Machine Learning / Deep learning is a subset of AI has been making a lot of waves recently. In previous videos and posts we’ve seen how deep neural nets are progressing rapidly within many fields. We’ve seen this technology and Machine Learning Applications diagnose medical conditions with more accuracy than trained experts.
AI or Artificial Intelligence is able to learn from numerous amounts of data, that humans can’t do. Since the Application of Artificial Intelligence the accuracy of early detection of esophageal cancer has reached 90%. AI has learned how to produce sound from scratch and recreate the visual aspects of speech.
Latest Machine Learning Applications
But what are the less known latest Machine Learning Applications of deep neural nets that could make their way down to the average consumer like us? In this post we’ll count down five interesting Machine Learning Applications of neural nets in the context of images and video.
Black and White Photos to Color
Normally when you see a black-and-white image of a time long past it looks dull distant and sometimes unrelatable. If you wanted to colorize that image and make it come to life it usually requires the effort of a keen artist who has a knowledgeable understanding of the cultural setting in order to choose the right colors.
What if there was a way to do this instantly. Well thanks to latest deep neural networks this is now possible.
“Let there be color” is a neural net system with latest Machine Learning Applications that automatically restores color to black and white photos. It can also do the same with videos. It uses a convolutional neural network and can process images of any resolution. Unlike previous approaches the Deep Learning Network actually learns patterns that occur naturally in photos; things like the sky is usually blue and clouds of white and gray and grass is typically green etc.
But it did this itself from past experience without human intervention. Let there be color, colors up the images much better than the state of the art for 100-year old images. Just take a look at some samples here.
Pixel Enhancing CSI Style
You know that thing that they do in movies and TV crime shows such as CSI, where an investigator demands that a grainy image be enhanced and somehow magically it becomes HD all that second. I know many people get annoyed with such an unrealistic portrayal of technology but now this kind of image enhancing is possible. Thanks to neural networks and machine learning.
Google Brain researchers have trained a deep learning neural network to take very low resolution images of faces and predict what those faces will most likely look like. They call the method Pixel Recursive Super Resolution and it enhances the resolution of photos dramatically.
In the image below you can see the original 8×8 photos and in the middle you can see the guess from the system. Obviously it’s not perfect and I don’t think it can be but it’s pretty unbelievable that this neural network can estimate so well the features of the person within a photo.
Generating New Images
Pix to Pix is a Machine Learning Application with deep learning neural network that generates new images based on input but this neural network was trained to perform multiple but specific tasks either create real street scenes from colored shapes or create a drawn map from an aerial photo or turn day scenes into night and finally make a photorealistic image just from an outline.
The Pics to Pics system is as to neural networks arranged in such a way that one network generates images and the other network is the judge of, if these images are real or not.
These two networks try and fool each other and improve each other in the process. This setup is known as a Generative Adversarial Network.
If you want to try out some of this stuff there’s a simple web app created for you to have a play with the edge outline part of this neural network
Lip – Reading
Lip net is a neural network developed by oxford university and google deepmind scientists. This network can watch a silent video of a person talking and convert the mouth movements directly to text.
Lip Net has raised 95% accuracy in reading people’s lips. An average lip reader has an accuracy of about 50 or 60 percent. So that’s not bad at all for Lip Net.
Creating a Scene from Scratch
A team of computer scientists from institutions such as the University of Wyoming and ubers AI department have created what they’re calling plug-and-play generative networks. This system is able to generate photorealistic images from a selection of a thousand categories.
I guess it’s early days for this stuff and some of them do look a little weird but it’s still cool to see the network try and figure out how to create a scene from scratch here’s a video of the neural Network running.
The goal is to generate the image that’s written in text below and each frame you see in this video is a different iterative step of trying to get there it actually kind of looks beautiful to look at.
Did you find this page useful? Help us to improve by rating this page.
[RICH_REVIEWS_FORM]
[RICH_REVIEWS_SNIPPET stars_only=”true”]
