Recently, four researchers from Google invented a program called “AutoML-Zero”. It is a software that is capable of creating new artificial intelligence (AI) programs without human interventions by relying only on rudimentary high school mathematical concepts. Borrowing the concept of Darwinian evolution, researchers began by combining mathematical computations at random to develop a population of 100 candidate algorithms. They then tested each of them on a simple task like image recognition or differentiating between one item from the other.

Biological evolution in machines

In each round of testing, AutoML-Zero will compare the performance of these algorithms with other hand-designed algorithms. Copies of the best performing algorithms will be re-created with some of their codes either be substituted, revised or removed in an unsystematic way, so that their “copies” or “children” will not be entirely the same as them. These new algorithms will be added back to the population while the old ones are destroyed as the cycle repeats.

To facilitate the entire process, AutoML-Zero will generate thousands of such populations at a go. It will also interchange algorithms between populations to avoid coming to an evolutionary dead end and spontaneously wipe out any identical algorithms. All these fall perfectly into the four key points of Darwin’s Theory of Evolution and they are: traits are passed down from one generation to the next; no identical individual within a species; more offspring ensure a higher chance of survival and survival of the fittest.

It takes tremendous effort and time to create an algorithm. For example, in driverless car systems, developers may need several months to work out a sound neural network system that can read and identify various road signs. Although some scientists had tried to simplify the process by introducing no coding knowledge required or ready-to-use AI tools, helping novice to master the technology in no time. However, outputs from these off-the-shelf AI tools are often subjected to biases.

Uncover new AI capabilities

At the moment, AI algorithms created under such evolution remain simple, but researchers involved in the project are confident they can scale AutoML-Zero up to develop more complex AI. One suggestion given to the team was instead of creating AI from scratch, perhaps the first generation of AI can be seeded with basic knowledge or know-how that human have found earlier on, so that algorithms are primed with learned concepts and can evolve at a more rapid manner.

The researchers agreed. In fact, they are planning to increase the number of mathematical operations within AutoML-Zero and also to devote more computing resources to it. They hope that eventually, AutoML-Zero will be able to unleash new AI capabilities; something that may be fundamental but taking up a long time for human to figure out. Details of the project is made available in a preprint paper on arXiv.

*

Author Bio

Hazel Tang A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.