DeepMind Technologies (a British artificial intelligence subsidiary of Alphabet Inc.) have released a paper which describes how an agent named Alpha Tensor has discovered a technique of matrix manipulation that has increased AI speed by 5% to 10%.
‘So what’ you may say. Well, for the first time since it’s inception, one of the most fundamental algorithms used in AI has been made more efficient and this was achieved by AI (not humans).
Data structures and algorithms have been developing since the invention of computer science and this has always fallen back on the application of one basic idea, matrix multiplication. It’s used in all facets of the scientific world including complex physics problems because all types of data can be represented as matrix images and also in machine learning.
So, if an algorithm that is this fundamental receives a 5% to 10% boost then it means that a large amount of ‘time can be saved’ as it looks for the shortest and most efficient way to achieve a goal or solve a problem.
The multiplying of two matrices has a basic problem and that is that there are literally an infinite number of possible ways to achieve this. Finding the quickest way is now the work of AI and it has already found a 5% to 10% increase, so one can imagine that as it keeps working at sorting the wheat from the chaff, then over time more increases are a real possibility.
This is definitely interesting for those of us who are unafraid of AI, but as a cautionary note, DeepMind Technologies were purchased by Google in 2014.