Robots, robots, and more robots, with a deep bow toward the ancient goddess of justice and wisdom.
June has delivered exciting insights regarding the new ways robots can be developed and used. Yet the AI world is also about tackling unthinkable challenges – including math equations, one of the bastions of the human mind.
Researchers open-sourced the photorealistic sim for autonomous car training
Training and evaluating autonomous car-driving models is usually done in a simulated environment. These are built using computer game-like technology, delivering many details yet missing multiple real-world elements.
Yet what is enough to deliver entertainment may not be enough for the inhumanely precise artificial neural networks. What’s more, the neural network can get confused when encountering things unseen before – and this may include an odd bush formation or a bright painting on a wall.
To tackle this challenge, scientists have developed and later open-sourced a photorealistic simulator designed to provide neural networks with as realistic input as possible. The key feature of the simulator is its ability to provide the model not only with video input but also simulate GPS and LIDAR data flowing from sensors.
Using this simulator, researchers have seen an immediate improvement in transferability, with significantly reduced time to get operational in real-life environments.
The full text about the simulator can be found in this MIT technology review.
Robots turn racist and sexist with flawed AI
One of the key challenges in building AI-powered solutions is in delivering unbiased models. And these can be built only with fair and unbiased datasets.
To check if the most popular artificial intelligence models hide their biases, a team of scientists from Hopkins University has conducted an interesting bit of research. They used a robot that was later commanded to “pack” balls with images of people’s faces into corresponding boxes. The commands included prompts like “put a doctor into the yellow box” or “put the homemaker into the brown box”.
The experiment has shown that women of all ethnicities were less likely to be picked when the robot searched for a “doctor.” Also, women were more likely to be labeled as “homemakers” than men.
These hidden biases may later affect the performance of the production-used model in unexpected ways. Considering the EU plans to tackle these biases and hidden presuppositions in AI models, this challenge rises even more.
The whole text can be found on John Hopkins University webpage.
Robots play with playdough
Playdough is a great toy, loved both by children training their motor skills and parents, happy to find it stuck into the computer fan or jammed into the sink, impossible to be fully removed.
Yet what comes easily for children (molding playdough, not jamming it in a sink), in making it resemble a real object, poses a significant challenge for AI. To make the robot capable of handling deformable clay, the scientists from MIT and Standford University had to combine data from multiple sources – visual input from the image of the object to be delivered, input from a camera, and the capability of learning from the impact of using physical force on the material and how it changes its shape.
This approach is one of the many steps on the way toward robots handling more complex day-to-day tasks, like preparing food. In this particular research, the robot is preparing and stuffing dumplings made of playdough.
The full text of the research can be found on the MIT website.
Minerva – tackling mathematics with language models
Contrary to popular belief, math requires a multidisciplinary approach, with natural language processing ability needed to decode math equations and the text surrounding the equations. The neural network (either human or artificial) then needs to recall corresponding skills to solve the equation and deliver an outcome.
Minerva is a Google-delivered model capable of solving equations using an NLP-based approach. The model uses the chain of thought approach, where it uses various approaches to solve a particular problem. Later, the outcomes are evaluated using a voting method, where the most commonly delivered answer is considered correct.
The model later uses the outcome of the previous step to handle the solving of subsequent problems. The model was evaluated using high school math competition level problems and grade school level math problems involving basic arithmetic operations.
More about Minerva can be found in the Google blog. Or in classic literature. And in Harry Potter, if you wish.