Unlike the human brain, which can handle multiple tasks at once, computers must “think” linearly to achieve intelligence. Despite this limitation, there are many areas in which AI has already advanced beyond human intelligence. With technologies such as deep neural networks, machines have learned how to talk, drive cars, win video games, paint pictures, and assist in making scientific discoveries.
In this blog, we will look at the six areas where artificial neural networks have proven that they can go above and beyond the limits of human intelligence.
Machine intelligence has a good track record of image and object recognition. The capsule networks created by Geoffrey Hinton have almost halved the best previous error rate on a test that challenges software to recognize toys such as trucks and cars from different angles. Even if the angle of view is different from the previously analyzed views, these capsules use generalization of objects in a geometric space to allow the system to better identify objects while also requiring fewer images to do so.
Another example comes from a state-of-the-art network which has been trained to mark images in a database such that it can classify them better than a doctor with over 100 hours of training hours on the same task.
You may have heard of IBM's Deep Blue and DeepMind's AlphaGo, both receiving global attention by beating world champions in chess and Go, respectively. But did you know that AI is also well adapted to video games?
Researchers have used deep learning to teach computers to play games such as Atari’s Breakout. The researchers in this experiment did not teach or pre-program the computers to play the games in a specific way. Instead, the computer is given control of the keyboard while it keeps track of the score. The computer will then learn autonomously, with the goal of maximizing the score. After playing only for two hours, the computer became an expert at the game.
The deep learning community is racing to train computers to beat humans at almost every game imaginable. This includes games such as Space Invaders, Doom, and World of Warcraft. With most of these games, the deep learning network has surpassed even the most experienced of players. Computers are not initially programmed to play these games; they learn them on the go by playing the game.
Last year, Google released WaveNet and Baidu launched Deep Speech. Both are deep learning networks that automatically generate human voice. The system learns to imitate human voices and, over time, improves its own ability to imitate them. It has grown increasingly difficult to distinguish their words from the speech of a real human.
LipNet—a deep network created by Oxford University with funding from Alphabet’s DeepMind— has achieved a 93% success in reading people's lips. The best of human lip readers have only a 52% success rate. A team at the University of Washington used lip sync to create a system that adds synthetic audio to an existing video.
While the previous three areas may not come as a surprise, AI has also been making significant progress in the field of arts. You can use neural networks to study a given piece of art’s strokes, colors, and shadows. You can create a new image based on the original style of the artist, or even recreating a piece with a different style.
For example, Deepart.io is an example of a company-created application that has used deep learning techniques to learn hundreds of distinctive styles. You can apply these styles to photos. Artist and programmer Gene Kogan has also used stylistic transformations to modify the Mona Lisa based on algorithmic styles learned from Egyptian hieroglyphics.
Timnit Gebro, a researcher at Stanford University, took 50 million photos from Google Street View to explore the ability of a deep learning network. The computer quickly learned to locally identify cars. Moreover, it individually identified more than 22 million vehicles including their manufactures, styles, models and years. One example of the applications this system has is figuring out the beginning and end of voter routes. According to the analysis provided, “if the number of sedans seen in a 15-minute drive exceeds the number of pickup trucks seen, the city has an 88% probability of voting for Democrats in the next presidential election.”
Another example of a machine intelligence that provides far more accurate predictions than humans would be Google’s Project SunRoof. The technology uses aerial photographs from Google Earth to create a 3D model of the roof and to distinguish it from the surrounding trees and shadows. It then uses the sun's trajectory to predict how much energy the solar panel can generate from this roof according to its position and specifications.
When it comes to website design, user behavior analysis is one of the key elements in providing optimal user experience. One can use the integration of artificial intelligence into the building of websites to efficiently modify the site and may even be more accurate than work done by human designers. The underlying technology of a system like this provides an average user’s opinion of the site's appearance. This allows the designers to determine whether the site is well designed or not. Today, web designers may be using a deep network to modify their designs, or they may be planning to use deep networks in the very near future.
With the endless possibilities of AI, it is no surprise that AI is perceived as a threat to humanity. However, the advancements in AI can also be beneficial in helping us making new scientific discoveries and improving the quality of life for the society. Applications such as Alibaba Cloud's ET Brain are aimed at tackling the most solving complex business and social problems. It is undeniable that AI is powerful, but ultimately, it all comes down to how we design and use it.
In this blog, we looked at six distinct aspects, where artificial neural network shave surpassed human intelligence. From speech generation to website modification, artificial neural networks have shown its possibilities. I firmly believe that this is just the tip of the iceberg to its vast capabilities.
Original article: https://mp.weixin.qq.com/s/DehAUE2uxBSjoPaFl4jFVQ
Alibaba Clouder - November 19, 2018
Alibaba Clouder - December 4, 2018
Alibaba Clouder - September 14, 2017
Alibaba System Software - December 25, 2018
Alibaba Clouder - January 24, 2017
Alibaba Clouder - September 30, 2017
An end-to-end platform that provides various machine learning algorithms to meet your data mining and analysis requirements.Learn More
Relying on Alibaba's leading natural language processing and deep learning technology.Learn More
A secure solution to migrate TB-level or PB-level data to Alibaba Cloud.Learn More
A premium, serverless, and interactive analytics serviceLearn More
More Posts by Alibaba Clouder