Generative Adversarial Networks - FUTURISTIC & FUN AI !

our world has moved into a phase where we see new technology that fascinates us every day for example Nvidia's developed an AI that produces highly detailed images of human looking faces but the people depicted don't actually exist while a I created faces are absolutely remarkable they are also incredibly eerie these are faces that you would believe to be real people had you not known they were actually machine generated in the artistic front we have the drawing tool pix depicts it's an image to image translator where an image is synthesized from a drawing on a canvas it's pretty interesting to see the generated images adding texture and depth perception based on pencil strokes this type of synthesis is not only localized to image data in September 2016 google's deepmind released a paper on wavenet it's a speech synthesizer that generates voices that are nearly indistinguishable from human speech the primary tool used in all of these applications are cans generative adversarial networks and that is what we're going to talk about today the idea of Ganz was conceived in 2014 by young good fellow so why do we use them there are many times where we just don't have enough data to create a model ganz can learn about your data and learn to synthesize or generate never-before-seen data to augment your data set consider them as an approach to unsupervised learning and even send me supervised learning I say this because semi-supervised learning is used when some of your data is labeled while the rest are not we may have a labelled data set but it may not be enough to train a model in this case we could generate unlabeled data using Ganz and so a semi-supervised approach is appropriate furthermore the generated data could also be just used as is like in the examples I mentioned before about face image synthesis so we got the generative part of Gans what about the adversarial part when I hear the term adversarial I think of reinforcement learning and game play so who exactly is competing against who a base again consists of two models that play against each other a generator and a discriminator think of the generator as the robber or more accurately the counterfeiter who tries to replicate the input data to produce counterfeit data or fake data the discriminator on the other hand is like a cob we needs to be able to distinguish between the real data which is from the data set and the counterfeit data which is produced by the generator the loss of the discriminator is used as the objective function for the generator in the next time around now we have six basic steps to train the gun so first of all is define the problem what exactly are we trying to do do we need to synthesize images from a caption or do we need to synthesize an image from another image or audio synthesis from sentences what is it the next thing we need to do is to define the architecture of your again that is answering questions like what is the generator what is the discriminator these are basically modeled depending on the complexity of the problem is it a multi-layer perceptron a neural network or a simpler model the third step is to train the discriminator model to distinguish between real and fake data when we speak of training we need to feed it both types we need to train it from the data in the data set and label it as real data and to train it against fake data we feed it fake data generated by our generator or the counterfeiter with the label fake data in step four we need to train the generator now so this uses the loss of the discriminator as the objective function to the generator in other words we need to modify parameters of the generative model to maximize a loss of the cop or the discriminator we then repeat the training of the generator and discriminator over n epochs after every iteration the generator will get better at fooling the discriminator finally the discriminator will not be able to tell the real images of the data set from the fabricated ones by the generator once the training is complete we synthesize data from the generator and this can be used to augment or true data set or just use as is because it's really cool so now we know the basic components of again and how it works that's great now let's get some more intuition on the loss function for our discriminator the discriminator uses a cross-entropy loss so why is that for one it's because that we are dealing with a classification model cross-entropy loss is a better performance metric than classification error or mean squared error MSE classification error computes the number of samples misclassified but does not take into account how off those predictions were it only cares about the ending classification result in cross-entropy loss on the other hand it is possible to have a model which misclassify samples more as a better model simply because when it was wrong it wasn't really wrong by that much another reason for cross-entropy loss in neural network architecture is to eliminate the vanishing gradient during training in other words the change in the weights does not become zero when using cross-entropy error loss and hence training is not stalled so to speak the cross entropy loss between the true distribution P and estimated distribution Q is given by this formula P and Q are thus vectors of M dimensions where m is the number of classes now a given sample can only belong to a single class so the true distribution is a one hot vector with one one and the rest zeros this is because we are certain that this sample belongs to a specific class the discriminator in again is a binary classifier it needs to classify data as either real or fake so in this case M is equal to 2 and the true distribution is a 1 Haunter only consisting of two terms so we can write the equations as shown this represents the loss for a single sample we can sum over the losses for n samples to get the overall loss now we know that half of the samples come from the data set and the other half is from the generator in other words we have half of the samples that follow the distribution of the real data and the other half of the samples which follow the distribution of the fake data from the generator the tilt sign represents is distributed as in mathematics let D of X be the predicted classification of the discriminator since we don't know how the samples are fed into the discriminator it makes sense to mathematically represent them as expectations rather than sums and so we end up with this form if you find it hard to remember which term of the expectation is distributed from the real data and which term is from the generator take a look at the D of X term this is either 0 or 1 depending on whether predicts the data as fake or real consider the first term with just D FX when does this term contribute to the loss function so it contributes a loss function when it is not 0 or when it is equal to 1 and this happens when the data is real the expectation of D FX when X is sampled from the generator would have been 0 and hence it isn't included now the reverse is true for the next term 1 minus D FX this term contributes to the loss function when it is 1 or when D FX is equal to 0 in other words when the data is fake and data is fake when it is sampled from the generator by the same argument the expectation of 1 minus D FX when X is sampled from the real data would be 0 and hence it isn't included I hope that was more of a comprehensive explanation of the loss function so now I'm moving on what are the types of Gans out there well the first one is the original vanilla again which I just outlined above but what about something a little more complicated like deep convolutional ganz D siegen's when we think of convolutional neural Nets we tend to think about label training data and supervised learning however DC gans demonstrate expanded use of convolutional neural networks in unsupervised learning by using CN NS in generators and discriminators more specifically consider the problem of generating images of faces like we see in Nvidia's AI the discriminator will take an input face image either from the generator or the actual data set and output real if it believes the image is that of an actual face or fake if it believes the face is not that of a human it is a binary classifier that can be implemented with convolutional neural nets the generator on the other hand will be given some data as input and we'll have to come up with a face now this is done through a D convolutional neural net the next type of gaen that we have our conditional gans we know the gans are a novel way to Train generative models but the type of data generated by the generator can be anything what if we input a condition that could dictate the type of image generated this could be done through conditional gans to give you a better example consider the Emnes data set which consists of images digit images 0 through 9 a typical Gann would be able to generate random digit images but through siegen's we can specify a condition for the generated image this condition is the label digit in this case so by feeding the digit 5 to our C again we can generate an image with the digit 5 in it in other words we can direct the generator to synthesize specific images another type of Gann that we have is an info again so this type is not only able to generate images but also learn meaningful latent variables without any labels and the data one example given in the paper is that when an info gun is trained on the Emnes data set variables representing the type of digits 0 through 9 the angle of the digit and the thickness of the stroke are all inferred automatically here's an example of a possible output when a salient variable is varied over every row you can see that as the variable value changes there is a difference in the thickness of the brushstroke another type again that we have is the Wasserstein Gann w Ganz one of the main problems of Ganz and DC Ganz is the objective function of the generator recall the objective function is to increase the loss of the discriminant but there is no clear sign of when to stop you need to keep looking at the samples and see if there are satisfactory enough to pass for real data the old method of minimizing generative loss was through minimizing a metric called the Jenson Shannon divergence it is a method of measuring the similarity between two probability distributions this new Wasserstein can however seeks to minimize a new metric called the Wasserstein distance between points in the probability distributions using the wah sourcing distance you can train the discriminator to convergence and this leads to higher quality generated data samples a link the main paper and a more accessible blog post to this in the description below there are many other ganz out there and it is a very hot topic of research now consider for example Microsoft's attention again that is attention generative adversarial Network this paper was released a few days ago this attention gain can create images from text through natural language processing and performs fine-grained tasks like generate parts of an image from a single word in the description it is able to do this through an attention mechanism where the generator focuses on generating parts of an image or data in high resolution and other parts in lower resolution as the context of the words becomes more known the surrounding parts of the image increase in resolution over time until finally we get an image that closely corresponds with the words in text there's so much fascinating technology sprouting from these Gans every day and it is an exciting topic to discuss and there's so much potential we'll come back to this topic in a future video I'll include a list of links to papers and interesting blog posts in the description down below I'll also include a cool application links like the pics to pics application the hand-drawn image synthesis tool thank you all for stopping by today and if you liked the video click that like button and subscribe for more content on data Sciences machine learning and deep learning did you click Subscribe please click Subscribe

Channel: CodeEmporium

Views: 37,367

Keywords: GANs, Generative Adversarial Networks, Technology, Deep Learning, Machine Learning, Wavenet, Research in Machine Learning, artificial intelligence, machine learning, neural networks, neural net, neural network, variational autoencoder, autoencoder, lstm, recurrent neural network, convolutional, convolutional neural network, generative model, discriminative model, deep learning, programming, computer science, coding, python, gaussian

Id: O8LAi6ksC80

Length: 14min 20sec (860 seconds)

Published: Tue Jan 23 2018