Al dives deep into boiling water
July has shown us that artificial neural networks can deliver impressive results when it comes to breaking challenges deemed unsolvable, and yet struggle with children’s games. This edition offers a glimpse into the social context of AI at work.
Computer and video games, excluding the earliest examples like Pong, have always utilized some form of artificial intelligence – usually in controlling foes, from Pac-Man’s ghosts to sophisticated masterminds in strategy games. Hide and seek is a new challenge in which AI shines – not to mention protein folding prediction.
AI succeeds in Hide and Seek and Capture the Flag
Following its triumphs in chess and go, artificial intelligence continues to build its presence in a number of games that were thought to be human only. The latest example comes from Deepmind’s research that examines the performance of reinforcement learning agents in hide and seek, capture the flag, and similar children’s games.
While the game itself appears to be child’s play, for artificial intelligence the task is highly complicated. It consists of changing one’s environment (procedurally generated in an XLand simulator) and while the goal remains constant, there are countless ways to achieve it.
What’s even more interesting, the paper presents an agent that is capable of playing multiple games in a changing environment. This is significant progress when compared to models like AlphaGo or AlphaStar, which achieved godlike performance in one game while being entirely incapable of playing any other.
In other words – while AlphaStar is a killer player when it comes to Starcraft, it would completely fail while tasked to play Age of Empires – while a human RTS-lover would take no time to switch from one to another.
More details on generally capable agents researched via open-ended training can be found in deepmind’s blog post.
Employees actually want more AI
AI is frequently shown as either a threat to employment or as the most significant of historical discoveries, one which will liberate human creativity and flexibility by taking over dull and repetitive tasks.
According to a SnapLogic study conducted by 3GEM, 89% of surveyed employees believe that AI could support their daily duties and support their productivity in up to half of their work. Also, 61% say that the biggest advantage of AI in the workplace would be increased productivity and 49% say it would improve decision-making processes.
Last but not least, more than half of surveyed employees believe that using AI will help them to achieve more work-life balance.
The study can be found on the snaplogic website.
Deepmind’s AlphaFold 2 goes open source
Protein folding used to be one of the most complex tasks, on the verge of the unsolvable. Predicting the way a protein will fold, especially considering the complicated and sophisticated shapes it can take, has long haunted the minds of chemists. And Deepmind’s AlphaFold has fielded the first effective attempt to truly solve the challenge.
The company decided to make the solution open-source, which comes with various real-life applications. For example, the software was used to identify the shape of the COVID-19 ORF3a protein, and by doing so helped us to understand the way the virus is spreading.
In the long term, the technology is predicted to be used to find new drugs and treatments for diseases. More on the topic can be found in Nature’s publication.
Neural networks prothesis enables paralyzed man to speak again
Breakthroughs in medicine are frequently cited as a core advantage of AI development. The ability to process gargantuan amounts of data and spot patterns can lead to breathtaking results.
A “speech neuroprosthesis” is a great example of such. By using neural networks that read signals from the brain and translate them into words that appear on a screen, the researchers from UC San Francisco were able to literally give speech back to their paralyzed patient.
The solution uses the signals in the brain that are responsible for controlling the muscles responsible for speaking. By that, the neural network delivers words in a much more effective way than typing usually does – especially for patients with movement disabilities.
The full research can be found on ScienceDaily.
AI dives deep into boiling water
Boiling water is not only the key element of tea, coffee, or numerous other beverages. The process of boiling is crucial when it comes to cooling systems, including those used in computer chips and nuclear reactors as well.
While cooling with liquids is generally a good idea, the problem comes with a “boiling crisis” – a situation where a cooled surface gets so hot that bubbles formed during the boiling process meld together to create an isolating layer which makes the cooling process ineffective and raises the temperature even further.
Researchers from MIT designed a system that uses high-speed infrared cameras and neural networks which can track up to 17 factors in an effort to determine how close the fluid is to a “boiling crisis.” By doing so, the safety of every cooling device, from microprocessors to nuclear reactors, can be significantly increased.
More about the research can be found on MIT’s News website.
