14 min read
Artificial Intelligence is a central technology present in every industry and company. Research in machine learning and AI is very extensive and impossible for anyone to read through it all. Perceptron (previously known as DeepMind) collects some recent discoveries, in Artificial Intelligence, and seeks to explain why these matter.
A team of engineers at the University of Glasgow have developed an ‘artificial skin’ that has the ability to not only experience but react to simulated pain. Researchers at DeepMind have created a machine learning system that anticipates and predicts where soccer players will be present while running on the field, which groups present at Tsinghua University, and The Chinese University of Hong Kong (CUHK) developed and created algorithms that are capable of generating realistic photos, and even videos, of human models.
The Glasgow team released a press statement that stated that their artificial skin utilized a novel processing system that was based on “synaptic transistors” designed in a way to mimic the neural pathways of the brain. These transistors are made using zinc-oxide nanowires that are printed onto the surface of a flexible plastic, which is then connected to a skin sensor responsible for registering changes in electrical resistance.
While artificial skin has been created and utilized before, the team stated that their design differed from previous attempts because it used a circuit built into the system to function as an ‘artificial synapse’ with the intention of reducing input to a voltage spike. This allowed the processing to be sped up by establishing an input threshold for the voltage which had a varied frequency based on the pressure applied to the skin.
The team sees an advantage of the skin being utilized in robotics. For instance, it can help prevent a robotic arm from coming into contact with high temperatures.
The team sees the skin being used in robotics, where it could, for example, prevent a robotic arm from coming into contact with dangerously high temperatures.
DeepMind states that they have developed an Artificial Intelligence model by the name of Graph Imputer. Graph Imputer can predict where soccer players will move on a pitch using the camera recordings of a subset of players. What is even more impressive is that it can make predictions beyond the view of the camera, which allows it to track the position of most, if not all, players on the field of play with good accuracy.
While the Graph Imputer is not perfect, DeppMind researchers state that it can be possibly used for other applications such as modeling pitch control, or also the possibility of a player controlling the ball assuming it is at a given position. (Several teams from the English Premier League of football use pitch control models not only during the game but in their pre-match and post-match analysis as well). Beyond football or other sports for that matter, DeepMind believes that the Graph Imputer techniques can be utilized in other domains such as pedestrian modeling on roads or crowd modeling in stadiums.
Where on one hand artificial skin and movement predicting systems continue to be impressive, photo and video generating systems are also progressing at a fast rate. Some high-profile works are OpenAI’s Dall-E 2 and Google’s Imagen. Let us take a look at Text2Human which was developed by CUHK’s Multimedia Lab. It can translate a description “the day wears a short-sleeve T-Shirt with pure color pattern and a short and denim skirt” into a picture of a person who does not actually exist.
Tsinghua University, partnering with the Beijing Academy of Artificial Intelligence, has created a model called CogVideo which is capable of generating video clips for text (for example, a horse drinking water). These particular clips may be rife with visual weirdness and other artifacts, but considering these are the work of completely fictional scenes, it would be too harsh of a criticism to make.
Drug discovery has employed the use of Machine Learning. In this, a near-infinite variety of molecules characterized in theory and literature are to be sorted through and categorized in order to find potentially beneficial effects. However, the volume of data is so large and the cost of false positives so high that even a 99% accuracy is not good enough. This is especially the case with unlabeled molecular data which constitutes a bulk of the data that is out there in comparison with molecules that have manually been studied over the years.
There has been a constant effort to create a model, by the CMU researchers, in order to sort through the billions of uncharacterized molecules, which was trained to make sense of these molecules without any extra information. This was achieved by making subtle changes to the molecule’s structure, like hiding an atom or perhaps, removing a bond, and then consequently observing how the molecule changes. This in turn allows to model to learn the intrinsic properties of how such molecules form and then behave. This led to the model outperforming other AI models in recognizing toxic chemicals in a test database.
Molecular signatures are also useful when diagnosing diseases. For example, two patients may present similar symptoms, but upon careful examination, it can be noticed that they have two completely separate conditions. While it may constitute standard doctoring practice, data from multiple tests and analyses pile upon and make it increasingly difficult to track all correlations. A form of clinical meta-algorithm is under work by the team at the Technical University of Munich which aims at integrating multiple data sources, and algorithms, to make a distinction between specific liver diseases and similar presentations. These models will not serve to replace doctors, though they can help to assist in approaching increasingly growing volumes of data that experts may not have the time or resources to interpret.
Real-Life Applications of AI....
14 min read