The artificial intelligence revolution can be both scary and exciting. Despite this, we’ve always felt safe in the assumption that there are certain jobs and tasks that only humans can do. But nothing brings this more into question than the achievements of Google’s DeepMind AI, which seems to be accomplishing new and unprecedented things each […] The artificial intelligence revolution can be both scary and exciting. Despite this, we’ve always
The artificial intelligence revolution can be both scary and exciting. Despite this, we’ve always felt safe in the assumption that there are certain jobs and tasks that only humans can do.
But nothing brings this more into question than the achievements of Google’s DeepMind AI, which seems to be accomplishing new and unprecedented things each day. Let’s take a look at some of the things that this advanced AI has achieved that many of us never saw coming.
DeepMind is a subsidiary of Google that focuses on the development of artificial intelligence and deep reinforcement machine learning. (What is artificial intelligence?
What Artificial Intelligence Isn’t
What Artificial Intelligence Isn’t Are intelligent, sentient robots going to take over the world? Not today — and maybe not ever.
Read More) While the company’s AI caught the world’s attention when its AlphaGo AI beat Go’s professional (human) world champion
The AIs Are Winning: 5 Times When Computers Beat Humans
The AIs Are Winning: 5 Times When Computers Beat Humans Artificial intelligence is getting good. In fact, computers are now beating the best and brightest minds that humanity can offer. What does that mean for us?
Read More in 2015, its aims and goals go far past beating humans at board games.
The deep reinforcement learning of its AI algorithms has been used in both research and applied contexts. Each year, new advancements are made thanks to the company’s AI developing increasingly complex abilities.
So far, the AI from DeepMind has found purposes within the medical industry, Google’s Android business, and other general AI experiments by Google. Here are five of its notable achievements.
One of the most enduring visuals to come from DeepMind’s AI achievements a video depicting a DeepMind neural network learning how to walk.
The AI was given parameters for torque-controlled virtual bodies, including the bodies’ number of joints, the degree of freedom of limbs, and obstacles it needed to traverse in a virtual environment. These environments were procedurally generated with a variety of obstacle types, such as hurdles or gaps.
Without being taught how to get past these obstacles, the AI needed to learn from scratch how to move and maneuver through its world. With this limited information, the AI taught itself how to walk in various bodies—including a humanoid, bipedal, and four-legged body.
Not only did it learn to walk and run, but it also could successfully tackle obstacles in its virtual environments—such as jumping over gaps and climbing ledges. This was also incredibly entertaining to watch since it came up with some inventive uses of its limbs.
Images generated by Google’s DeepMind AI (Image Credit: DeepMind)
An intriguing capability developed by DeepMind’s AI is the ability to create its own original, realistic images from nothing. To do this, researchers used ImageNet as a database to provide real-world sample images for the AI to learn from.
The neural network was then trained to not only generate images based off of what it learned from the data, but it could also distinguish generated images from real-world images.
The algorithm uses Generative Adversarial Networks (GAN), a type of AI algorithm that has existed for some time. But what makes the DeepMind AI’s image generation unique is how much it has improved upon and optimized the technology. In terms of the quality metrics used to assess generated images, the samples created by DeepMind’s AI outperformed other attempts by a significant margin.
Nothing makes images of SkyNet come to mind as much as realizing that DeepMind AI has already learned how to strategically out-think human opponents. You may have heard about DeepMind AI beating human opponents at board games, but it now knows how to work in a team.
DeepMind AI has figured out how to beat humans in Quake III Arena’s Capture The Flag matches. Its teamwork capabilities aren’t even limited to other AI—the bot was even able to work with human teams to beat opponents in the game.
“Through new developments in reinforcement learning, our agents have achieved human-level performance in Quake III Arena Capture the Flag, a complex multi-agent environment and one of the canonical 3D first-person multiplayer games,” DeepMind said in their announcement. “These agents demonstrate the ability to team up with both artificial agents and human players.”
The bots had to learn from scratch how to see and act in these procedurally generated (and therefore unseen) environments. They did this without even knowing the rules of the game. They then had to learn how to also cooperate and compete in order to win. Researchers appropriately named the AI “For The Win (FTW) agents”. This AI took part in a tournament with 40 human players.
Researchers even reduced the bots’ accuracy and reaction to lower their performance. Despite this, they learned human-like behavior such as camping bases and following teammates in order to win.
The performance of agents during DeepMind AI training in Quake III Arena. Credit: DeepMind
When measuring their Elo rating, a metric used to score the skill of players in zero-sum games, the FTW agents overtook the average score of regular players and players during the course of its training.
One of the most impressive achievements from DeepMind’s AI is its ability to navigate through a city without a map. The AI relied on learning from experience instead. This is a relatively simple task that humans perform all the time. But the underlying mental mechanisms that enable us to do this are very complex.
DeepMind AI had to navigate through major cities and reach a specific location without a map. The AI moved through its virtual environment with a first-person view derived from Google Street View images.
Over time, the AI memorized different routes and ways to get places as if it was a resident city-slicker itself.
“The agent is rewarded when it reaches a target destination (specified, for instance, as a pair of latitude and longitude coordinates), like a courier tasked with an endless set of deliveries but without a map,” DeepMind said in their statement on the project.
Over time, the AI learned how to navigate new cities. It was even able to apply what it learned to new cities. For example, after learning how intersections worked, it would use this knowledge for future cities.
The company’s AI has also previously learned how to successfully navigate 3D mazes, something that many of us humans still fail to do.
