Perceptron: Pain-sensing robots and artificial intelligence that anticipates soccer players' moves

News Sand DC
Source: Getty Images

Machine learning and AI research, which is now a critical technology in almost every sector and corporation, is just too extensive for anybody to study all. Perceptron (formerly Deep Science) is a column that seeks to gather and explain some of the most important recent findings and articles — notably in, but not limited to, artificial intelligence — and explain why they matter.

A team of engineers from the University of Glasgow produced "artificial skin" that can learn to feel and react to simulated pain this week in AI. Researchers from DeepMind built a machine learning system that predicts where soccer players would run on a field, while groups from Tsinghua University and The Chinese University of Hong Kong (CUHK) produced algorithms that can generate realistic images — and even films — of human models.

The Glasgow team's artificial skin, according to a news release, used a novel form of processing system based on "synaptic transistors" that mimicked the brain's neural networks. The transistors, which were created from zinc-oxide nanowires printed on the surface of flexible plastic, were linked to an electrical resistance sensor on the skin.

Source: University of Glasgow

While artificial skin has been attempted previously, the team says that their concept is unique in that it uses a built-in circuit to operate as an "artificial synapse," reducing input to a voltage spike. By selecting a threshold of input voltage whose frequency varied according to the degree of pressure applied to the skin, the team was able to "teach" the skin how to respond to simulated pain and "teach" the skin how to respond to simulated pain.

The skin might be employed in robotics to protect a robotic arm from coming into touch with dangerously high temperatures, according to the researchers.

DeepMind claims to have built an AI model called Graph Imputer that can predict where soccer players will go based on camera records of only a fraction of players, which is tangentially connected to robotics. More impressively, the system can generate predictions about players outside of the camera's vision, allowing it to pretty accurately monitor the position of the majority — if not all — players on the field.

Source: DeepMind

Graph Imputer isn't without flaws. However, the DeepMind researchers suggest it might be used for things like simulating pitch control or determining the likelihood that a player would be able to control the ball if it is in a specific spot. (Pitch control models are used by several top Premier League clubs during games, as well as in pre-match and post-match analysis.) DeepMind believes the principles underpinning Graph Imputer to be relevant to areas such as pedestrian modeling on roadways and crowd modelling in stadiums, in addition to football and other sports analytics.

To be sure, artificial skin and movement-predicting systems are amazing, but photo- and video-generating technologies are moving at a breakneck speed. There are obviously notable works like OpenAI's Dall-E 2 and Google's Imagen. Take a look at Text2Human, a programme developed by CUHK's Multimedia Lab that can turn a caption like "the lady wears a short-sleeve T-shirt with a pure colour pattern, and a short and denim skirt" into an image of a non-existent human.

Tsinghua University, in collaboration with the Beijing Academy of Artificial Intelligence, developed an even more ambitious model called CogVideo, which can produce video clips from the text (for example, "a guy skiing," "a lion sipping water"). The video is riddled with artefacts and other visual oddities, but given that they are entirely imaginary situations, it's difficult to be too critical.

The near-infinite diversity of compounds that occur in literature and theory must be sifted through and analysed in order to detect potentially positive effects, thus machine learning is frequently utilised in drug development. However, the volume of data is so huge, and the cost of false positives might be so high (chasing leads is expensive and time-consuming), that even 99 per cent accuracy isn't good enough. This is especially true with unlabeled molecular data, which makes up the vast majority of what's available (compared with molecules that have been manually studied over the years).

Source: CMU

CMU researchers have been training a model to sort through billions of uncharacterized molecules without any further information in order to construct a model that can filter through billions of uncharacterized molecules. It accomplishes this by making little structural modifications to the (virtual) molecule, such as concealing an atom or deleting a link, and then monitoring how the molecule changes as a consequence. This allows it to understand inherent aspects of how such molecules are produced and behave, which helped it outperform other AI models in a test database for identifying dangerous substances.

Molecular markers are also important in disease diagnosis; two individuals may have identical symptoms, but a close examination of their lab findings reveals that they have quite distinct illnesses. Of course, this is a typical medical procedure, but when data from various tests and analysis accumulates, keeping track of all the relationships becomes challenging. The Technical University of Munich is developing a clinical meta-algorithm that combines several data sources (along with other algorithms) to distinguish between liver illnesses that have similar symptoms. While such models will not be able to take the position of physicians, they will continue to assist in the management of the enormous amounts of data that even experts may not have the time or knowledge to understand.

#buttons=(Accept !) #days=(20)

Our website uses cookies to enhance your experience. Learn More
Accept !