Eat more nutritious foods like fruits, vegetables, grass-fed meats, fish, and oatmeal, but don’t forget the less nutriti

There are numerous applications for GANs and many new applications are coming out all the time. But since this article is about computer vision, two extremely interesting applications of GANs are:,Generative adversarial networks or GANs for short were introduced by Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio in their paper “Generative Adversarial Networks” which you can read here.,Second, YOLO reasons globally about the image when making predictions. Unlike sliding window and region proposal-based techniques, YOLO sees the entire image during training and test time so it implicitly encodes contextual information about classes as well as their appearance.,YOLO divides the input image into an SxS grid. And for each grid predicts if the centre of an object is present within the grid. If the centre of the object is in the grid, the grid will then predict a bounding box with 5 values, x,y,w,h,c. (x,y) are the coordinates of the centre of the object relative to the grid, (w,h) is the width and height of the object relative to the whole image and (c ) is the class of the object.,Third, YOLO learns generalizable representations of objects. Since YOLO is highly generalizable it is less likely to break down when applied to new domains or unexpected inputs.”,Deep fakes I’m sure everybody has heard of deep fakes from the media. Deep fakes are also GANs where the generator is trained to perform the faking operation, and the critic is tasked with detecting the fake. The generator can be trained for long enough to fool most humans. This is a somewhat dangerous technology and something to be aware of on the internet.,YOLO stands for you only look once. When the paper was released, the popular method for object detection was to reuse classifiers to classify local regions of an image and use a sliding window approach to check if each region of an image has an object. YOLO shifted the paradigm by proposing object detection as a regression problem, where they only use a single network for the entire pipeline and process the whole image at once rather than in regions.,Super-resolution Super-resolution refers to taking a low-quality image and generating a high-quality image from it. Nvidia’s new DLSS could be using this technique. Jeremey Howard from fast.ai has an extremely interesting approach called the noGAN approach for super-resolution. This process is a sort of pretraining for GANs, where high-quality images are converted to lower quality images for training data of the Generator, and the critic is pre-trained on the generated images. This way, both the generator and critic have a head start, and this method is found to significantly improve training time for GANs.,GANs are a neural network pair that are trained through an adversarial process. The 2 parts of a GAN are a Generator and a Critic/Discriminator. The role of the generator is to generate high-quality data that is similar to training data, and the role of the critic is to differentiate between the generated data and the real data. The objective function of the generator is to maximise the loss of the critic, and the objective function of the critic is to minimize its loss.,Data scientists do lots of exploration and experimentation. Jupyter Notebook (Notebook from here onwards) is a great tool for exploring and experimenting. However, things can get cluttered and messy quickly when using Notebook. Keeping your workflow clean, organised and easy-to-understand is an important skill and will serve you well in your professional career. In this post, I will share a few of my tips on how to keep your Notebook and workflow more organised.,The Synthetic Data Vault (SDV) was first introduced in the paper “The Synthetic data vault”, then used in the context of generative modeling in the master thesis “The Synthetic Data Vault: Generative Modeling for Relational Databases” by Neha Patki. Finally, the SDV library was developed as a part of Andrew Montanez’s master thesis “SDV: An Open Source Library for Synthetic Data Generation”. Another master thesis to add new features to SDV was done by Lei Xu (“Synthesizing Tabular Data using conditional GAN”).,However, in segmentation tasks, we want the output to be the same shape as the input image and the added features for labelling pixels. So the downsampling of a traditional Conv architecture is supplemented by an upsampling path, to add back the height and width of the image to the output, while maintaining the features. There are many upsampling methods, but the most common one used in most libraries is Transpose convolution upsampling. You can read about this method here.,YOLO was first introduced by Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi in their paper “You Only Look Once: Unified, Real-Time Object Detection” which you can read here. The paper was proposed as a fast, state of the art model for object detection in 2015. Over the years, YOLO has had 4 official versions (where papers were published). The first 3 were by the original authors and the last one was by a different author. I will not discuss the versions of YOLO now, maybe in another post ;-),Think of this process as analogous to a thief and the police. The thieves want to fool the police and keep improving their tools and techniques, and the police want to catch the thieves so they improve too. The generator is like the thief and the critic is like the police.,Deep fakes I’m sure everybody has heard of deep fakes from the media. Deep fakes are also GANs where the generator is trained to perform the faking operation, and the critic is tasked with detecting the fake. The generator can be trained for long enough to fool most humans. This is a somewhat dangerous technology and something to be aware of on the internet.