5 6 units and finally fully connected to all 10 unit output layer from from past cruelties after

23:53

Federation of the training set we get 99 21 % accuracy 1 7 months

23:59

and there is not too bad what about learned kernels this interesting to think about what the feature detectors is picking up so if we look at say a big dataset emission at this is that this is the preserve the paper I mentioned by the beginning these are the kernels they learned by the

24:16

neural network and for comparison a considerable over there no the reason the columns of the bottom is just because of the way they did the actual thing involving 2 ways but

24:26

up but if you look at the top right you can

24:28

see how it's picked up always versatile edge detectors of various sizes and orientations that's the 1st letter uh uh either Fergus to further and they figured out a way of busy lies in the kernels how they respond to the 2nd so you can see the book cells that respond to various of sort more complex features then squares and you kind of the this curve the texture Wilson online

24:54

features of circular features and further up on about layer 3 you get somewhat more complex features

25:01

still we've got things that repres River recognize parts of simple parts of objects the case this gives you an idea of roughly how the convolutional networks fit together uh the operators of the feature detectors where each letter builds on the previous 1 picking up ever more complex features OK now

25:23

move on to lasagna if you are

25:27

specifying that what using the mathematical expressions using the and it's really powerful was quite low level if you have to write

25:34

I on your network mathematical expressions and on the expression each time it could get a painful but was only built on

25:40

top of it and makes it nicer to build networks using the

25:46

and its API rather than slide to specify mathematically expressions you can construct layers of the network and the Council and get the expressions for the output for its output loss this Python layer on top of the utterances with them so in the end but the cool thing about it is if you have 1 of these mathematical expression compiler if you are constant some crazy new loss function will do something new and crazy whatever years you like all you wanna beats of inventive you can just go right of maps and let the honor take care of figuring out how to run that using CUDA using uh nvidia's CUDA and what part worry about itself is quite easy to get going as old as you just do all in Python Annals what's great so that's that's what happened to like it um from once again uh slides available there you want to go go and dive in more detail as for how to build and train neural

26:42

networks having starts out with

26:46

a bit about the architecture if you want to know what I want to use it to get a nice new I was gonna work going to try and give you some rough ideas of all colors you wanna use and where the nose to get some this guy is going to give you good results so you're a part of the

27:02

network is going to be

27:04

stuff your input layer is going to be blocks going to consist of some number of convolutional layers to 3 for convolutional layers followed by max-pooling where that effectively down

27:16

samples alternatively you can also use starting as well and we have

27:21

another block the same and you'll notice that limitation is that's quite common never in the academic literature is you specify the number of filters the number of the number of the number of kernels and then the 3 specifies the the size they often used quite small

27:37

filters in the 3 by 3 kernels and then MPT means max pooling downsample strapped 2 and notes that after we've we've done the downsampling you double the number

27:50

of filters in a convolutional layers and then finally at the ends of

27:56

blocks of convolutional max-pooling layers you can have a fully connected than older Austin is dense layers where you'll typically and if you go the large resolution coming

28:07

out of the you want to you want to work out what was of dimensionality is points and roughly maintain that tall reduce perhaps a bit in fully-connected layer and you can uh because of 2 or 3 fully connected layers we like and finally value out of

28:24

and this notation fully connected layers of just mean

28:27

206 channels and OK case so overall as discussed previously convolutional as a of the text

28:36

features in various locations throughout the image the fully connected layers are gonna pull information together and finally produce the output

28:46

but there are also some architectures you could look at the inception that works by Google resonance Microsoft for inspiration from a guy with a somewhat smaller people been up to

28:56

I gunsight more complex topics the batch normalization it's recommender most cases it makes things better and necessary deep networks follow a larger tell you deep learning neural networks a a deep neural network is simply never roughly more than 4 layers lets us all Islamic and so

29:16

if you want to take a deep networks of more than 8 layers you 1 batch normalization of always the just 1 trained very well and the closer speed up training so that of your loss drops you cost drops faster prebook um well can take more lead from Connecticut run on you can reach their rights as well and the reason why it's good

29:37

is sometimes you think about the the magnitude of the numbers you might start out making numbers of certain magnitude your input layer that's that that market you might be increased or decreased by I 1 by the multiplying by the weights

29:52

to get to the next layer and what the people at the staff-level layers on top of each other you can find the magnitude you you'll values either the exponential increases or explanatory to shrink toward 0 and be 1 of those bad the training of completely so batch normalization it's a standardizes by dividing by the standard deviation subtracting the mean of 3 July so you want inserted into your convolution fully connected layers off the matrix multiplication and the before in bias and the before nonlinearity so the

30:24

nice thing is lasagna with a single call does that for users to do too much surgery cellphone network but dropout is

30:34

pretty much necessary for training uh you don't use a train time that you don't use it prediction and test time into random sample through the never to see what happens it reduces also

30:44

has over-fitting overfitting is a particularly horrific problem in machine learning and it's going to buy you all is gonna by you hold time machine learning at this is what you get when you

30:58

train a model on the training data is a very very good at that samples that are in your initial training set but when you are shown the example is never seen before the does dies it fails completely and so essentially what music is

31:13

particularly goes examples of picks out features of those particular training samples and fails to generalize so drop out there back this we

31:27

gotta do using a randomly choose units and layer and multiply random subset of them by 0 usually about around half of them and so you gotta keep in the magnitude of the

31:39

output the same by scaling up by a factor of 2 and then during test predict he just run is normal the dropout turned off you know apply after the fully-connected layers normally

31:51

don't bother you can do off the convolutional as well that um the fully counted as tools and strongly we applied that's I put it aside and Sonya and show you what it actually does this is with the drop and also you see all the outputs going through little diamonds represent all drop out so we take half of them we think in terms of an EC the gray

32:12

the gray white lines while that essentially means is when during training the uh the back-propagation won't affect those weights because the drop that kills and then the next time around you do you turn off a different subset of them and furthermore and the reason it

32:30

works is it causes the net causes the units to learn more about set features rounds of learning of can adapt and develop features that a bit too specific to those those units so that's that's roughly how it sort of combats overfitting dataset

32:46

augmentation because train your networks is notoriously data hungry you want to

32:53

reduce the overfitting and you reach a larger training set and you can do that by artificially modifying existing to you adjusting training set by taking a sample modifying it somehow and adding that modified version

33:06

to the training set so images it is going take

33:11

shift over by a certain amount of down by that you're gonna rotates at that against skeletal but horizontally flipped it be careful of that 1 safe example go images of people vertically fitness upside down that you that will describe the training set so

33:28

you got when when you're doing dataset augmentation think about what you need for your

33:32

dataset and what it should output and think about what whether your transformations are a good idea I can finally dataset

33:40

states standardization 1 your networks trained more effectively when you'll dataset has a mean of 0 would

33:50

polarize amuse 0 and unit variance both the deviation of 1 so and also

33:58

with regression you wanna standardize your input data and regression you wanna standardized

34:02

output I don't remember that in regression we also quantifying the server using real valued outputs we will make sure that standardizes well I've I've I've I've personally found that means that no 1 I haven't done that so uh but when you use your network when you deploy don't forget to uh do the reverse summarization together back into this into the space you back into this of scale range that you

34:27

want to be in in the 1st place and to do that salvation straddle samples and into the right and the case of images

34:34

he is going to go through all the images and extract all the pixels and splash Mountain big people the RGB channels

34:41

suffer and you're going to compute the the and compute the mean and standard deviation in red green and blue again 0 the mean by

34:51

subtracting it from divide by the standard deviation and that's standardization OK when training goes wrong as often will um you'll notice we wanna do is as you as he train you and and you might get an idea of what a lot what the value loss functions

35:13

when it goes crazy and start heading towards near 10 to the 10 and eventually goes everything's there is going to hell

35:21

so you try you lost as the training our network so you can watch for this OK if you have error rate of

35:31

growth of a random guess like this is throwing about 3 point it's not downloading anything essentially it's learning

35:40

to predict a constant value a

35:42

lot of the time um sometimes it is it is just there is enough data for it to pick up patterns and they can also learn to predict a constant value let's say you're for instance that

35:54

you have a dataset where say you've gotten divided say 10 classes let's say class the last class only has about 0 . 5 % of the examples and 1 of the best ways this sneaky hard Italy on that will figure out where to figure out is to simply say that that that this is something never predict that last class because it's only gonna be wrong in 0 . 5 % of the cases and that's actually pretty good way of getting the last stanza of the preload value by concentrating on all the other classes in getting those right and the and the problem is is a local minimum is a local minimum of the weakening of

36:30

the local minima of the cost function and your numbers get stuck in there a lot and it will be the end of your existence in but the most often don't learn what you expect to all you want to you look at it and think I has a human and a minor the result is this and

36:48

unknown elements to pick of features in detect something quite different on so yeah looking at

36:57

the bane your existence and illustrate this is a really nice cool example that is available online and talk about how I design a computer vision problem using networks with a simple problem like handwritten digits you could just your network 1 your network will doing great

37:14

wonderful it's a more complex problems the often just not

37:18

enough and your works on not a silver bullet so please don't believe all height that surround deep learning right now and its theoretical foot fairly possible to use a single neuron complex problems if you have enough training data which often in an

37:32

incredible amount so for more complex problems you break the

37:37

problem down into smaller steps and to talk a bit about Felix allows 2nd place solution to the cow competition on defined right whales so His 1st naive solution was to

37:53

train a classifier to identify individuals so long enough couple of his website OK cool

38:02

OK so effectively these

38:06

patterns on had a well as what used to identify an individual and the challenges of pick out figure out do an image of a whale figure out which individually is and this is the kind of image you get in the training set you've got the ocean surrounding a little while as he breaches as the prospectus had over the surface you gotta figure out who is from the picture so if it's a 1st solution was to to stick that through a classifier and see what happens so we skull find out OK baseline naive approach and what he found was is that it gave no better than random chance Saudi and that is a use what's called saliency detection where try he got the use detect trick to figure out which parts of the image of influencing the network's output the most and he found out that actually the bits of the ocean reflecting how to do that

39:02

OK trifles experiment owing to imagine that I give you

39:06

this problem you a bunch of images arrive wells might say that's number 1 that's number 7 that's in the 13 box evil something given really really horrendous horrible amnesia that has completely 1 2 minds of the concept well is what the ocean is just about every human concept you have to of the tree starting out with images this 0 knowledge of told no semantic knowledge about the problem you can't even guess what it is is given images and given numbers mental from this training set to figure out what these are what do you know what's what more

39:35

how gonna make that decision is that the ocean is the whale you what part of what what part of the images actually help you make a decision we think about it From this better than you know what that's where you 1 and starting out from starting

39:50

out from 0 knowledge and that's why the initial solution didn't work very well you can't if you had a billion images of all the ground truths and well below of the marine biologists have gone in your hand hand classified a billion images of them and put an enormous amount of human effort because

40:06

then the signal will eventually come through the noise that we compactly do that in real life so and his

40:18

solution is the armaments in the

40:20

region by saliency so found out they had locked onto the wrong features so you train what's called the localizer now I've told about classifying regret classifiers

40:29

and regressors localizes what they do is they look at image and they find might might target a point of interest is over there the image and so what he did is he found he got the localiza to take take the limit of the well and found out who had is there and after that you want for classifiers the idea is you 1st gets trained in it what to look

40:54

for whale they could outcropping of an image and then just work on that piece and furthermore he turned the keypoint

41:05

the only will have a learning you again you trade keypoint funded to find the front of the head and the back of the head so you could then take the image of a whale and takes so they're all the same orientation and position and after that I haven't got some really uniformly images of

41:23

Wales a given monitor the classifier and eventually that's a train the classifier on oriented and crop well had images bottom second-place Macao competition so I was

41:36

kind of this of nice dismissed that nice illustration of how so if you got

41:41

a really careful about how use these things alright so modern

41:47

time great OK there might have been

41:52

vector because of your choosing them signal sequencing such oxidative Georgina um and transfer learning using a

42:03

pre-trained network is often a good idea the Oxford heavy-duty 19 the 19 then your network it was trained on about big a million image dataset call image and the great

42:15

thing is the play of generously made the new what weights for all available under Creative Commons license Korea CC

42:21

Attribution and you can get of there you there's also a pair Python pickled version you grab well

42:31

they're very simple and effective models like consists of 3 by the convolutions max pooling and fully connected layers that's the architecture and a classified image of EEG 19 Russia you IPython nobody will do that alright

42:56

so wouldn't take an image to classify resolvable

43:00

Peacock here we let that work 0 sorry but you problems FIL the goal of the court again classified we're going to load in our pre-trained network um I'm living above skip over the cared about is inhibited all but effectively and you can go through the nobody itself it's almost only get help again 1 of us been through this quite quickly so reasoning through but above all the models like um OK this is where we have to build a architecture so you can see the input layer we go all convolutional layers max-pooling is always on your appeal and will skip all this will go down find the output which is the softmax nonlinearity building it uh when drop all parameters in they keep pardon OK sorry by this is originally from our tutorial so anyway finally we show our image of a guy classified and we predict all probabilities here we notice the output is a 1 a vector of thousand probabilities and we find out the predicted classes

44:11

84th probability 98 100 % is because and you can run that yourself and you find out the work so the cool thing is

44:19

you can take the pre-trained network and using itself um transfer

44:32

learning is a cool trick in this is the last 4 columns show you training and from scratch in data hungry the reason is the atomic training data and preparing a let's time-consuming expensive what if we don't have enough training data to get good results without money to repair the repair Will the ImageNet dataset is really huge millions of images with ground truths and what if we could somehow use

44:58

it's to how what we could somehow use the image that dataset models fast data to help us with a different task for the good news is we can and the trick is

45:10

this rather than try and reuse the day so you train your net well IVG 19 we download BG 19 and you can take part of that network and retain throw away the end part it and and and stick sticks in the stuff on the and the lot of what

45:30

we want and that way you effectively to be trained as the bit that you've added and then fine-tune the analog over but essentially we can use you can reuse part of EEG 19 to say classify images the weren't image two-dimensional and 4 classes

45:44

and the different kinds of object category the 1 mentioned in image now so you can reuse it you can reuse it for localization say wanna find location knowledge about the location of that well that may be

45:56

uh all segmentation we wanna find exact outline of the boundary to the transfer learning what we do this without a region

46:05

19 that looks like that as well as we chop off those last 3 that stuff on the left gets for so we can just time but we chop off the last 3 letters we create a new ones randomly initialized on the end then all we do is to train than ever was any all training data the only gonna learn the promising train the parameters of the new layers that you created many we train

46:38

proximal on all layers having trained initial new ones you

46:41

train you then find him the whole lot it is due to train this time of abatement parameters of all layers and this will be some better accuracy and the result is a nice shiny new network with good performance on your particular target domain this can be solved better than you could get with so uh starting out with your interest with you datasets OK so finally

47:05

some cool work in the field of might be of interest to you as I I think I mentioned this briefly ready that they visit lists and send uh the

47:14

invisibility understanding convolutional networks they decide to visualize the responses of the convolutional layers to very

47:20

simple scenes images well inside to visualize what's going on there we wanna find all we know it's picking up this is this is a good place to look for how to how to work out what your network is detecting and these guys decide to figure out the conformal neural network so decided generate images directly unrecognizable to humanize the recognized by the

47:41

network so for instance the

47:45

network has a high degree of the a high confidence that is in fact Robin it looks like man looks like a horrible noise but it thinks acid-treated but that's an armadillo but that's a peacock really got simply color then went

48:03

on to that was so hot we generate images that they may still make sense to humans that's a king penguin

48:10

that starfish kind words picking things up it's looking for texture but is not really looking for the actual structure of the object so it's picking up certain means ignoring the ignoring other quite important features and you can run your networks works in reverse you can get to generate images as well as as well as classify them so these

48:29

guys decide to make them generate chairs so the give the

48:33

orientations on the top of the parameters of the chair and the trying to generate an image

48:38

Sinhalese chairs to more from and this 1 got a lot of press no artistic style user you if you got the prison and people also about 1 more assurance they took Oxford that from extract texture

48:55

features from 1 image and apply them to the other so you take the efforts of the so

49:05

this waterfront and you take a painting like say starring of of anger and repaints the image in the style of anger on the start of the government is screen on these others is very cool and nice thing is that the iPhone apps that do this now and what these guys did is this

49:24

is this is this is a bit of a masterpiece of work they are they

49:27

generate these these images of Belgium generated by on your network and why they did it is they're trying to tune your networks want to be a master forger on there to be the detective the master folder tries to generate an image and the detective tries to tell us our real Benjamin image of a better word that 1 has been generated by the by the forger and the idea is you can adapt them to get both better so that the

49:49

master folding spectrum better match until it generates pictures but that which is comical and even to give further by

49:56

figuring out what the so by combining some of parameters

50:00

and the it if you seen some of the and results from the universe of the King minus mental only screens stuff has been done on some of the bottom word to rectangular were to the most similar things with the facial expression as well anyway have you

50:18

found helpful I have being good a given a great audience

50:23

thank you very much

50:26

thank thank you and we have about 2 and 9 minutes and the questions the it was a great dog thank you I have actually several questions from the 1st 1 uh when you are modeling a neural network how do you chew or is there a way to to was uh how many heater layers and neutrons IUD airing them because know debts and that was an issue for me when I was modelling some I'm not aware of

51:26

any particular so the rule of thumb choose how to design a network architecture was a little thumb I used to look at something with things that work for other people and and all of that so the occident architecture where you've got the I found a small convolution kernels for people from the small convolutional kernels were well a few of those letters followed by max-pooling starting on those blocks repeats I think that I think there's some people that probably try things like grid search for used to try all the public and automatically also the architecture of a given the fact for semantic image pneumonia training on can extend even in 2 weeks all hours at least on a really big GPU is that can be impractical so uh afraid to say is this rule of thumb is for us and just right out and see it for a year while about a couple exceptions to all the people done this adaptive firms I'm sorry I additional information on my 2nd question which we

52:23

saw that you guys are analyzing images and numbers and is the way that you can like make strings input and to recognize patterns in them how would you do that would you have to like transform them some call mean for text processing in so 1

52:44

I think what people tend to do that and user might want to write to convert each word into an embedding which is like a 611 vector um and then use what's called a recurrent neural network where rather than just having it goes to the output because partially through and feed back into nearly aligned and so that it is of has an idea of time and I not you of non-payment abysmal sounds and figure and I'm lots of my comfort zone and terms of use of ICT in the look of recurrent neural networks of but yeah they didn't use the word embeddings tend to use the word embeddings to convert the words into vector this sort of cultural way of doing that is to just have turned into 1 hot representation I forgot 2000 words in new vocabulary is simply have a reason have 1 1 you have a 1 on 2 groups and the work of the vector of all zeros except 1 for the particular what that is but given the sparsity of the often causes problems which will use the embeddings and the last question

53:46

sorry uh key could you train the neural network to do like Matt like addition maybe multiplication and if you can't would it be may be faster than the usual way of that process of doing it um you can try

54:01

to do addition I think actually there some people managed to take the image that dataset we take 200 digits in an image the figures out what they are remember and then is trained to produce the some of it can work multiplication but I do know that should be above figure out again units do that so the models actually called extends certain things which is interesting and so there's certain things that is don't do very well well so things quite limited as for would be fostered where the fast and so using however the mathematical operations to do something that is a is a 1 instruction operation processes true that

54:43

thank social career really interesting and great stuff him around the images and we'll thoughts on how neural networks could be applied to text analytics because most people don't do that some text analytics firm is

55:03

also my area so I don't know but uh I would speak to Catherine Jamal she's here and she did a very very good talk uh describing the sort of the students to give a really really good insuring really really good sort of overview of what this text processing world is like and she gave quite few non-animals new networks or some of the best models right now but it's also my expertise but she knows the stuff on the outside speak to come any other questions I

55:40

will then be the name of neural networks comes to the size of the plane dual if it's a used widely in in science of not sure and I think that

55:52

the and in the model that we use for our new networks that have been talking about here is quite different from my neurons in the brain work I think that my my very very basic layman's understanding of brain neurons is they they operate on spike rate so that they they they generally output spikes in the frequency of that is often the strength of the output of thing I don't know uh so I think that so trying to like in these 2 others so I I don't think of it that much alike I think that the way where the similarity is is that people look to how neurons in the brain up to recover the child and they said Hakim makes and models this is the what we've got something that seems to work well given of processes and systems produce very good pattern recognition so that uh as for similarities to bring beyond that controlled same or than any other questions hi how have you heard of the

56:58

self driving car using deep learning to you implement how the driving I wonder how the would update of the cost function because is a stream of we DO rather than a fixed static output I've heard about

57:14

burned from motion of the whole of doing it and at the future if you're to try and do something about that 1 of the things you could do is you could prepare a bunch of footage we say the human who's driving this call has done well as may have and crash to kill me 1 of them's history like that so the idea is that what's good and maybe if there some footage accidents that's a that's bad don't do all they do is to what will you probably want to do it is you want to say given this video however these outputs as and still like this acceleration brake this produce these decisions and so that's actually bit like a little bit like the Atari Game Playing networks works Google developed the stuff for the global of really good schools in the video games will take the input screens and they decide whether to move up down left right shoot but a similar thing where a set of deciding whether to move up down left right show you want you control the steerable accelerator and the brakes you could do it like that but given my experience of and given the father of as I mentioned we have particularly rare examples of rare situations where coffin your neighborhood as treaty not because they might make up 0 . 0 0 0 1 % your training set uh it'll never but it'll never actionable to learn anything from those list of the cost function you it'll it'll discover local minima the signals that I would not be very comfortable getting into a call that was just controlled on your network I would not want my life in the hands of the but does not behave could build the whether it would undergo the very good and the

59:07

questions hi do you ever combined neural networks of our techniques like an approximation algorithms approximation algorithms yeah like optimization techniques boasting about travel salesman problem for example of

59:31

and know it's even even tried for the unknown not were unaware that I would be surprised instructive unaware of some looks is afraid and this the going on forever sorry uh it argues that we have to figure out a way of constructing a cost function of a some measures of the solution so how long ago about symbolism I have time for

1:00:00

1 last question the you may be asking about technical question about the ads but uh when you apply dropout does the Ys expression get real recompile recognize tools the efficient need not to take account of the uh sort of of also weights or a uh the gets sequelae in all you be floating point operations and get to the GP or CPU all but there is you so that the effect the gradient I think it's a 2nd because what you do is you get to because random number generator

1:00:52

generate 0 1 and the multiple you put that multiplying expressions I think it's just selection optimizing something we quite difficult to optimize because the problem is on and for a single sample and the budgetary blocking out difference of subset of the the the units are not even sure how what actually go optimizing an efficient way because you gotta always select which user dropping elements the from that decide what you can do operations you can save you can't to go do it on the fly and can be quite tough some guessing it I would guess that doesn't it so now since

1:01:29

then on the question so and probably time chip was 1 of them talk and fj and probably as AS giants

00:00

Resultante

Multiplikation

Selbst organisierendes System

Güte der Anpassung

Gebäude <Mathematik>

Ruhmasse

Bildanalyse

Überlagerung <Mathematik>

Code

Computeranimation

Soft Computing

Informationsmodellierung

Code

Programmbibliothek

Vorlesung/Konferenz

Maschinelles Sehen

Neuronales Netz

Maschinelles Sehen

Neuronales Netz

01:11

Resultante

Knotenmenge

Informationsmodellierung

Gewicht <Mathematik>

Gewicht <Mathematik>

Wärmeübergang

Faltungsoperator

Programmbibliothek

Vorlesung/Konferenz

Neuronales Netz

Computeranimation

Neuronales Netz

02:11

Rechenschieber

Informationsmodellierung

Verknüpfungsglied

Materialisation <Physik>

Browser

Vorlesung/Konferenz

Wiederkehrender Zustand

Elektronische Publikation

Code

Computeranimation

03:03

Subtraktion

Transinformation

Klasse <Mathematik>

Profil <Aerodynamik>

Unrundheit

Code

Teilbarkeit

Computeranimation

Rechenschieber

Spezialrechner

Graphikprozessor

Virtuelle Maschine

Gewicht <Mathematik>

Menge

Datentyp

Vorlesung/Konferenz

Bildgebendes Verfahren

Einfügungsdämpfung

04:02

Kraftfahrzeugmechatroniker

Klasse <Mathematik>

Vektorraum

Objektklasse

Term

Computeranimation

Spezialrechner

Soft Computing

Gewicht <Mathematik>

Vorlesung/Konferenz

Bitrate

Maschinelles Sehen

Bildgebendes Verfahren

Auswahlaxiom

Schreib-Lese-Kopf

Neuronales Netz

04:50

Compiler

Regulärer Ausdruck

Übergang

Physikalisches System

Bitrate

Punktspektrum

Term

Computeranimation

Übergang

Netzwerktopologie

Spezialrechner

Software

Arithmetischer Ausdruck

RIS <Medizin, Informationssystem>

Gewicht <Mathematik>

Spieltheorie

Faltungsoperator

Vorlesung/Konferenz

Compiler

Bitrate

Neuronales Netz

Fehlermeldung

Neuronales Netz

06:11

Softwareentwickler

Web Site

Bit

Selbstrepräsentation

Ausbreitungsfunktion

Vektorraum

Zentraleinheit

Information

Datenfluss

Computeranimation

Spezialrechner

Deskriptive Statistik

Software

Neuronales Netz

Bildgebendes Verfahren

Gammafunktion

Neuronales Netz

07:20

Matrizenrechnung

Gewicht <Mathematik>

Klasse <Mathematik>

Zahlenbereich

Vektorraum

Ein-Ausgabe

Term

Computeranimation

Einfache Genauigkeit

Datensatz

Neuronales Netz

Gerade

Neuronales Netz

Funktion <Mathematik>

08:08

Resultante

Matrizenrechnung

Gewicht <Mathematik>

Element <Mathematik>

Matrizenrechnung

Automatische Handlungsplanung

Zahlensystem

Term

Computeranimation

Ausdruck <Logik>

Multiplikation

Einheit <Mathematik>

Sigmoide Funktion

Stichprobenumfang

Neuronales Netz

Bildgebendes Verfahren

Normalvektor

Nichtlineares System

Funktion <Mathematik>

Parametersystem

Lineares Funktional

Pixel

Vektorraum

Objektklasse

Portscanner

Mapping <Computergraphik>

Rechenschieber

Funktion <Mathematik>

Mereologie

Parametersystem

Ein-Ausgabe

Neuronales Netz

09:46

Lineares Funktional

Bit

Subtraktion

Wellenpaket

Kategorie <Mathematik>

Wasserdampftafel

Backpropagation-Algorithmus

Extrempunkt

Ein-Ausgabe

Computeranimation

Linearisierung

Prognoseverfahren

Funktion <Mathematik>

Mittelwert

Kostenfunktion

Vorlesung/Konferenz

Neuronales Netz

Funktion <Mathematik>

Neuronales Netz

Fehlermeldung

Nichtlineares System

10:52

Sichtbarkeitsverfahren

Lineares Funktional

Exponent

Likelihood-Funktion

Klasse <Mathematik>

Zahlenbereich

Element <Mathematik>

Ein-Ausgabe

Extrempunkt

Computeranimation

Funktion <Mathematik>

Kostenfunktion

Vorlesung/Konferenz

Bildgebendes Verfahren

Neuronales Netz

11:39

Sichtbarkeitsverfahren

Lineares Funktional

Lineare Regression

Subtraktion

Gewichtete Summe

Kategorie <Mathematik>

Gewichtete Summe

Ein-Ausgabe

Computeranimation

Linearisierung

Funktion <Mathematik>

Lineare Regression

Endogene Variable

Nichtunterscheidbarkeit

Vorlesung/Konferenz

Bildgebendes Verfahren

Fehlermeldung

Neuronales Netz

Funktion <Mathematik>

12:35

Parametersystem

Gewicht <Mathematik>

Derivation <Algebra>

Computer

Gradient

Code

Computeranimation

Gradient

Portscanner

Arithmetischer Ausdruck

Vorzeichen <Mathematik>

Tensor

Maschinenschreiben

Parametersystem

Gradientenverfahren

Derivation <Algebra>

Computeralgebra

Datenfluss

Fehlermeldung

13:42

Parametersystem

Wellenpaket

Gewicht <Mathematik>

Stapelverarbeitung

Derivation <Algebra>

Bitrate

Computeranimation

Gradient

Einfache Genauigkeit

Wellenpaket

Mittelwert

Stichprobenumfang

Vorlesung/Konferenz

Bitrate

Tabelle <Informatik>

14:33

Mittelwert

Prozess <Physik>

Wellenpaket

Stapelverarbeitung

Iteration

Computeranimation

Multiplikation

Iteration

Einheit <Mathematik>

Mittelwert

Stichprobenumfang

Gradientenverfahren

Vorlesung/Konferenz

Neuronales Netz

Parallele Schnittstelle

Ganze Funktion

Einflussgröße

Parallele Schnittstelle

Leistung <Physik>

Parametersystem

Validität

Stichprobe

Uniforme Struktur

Funktion <Mathematik>

Menge

Rechter Winkel

Digitalisierer

Parametersystem

Computerarchitektur

Perzeptron

Neuronales Netz

Fehlermeldung

16:07

Ortsoperator

Randverteilung

Zentrische Streckung

Gewicht <Mathematik>

Ortsoperator

Invarianz

Einfach zusammenhängender Raum

Computeranimation

Spezialrechner

Einheit <Mathematik>

Softwareschwachstelle

Digitalisierer

Zahlzeichen

Translation <Mathematik>

Translation <Mathematik>

Neuronales Netz

Bildgebendes Verfahren

Neuronales Netz

16:56

Sichtbarkeitsverfahren

Kernel <Informatik>

TVD-Verfahren

Wellenpaket

Gewicht <Mathematik>

Computeranimation

Spezialrechner

Knotenmenge

Einheit <Mathematik>

Neuronales Netz

Grundraum

Maschinelles Sehen

Bildgebendes Verfahren

Ortsoperator

Nichtlinearer Operator

Pixel

Analoge Signalverarbeitung

Spider <Programm>

Ein-Ausgabe

Soft Computing

Menge

Digitalisierer

Faltungsoperator

URL

Computerunterstützte Übersetzung

Pixel

Lie-Gruppe

Faserbündel

Neuronales Netz

18:10

Nachbarschaft <Mathematik>

Kernel <Informatik>

Einfügungsdämpfung

Gewicht <Mathematik>

Ortsoperator

Wellenlehre

Computeranimation

RFID

Kernel <Informatik>

Knotenmenge

Vorlesung/Konferenz

Maschinelles Sehen

Bildgebendes Verfahren

Gerade

Funktion <Mathematik>

Lineares Funktional

Filter <Stochastik>

Multifunktion

Pixel

Green-Funktion

Benutzerfreundlichkeit

Ein-Ausgabe

Biprodukt

Design by Contract

Rechenschieber

Bildschirmmaske

Soft Computing

Rechter Winkel

Faltungsoperator

Mereologie

Maschinelles Sehen

Neuronales Netz

20:14

Kernel <Informatik>

Orientierung <Mathematik>

Matrizenrechnung

Subtraktion

Gewicht <Mathematik>

Mathematisierung

Matrizenrechnung

Computeranimation

Kernel <Informatik>

Eins

Demoszene <Programmierung>

Multiplikation

Spannweite <Stochastik>

Bildgebendes Verfahren

Funktion <Mathematik>

Pixel

Einfach zusammenhängender Raum

Schwach besetzte Matrix

Quick-Sort

Office-Paket

Quadratzahl

Funktion <Mathematik>

Faltungsoperator

Mereologie

Speicherverwaltung

Computerarchitektur

Pixel

Neuronales Netz

Varietät <Mathematik>

21:19

Punkt

Extrempunkt

Matrizenrechnung

Extrempunkt

Computeranimation

Übergang

Gradient

PCMCIA

Stichprobenumfang

Konstante

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Stochastische Abhängigkeit

Nichtlinearer Operator

Pixel

Stichprobennahme

Ruhmasse

Bildauflösung

p-Block

Konfiguration <Informatik>

Portscanner

Energiedichte

Funktion <Mathematik>

Kantenfärbung

Maschinelles Sehen

Pixel

22:38

Resultante

Pixel

Stichprobennahme

Singularität <Mathematik>

Versionsverwaltung

Extrempunkt

Ein-Ausgabe

Faltungsprodukt

Quick-Sort

Computeranimation

Kernel <Informatik>

Portscanner

Funktion <Mathematik>

PCMCIA

Zahlzeichen

Faltungsoperator

Vorlesung/Konferenz

Pixel

Bildgebendes Verfahren

Funktion <Mathematik>

Neuronales Netz

23:31

Kernel <Informatik>

Wellenpaket

Datenmodell

Programmierumgebung

Paarvergleich

Computeranimation

Kernel <Informatik>

Spezialrechner

Metropolitan area network

Einheit <Mathematik>

Iteration

Minimum

Zahlzeichen

Bildgebendes Verfahren

Funktion <Mathematik>

Neuronales Netz

24:24

Kernel <Informatik>

Nichtlinearer Operator

Orientierung <Mathematik>

Zellularer Automat

Quick-Sort

Computeranimation

Kernel <Informatik>

Objekt <Kategorie>

Spezialrechner

Textur-Mapping

Quadratzahl

Rechter Winkel

Mereologie

Vorlesung/Konferenz

Kurvenanpassung

Lie-Gruppe

Neuronales Netz

25:19

Lineares Funktional

Einfügungsdämpfung

Mathematik

Compiler

Regulärer Ausdruck

Übergang

Computeranimation

Übergang

Rechenschieber

Arithmetischer Ausdruck

Mereologie

Vorlesung/Konferenz

Datenfluss

Neuronales Netz

Neuronales Netz

Funktion <Mathematik>

26:41

Resultante

Kernel <Informatik>

Bit

Mereologie

Zahlenbereich

Extrempunkt

MP3

Zahlensystem

Computeranimation

Kernel <Informatik>

Metropolitan area network

Stichprobenumfang

Inverser Limes

Vorlesung/Konferenz

Neuronales Netz

Filter <Stochastik>

Stichprobennahme

p-Block

Digitalfilter

Ein-Ausgabe

Faltungsoperator

Mereologie

Computerarchitektur

Kantenfärbung

p-Block

Neuronales Netz

27:36

Filter <Stochastik>

Bit

Punkt

Extrempunkt

Zahlenbereich

p-Block

Digitalfilter

MP3

Computeranimation

Kernel <Informatik>

Zahlensystem

Faltungsoperator

p-Block

Bildauflösung

28:27

Ortsoperator

Softwareentwickler

Resonanz

Stapelverarbeitung

Zahlensystem

Computeranimation

Spezialrechner

Vorlesung/Konferenz

Information

URL

Computerarchitektur

Normalvektor

Stapelverarbeitung

Neuronales Netz

Bildgebendes Verfahren

Funktion <Mathematik>

Neuronales Netz

29:15

Größenordnung

Matrizenrechnung

Einfügungsdämpfung

Wellenpaket

Gewicht <Mathematik>

Güte der Anpassung

Stapelverarbeitung

Zahlenbereich

Ein-Ausgabe

Computeranimation

Arithmetisches Mittel

Multiplikation

Rechter Winkel

Konstante

Faltungsoperator

Vorlesung/Konferenz

Größenordnung

Stapelverarbeitung

Normalvektor

Neuronales Netz

Neuronales Netz

Standardabweichung

30:23

Softwaretest

Wellenpaket

Datenmodell

Systemaufruf

Ausgleichsrechnung

Zeitreise

Computeranimation

Chirurgie <Mathematik>

Informationsmodellierung

Softwaretest

Prognoseverfahren

Gruppe <Mathematik>

Stichprobenumfang

Algorithmische Lerntheorie

Neuronales Netz

Neuronales Netz

Normalvektor

31:12

Softwaretest

Wellenpaket

Multiplexbetrieb

Term

Teilbarkeit

Computeranimation

Teilmenge

Rhombus <Mathematik>

Multiplikation

Einheit <Mathematik>

Funktion <Mathematik>

Stichprobenumfang

Randomisierung

Größenordnung

Tropfen

Normalvektor

Funktion <Mathematik>

32:12

Bit

Gewicht <Mathematik>

Wellenpaket

Backpropagation-Algorithmus

Versionsverwaltung

Unrundheit

Überlagerung <Mathematik>

Quick-Sort

Computeranimation

Teilmenge

Erweiterte Realität <Informatik>

Einheit <Mathematik>

Menge

Stichprobenumfang

Tropfen

Gerade

Neuronales Netz

33:05

Wellenpaket

Transformation <Mathematik>

Transformation <Mathematik>

Computeranimation

Arithmetisches Mittel

Erweiterte Realität <Informatik>

Einheit <Mathematik>

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Varianz

Informationssystem

Fitnessfunktion

Neuronales Netz

Aggregatzustand

Standardabweichung

33:57

Distributionstheorie

Zentrische Streckung

Lineare Regression

Pixel

Green-Funktion

Ablöseblase

Computer

Ein-Ausgabe

Speicherbereichsnetzwerk

Raum-Zeit

Computeranimation

Arithmetisches Mittel

Spannweite <Stochastik>

Funktion <Mathematik>

Reverse Engineering

Rechter Winkel

Lineare Regression

Ein-Ausgabe

Stichprobenumfang

Server

Vorlesung/Konferenz

Neuronales Netz

Pixel

Bildgebendes Verfahren

Standardabweichung

Funktion <Mathematik>

Neuronales Netz

34:50

Lineares Funktional

Einfügungsdämpfung

Wellenpaket

Punkt

Weg <Topologie>

Bitrate

Teilbarkeit

Computeranimation

RFID

Portscanner

Vorlesung/Konferenz

Bitrate

Standardabweichung

Neuronales Netz

Fehlermeldung

35:38

Konstante

Extrempunkt

Konstante

Klasse <Mathematik>

Mustersprache

Vorlesung/Konferenz

Extrempunkt

Neuronales Netz

Computeranimation

Instantiierung

36:28

Resultante

Extrempunkt

Zahlenbereich

Element <Mathematik>

Extrempunkt

Computeranimation

Soft Computing

Kostenfunktion

Digitalisierer

Konstante

Vorlesung/Konferenz

Maschinelles Sehen

Maschinelles Sehen

Neuronales Netz

Expertensystem

Neuronales Netz

37:17

Distributionstheorie

Bit

Wellenpaket

Einfache Genauigkeit

Komplex <Algebra>

Computeranimation

Einfache Genauigkeit

Mehrrechnersystem

Flächentheorie

Mereologie

Mustersprache

Hypercube

Randomisierung

Vorlesung/Konferenz

Neuronales Netz

Figurierte Zahl

Bildgebendes Verfahren

Funktion <Mathematik>

Neuronales Netz

38:59

Netzwerktopologie

Wellenpaket

Quader

Mereologie

Zahlenbereich

Vorlesung/Konferenz

Bildgebendes Verfahren

Computeranimation

Entscheidungstheorie

39:50

Spezialrechner

Quader

Metropolitan area network

Videospiel

Punkt

Wellenpaket

Endogene Variable

Stellenring

Inverser Limes

Geräusch

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Computeranimation

40:52

Spezialrechner

Metropolitan area network

Orientierung <Mathematik>

Wellenpaket

Ortsoperator

Wellenpaket

Minimum

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Computeranimation

Schreib-Lese-Kopf

41:40

Gewicht <Mathematik>

Versionsverwaltung

Datenmodell

Systemaufruf

Wärmeübergang

Objektklasse

Computeranimation

Spezialrechner

Gewicht <Mathematik>

Wellenpaket

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Neuronales Netz

Attributierte Grammatik

42:28

Dualitätstheorie

Klasse <Mathematik>

Regulärer Ausdruck

Turing-Test

Extrempunkt

Computeranimation

Spezialrechner

Metropolitan area network

Informationsmodellierung

Last

Mailing-Liste

Vorlesung/Konferenz

Tropfen

Neuronales Netz

Bildgebendes Verfahren

Gammafunktion

Funktion <Mathematik>

Nichtlineares System

Demo <Programm>

Parametersystem

Datenmodell

Magnetooptischer Speicher

Vektorraum

Ein-Ausgabe

Variable

Menge

Faltungsoperator

Parametersystem

ATM

Computerarchitektur

Modelltheorie

Personal Area Network

Neuronales Netz

44:11

Resultante

Wellenpaket

Güte der Anpassung

Wärmeübergang

Ordinalzahl

Computeranimation

Spezialrechner

Dämpfung

Gewicht <Mathematik>

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Neuronales Netz

44:58

Inklusion <Mathematik>

Bit

Subtraktion

Mereologie

Gewicht <Mathematik>

Kategorie <Mathematik>

Klasse <Mathematik>

Stellenring

Computeranimation

Portscanner

Objekt <Kategorie>

Task

Spezialrechner

Metropolitan area network

Informationsmodellierung

Task

Gewicht <Mathematik>

Mereologie

Vorlesung/Konferenz

URL

Neuronales Netz

Bildgebendes Verfahren

Dimension 2

45:56

Resultante

Parametersystem

Wellenpaket

Güte der Anpassung

Wärmeübergang

Extrempunkt

MP3

Computeranimation

Eins

Quader

Randwert

Domain-Name

Font

Wellenpaket

Parametersystem

Wärmeübergang

Vorlesung/Konferenz

Abstand

Neuronales Netz

Neuronales Netz

47:00

Geräusch

Computeranimation

Kreisbogen

Endogene Variable

Demoszene <Programmierung>

Spezialrechner

Uniforme Struktur

Datenfeld

Minimalgrad

Bereichsschätzung

Endogene Variable

Faltungsoperator

Vorlesung/Konferenz

Neuronales Netz

Bildgebendes Verfahren

Neuronales Netz

Instantiierung

Metropolitan area network

48:00

Parametersystem

Orientierung <Mathematik>

Reverse Engineering

Datenmodell

Computeranimation

Digitale Photographie

RFID

Arithmetisches Mittel

Objekt <Kategorie>

Gefangenendilemma

Textur-Mapping

Spezialrechner

Iteration

Gewicht <Mathematik>

Funktion <Mathematik>

Reverse Engineering

Ein-Ausgabe

Parametersystem

Vorlesung/Konferenz

Wort <Informatik>

Datenstruktur

Neuronales Netz

Bildgebendes Verfahren

Neuronales Netz

48:55

Inklusion <Mathematik>

Spezialrechner

Bit

Reelle Zahl

Vorlesung/Konferenz

Wort <Informatik>

Bildgebendes Verfahren

Computeranimation

Neuronales Netz

Touchscreen

49:48

Resultante

Sichtbarkeitsverfahren

Parametersystem

Matching <Graphentheorie>

Punktspektrum

Computeranimation

Spezialrechner

Metropolitan area network

Arithmetischer Ausdruck

Informationsmodellierung

Vorlesung/Konferenz

Wort <Informatik>

Faltung <Mathematik>

Grundraum

Touchscreen

Neuronales Netz

51:25

Prozess <Physik>

Wellenpaket

Thumbnail

Zahlenbereich

Schlussregel

Ausnahmebehandlung

p-Block

Ein-Ausgabe

Formale Semantik

Mustersprache

Faltungsoperator

Computerarchitektur

Information

Bildgebendes Verfahren

Zeichenkette

Neuronales Netz

52:41

Addition

Topologische Einbettung

Prozess <Physik>

Selbstrepräsentation

Gruppenkeim

Vektorraum

Schwach besetzte Matrix

Term

Zeitzone

Quick-Sort

Differenzengleichung

Wort <Informatik>

Schlüsselverwaltung

Figurierte Zahl

Funktion <Mathematik>

Neuronales Netz

54:00

Addition

Nichtlinearer Operator

Informationsmodellierung

Multiplikation

Einheit <Mathematik>

Digitalisierer

Analytische Menge

Figurierte Zahl

Bildgebendes Verfahren

Informationssystem

Neuronales Netz

54:59

Ebene

Informationsmodellierung

Prozess <Physik>

Flächeninhalt

Güte der Anpassung

t-Test

Dualitätstheorie

Quick-Sort

Neuronales Netz

55:48

Hydrostatik

Streaming <Kommunikationstechnik>

Informationsmodellierung

Prozess <Physik>

Kostenfunktion

Güte der Anpassung

Ähnlichkeitsgeometrie

Physikalisches System

Mustererkennung

Frequenz

Neuronales Netz

Funktion <Mathematik>

57:13

Nachbarschaft <Mathematik>

Videospiel

Approximationsgüte

Bit

Approximation

Schießverfahren

Wellenpaket

Extrempunkt

Minimierung

Güte der Anpassung

Systemaufruf

Mailing-Liste

Ein-Ausgabe

Videokonferenz

Entscheidungstheorie

Travelling-salesman-Problem

Computerspiel

Menge

Spieltheorie

Kostenfunktion

Touchscreen

Neuronales Netz

59:30

Soundverarbeitung

Nichtlinearer Operator

Punkt

Gewicht <Mathematik>

Compiler

Symboltabelle

Automatische Differentiation

Zentraleinheit

Zufallsgenerator

Gradient

Arithmetischer Ausdruck

Reelle Zahl

Kostenfunktion

Einflussgröße

1:00:51

Teilmenge

Nichtlinearer Operator

Arithmetischer Ausdruck

Subtraktion

Einheit <Mathematik>

Trennschärfe <Statistik>

Stichprobenumfang

Element <Mathematik>

1:01:43

Metropolitan area network

Computeranimation

### Metadaten

#### Formale Metadaten

Titel | An Introduction to Deep Learning |

Serientitel | EuroPython 2016 |

Teil | 166 |

Anzahl der Teile | 169 |

Autor | French, Geoff |

Lizenz |
CC-Namensnennung - keine kommerzielle Nutzung - Weitergabe unter gleichen Bedingungen 3.0 Unported: Sie dürfen das Werk bzw. den Inhalt zu jedem legalen und nicht-kommerziellen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen und das Werk bzw. diesen Inhalt auch in veränderter Form nur unter den Bedingungen dieser Lizenz weitergeben |

DOI | 10.5446/21145 |

Herausgeber | EuroPython |

Erscheinungsjahr | 2016 |

Sprache | Englisch |

#### Inhaltliche Metadaten

Fachgebiet | Informatik |

Abstract | Geoff French - An Introduction to Deep Learning Deep learning: how it works, how to train a deep neural network, the theory behind deep learning, recent developments and applications. ----- (length: 60 mins) In the last few years, deep neural networks have been used to generate state of the art results in image classification, segmentation and object detection. They have also successfully been used for speech recognition and textual analysis. In this talk, I will give an introduction to deep neural networks. I will cover how they work, how they are trained, and a little bit on how to get going. I will briefly discuss some of the recent exciting and amusing applications of deep learning. The talk will primarily focus on image processing. If you completely new to deep learning, please attend T. Rashid's talk 'A Gentle Introduction to Neural Networks (with Python)'. His talk is in the same room immediately before mine and his material is really good and will give you a good grounding in what I will present to you. |