Machine learning (ML) is an amazing field that enables a huge number of powerful and interesting techniques. ML is a broad field that has applications in many areas. From image processing to conservation, ML provides unique solutions to problems old and new. Here are some interesting and cool applications of machine learning.
1. Seeing Through a Computer’s Eyes: Interpreting Neural Networks
Neural networks (NNs) and deep neural networks (DNNs) are very popular machine learning techniques. This type of modeling is used in many of the best-known applications of ML. Image classification, face identification, and speech recognition are just a few examples.
DNNs are inspired by animal brains. They are made up of “artificial neurons” that form complex connections (weights) with each other. The internal structure of these networks is incredibly complex and is not programmed by a human. Instead, DNNs are trained on examples, adjusting their weights to produce the desired output. Therefore, neural networks work as a sort of “black box.”
This means it is difficult, if not impossible, to understand exactly WHY each model produces good results. Researchers are now trying to peek inside the “black box” of neural networks. This field is called “interpretability” and has produced some strange images.
The earliest examples of this are the Google DeepDream images. The DeepDream framework visualizes the way a neural network “sees” images. This results in some wild, psychedelic images.
Since the release of DeepDream, researchers at Google have continued to try and show us what machines see. A team at Google’s research lab has created an interactive website that does just this. Try it out for yourself to get an idea of what and how a neural network sees – a firsthand look at a machine learning application!
2. Style Transfer and Neural Image Analogies
Neural networks can do far more than just see, they can also create! The idea of using machines to generate images may sound odd, but it creates some incredible results. These techniques allow for a few different results. Image analogies can be used to rearrange parts of an image, maintaining the overall appearance and patterns.
Style transfer is a related technique that allows the “style” of one image to be applied to another. This allows you to imitate an artistic style in images. It can also be used to make a fall landscape into a winter landscape.
These techniques produce amazingly good results. Some of these images would be difficult to identify as being made by a machine. They can also produce beautiful and unique images. These are certainly very interesting applications of ML.
3. Colorizing Images
Old black and white images and films have been tediously colorized by hand for decades. Computers have provided some assistance with tools that simplify the process. However, artists have still needed to draw the outlines of each color, select the correct shades and adjust these shapes as the objects move. This is a tedious and slow process.
Colorization is another machine learning application that simplifies people’s lives. Neural networks (NNs) are being used to colorize images with great results. This type of technology uses a Generative Adversarial Network (GAN) to produce results. There are a number of different GitHub repositories with open-source code allowing anyone to try it out!
Newer techniques are combining human intelligence with this technique to allow users to select the correct colors for different areas. This is a human-in-the-loop strategy, which is an essential way to maximize the effectiveness of ML.
These techniques even work for videos as shown in the examples below!
4. Zoom and Enhance!
If you have ever watched crime TV shows like ‘Law and Order: SVU,’ you’ve seen the detectives point at a license plate pixelated beyond all recognition and shout “enhance.” In the magic of Hollywood, they are suddenly able to read the plates and bring the criminal to justice. This is so ridiculous that “zoom and enhance” has become a meme.
It is certainly true that you cannot find detail in an image that isn’t in the original. However, ML techniques do allow you to increase the resolution of an image! This does allow the user to “zoom and enhance” well beyond the normal limitations of an image.
This technique uses machine learning to detect edges, guess at shapes, and give pixelated images a smooth and believable appearance. Any detail gained is just the machine’s best guess. However, its guesses look very good and can be used to improve the appearance of low-res photos.
5. Identifying Individual Zebra
ML has applications beyond tech companies. Wildlife conservation efforts have also benefitted from ML techniques. Dr. Tanya Y. Berger-Wolf of the University of Illinois Chicago has been using ML for a number of conservation projects.
These include a project that uses computer vision to identify individual animals. For example, the stripes of a zebra are unique to that zebra, but difficult for humans to recognize. Machines can identify which zebra is in each photo.
This is extremely important for ecological research. It allows researchers to track the behavioral changes of individual animals through time. This is important for many techniques including estimating population size.
6. Image Synthesis
The ability for machine learning to generate images may seem odd. Though this task is not like a classic learning task, it is similar. By training a Generative Adversarial Network (GAN) on examples, the machine can learn the common features of examples.
This allows the computer to generate new images that fit the category. One of the most impressive examples of this technique produces unique human faces. Check it out here at thispersondoesnotexist.com. It may be hard to believe, but these images are not real people. Keep reloading the page, and you will never see the same face twice.
7. Computer Security
ML can be used to defend computers and networks from attack. Cybersecurity is a constantly evolving field. New threats appear every day. The software and professionals can have a hard time keeping up. ML can provide powerful tools to help keep computers secure.
ML is good at finding trends in labeled data. By training an ML algorithm on data about attack vectors, it can learn to identify them. This can make ML very good at tasks like spam detection and filtering. With this task, there is a huge amount of data available.
However, there are still extensive limitations to the use of AI and ML in security. Some tasks do not have enough training data available. Others simply detect “anomalies,” which may not represent real threats.
Creativity is the area where humans have the clearest advantage over machines. However, some researchers are working to change this. Art created by machine learning is still generally off the mark to a comical degree. However, ML can produce interesting creative works. Here are a few examples.
ML has been used to write poetry! The result is….. underwhelming. It can produce some pretty verse but lacks the cohesion and meaning of real poetry. But who am I to tell a computer how to express itself…. Maybe I just don’t get how the computer feels.
AI is also being used to produce written content. Though it is still in the early stages, this has the potential to have huge consequences. The creation of automated fake news is just one scary possibility. ML is good at creating interactive chatbots that seem human but are not as good at writing full stories.
OpenAI has created a project called MuseNet that is capable of producing music in a number of different styles. It is trained on MIDI files taken from a number of sources. The team that created it claims it has the ability to understand music as a composition. It can create movements and melodies like a real song.
You can try out MuseNet for yourself, creating unique compositions based on different starting conditions.
The idea of AI playing a song has been with us for a long time. In the classic film, “2001: A Space Odyssey” HAL, the ship’s AI pilot, uses music to demonstrate his humanity.
9. Machine Translation
Google translate has been a mainstay of any Intro to Spanish course since 2004. However, in 2017, Google made a huge change. They switched Google Translate from a rule-based program to a neural machine translator.
This change has led to far better translations overall. Google has trained their tool in order to translate books into many languages. This has led to good translations of many phrases. However, it has also had some odd consequences.
Some of the most heavily translated books in the world are religious texts. Because of this, Google Translate has made some oddly solemn translations. It has spoken of the doomsday clock and other strange topics. Read some of the funniest (and spookiest) examples on the subreddit r/TranslateGate.
10. Generating Inspirational Quotes
ML has the power to inspire! Sort of……
The bot’s ability to match the style, tone, and vocabulary of inspirational quotes is impressive. However, the insight it is trying to share is far less clear. But the absurdity of these quotes does a great job of uplifting you in another way.
Last but not least, ML is present in chatbots! By incorporating machine learning into chatbots, swivl is able to tag data, making customer service a seamless experience for both the customer and the business. The chatbot learns from its interactions and becomes more efficient the more people ask it questions.
To learn more about how swivl can help you apply machine learning to your business, schedule a demo and meet our Intelligent Agent, Hoover.
Please join our newsletter to learn about exciting new machine learning applications.