July was dominated by OpenAI and Deepmind with two outstanding AI systems – DALL-E and AlphaFold, both challenging the status quo in their fields.
Another pair of cherry-picked bits of research from July were about robots and their cooperation with other robots, or with humans, by either sharing goals or following the example of a human guide.
AlphaFold gradually reveals structures of all known protein
Proteins are basically the bricks that all living things are made of. The relationship between the protein encoding in amino acids and their 3D structure remains a challenge for scientists. To tackle it Deepmind, the Google-backed AI research lab, has delivered and open sourced AlphaFold, an AI-based protein-folding tool.
A protein’s structure is closely linked to its function. Able to predict these structures from amino acid sequences, the tool has been widely used in various scientific projects, including research on antibiotic resistance and plastic pollution.
In June, DeepMind shared that AlphaFold DB, a database containing information about the proteins mapped by AlphaFold, has cataloged over 200 million structures.
More about the tool and its potential applications as well as how to use the gathered data can be found in this Deepmind blog post.
Training robots to cooperate – without communication
Cooperation with open communication lines is (or at least should be) pretty straightforward. The less reliable and effective communication is, the harder it is to cooperate. And with communication completely blocked? Researchers of the University of Illinois Urbana-Champaign proved that it is also possible.
The agents trained using reinforcement learning were unable to communicate with each other by any means. To make the challenge even harder, the roles of each agent in the team also remained unobvious.
The developed ML-based technique allowed an agent to identify whether another agent was contributing to the overall goal of the team. The approach is comparable to reviewing a team of football players in which one particular player gets the goal, yet setting up that final kick is all about teamwork.
The algorithm can be useful in multiple situations, including in cooperation between autonomous vehicles using the same road or robots working in the same warehouse.
More can be found on the University page.
Robots learn household tasks by imitating humans
The idea of learning from observation is natural for humans, with children at first watching and then mimicking their parents while playing. Robots were largely incapable of doing so – but not entirely.
The research done at Carnegie Mellon University has shown that robots powered by neural networks can be capable of learning from watching videos with humans performing certain tasks and generalizing their observations later.
The method is called WHIRL, short for in-the-Wild-Human Imitating Robot Learning. The robots were trained in doing household chores – for example taking meals out of the refrigerator or taking a bag of trash out of the bin.
This approach is a big step toward automating the training of robots designed to work in environments used by humans to perform comparable tasks.
More about this research can be found on the Carnegie Mellon University website.
DALL-E now available in Beta
Last but not least – DALL-E is an OpenAI-delivered AI system that creates images from text prompts. The system is versatile enough to deliver multiple types of images – from photorealistic through mimicking certain artistic styles to complete abstractions. Also, it processes input no matter how weird it is.
The organization is increasing the availability of the system by inviting in the next batch of users who joined the waiting lists. These users will be able to use the system for free to generate a limited number of images and have the opportunity to pay an additional fee to increase the number of images.
Also, the delivered images will be usable for commercial purposes – for example as artwork, on mugs, or for clothing.
Join the waiting list and read more about the rules through the OpenAI blogpost.