Artificial intelligence malaysia (AI) is advancing at a breakneck pace. Researchers have developed software that incorporates Darwinian evolution notions, such as “survival of the fittest,” in order to construct AI programmes that improve generation after generation without human intervention. The programme recreated decades of AI research in a matter of days, and its creators believe that it may one day lead to the discovery of novel methods to AI.
“While the majority of people made tiny steps, they took a massive leap into the unknown,” says Risto Miikkulainen, a computer scientist at the University of Texas at Austin who was not involved in the research. “This is one of those publications that has the potential to spark a great deal of future study.”
Machine Learning and Algorithm
It takes time to develop an AI algorithm. Consider neural networks, a type of machine learning that is frequently used for language translation and autonomous driving. These networks have a loose resemblance to the brain’s structure and learn from training data by varying the strength of connections between artificial neurons. Smaller subcircuits of neurons do specific tasks—for example, detecting road signs—and researchers can spend months figuring out how to connect them in such a way that they work in unison.
Scientists have accelerated the process in recent years by automating certain procedures. However, these programmes continue to rely on human-designed circuitry. This means that the product is still constrained by the engineers’ imaginations and preconceived biases.
Computer Science And AI
As a result, Quoc Le, a Google computer scientist, and colleagues built a tool called AutoML-Zero that allowed for the development of AI programmes with virtually no human input, utilising just elementary mathematical principles that a high school student would understand. “Our ultimate goal is to create unique machine learning concepts that scholars have been unable to discover,” he explains.
The programme discovers algorithms by approximating evolution loosely. It begins by randomly combining mathematical operations to generate a population of 100 candidate algorithms. It next tests them on a straightforward task, such as image recognition, in which they must determine whether a picture depicts a cat or a truck.
Each cycle, the programme compares the performance of the algorithms to that of custom-designed algorithms. Copies of the best performing algorithms are “mutate” by randomly altering, editing, or deleting portions of their code in order to generate subtle variants on the best algorithms. These “children” are incorporate into the population, whilst earlier programmes are phase out. The cycle continues.
The system generates thousands of these populations simultaneously, allowing it to process tens of thousands of algorithms every second until it discovers a good solution. Additionally, the programme employs optimization techniques such as periodically exchanging algorithms between populations to avoid evolutionary dead ends and automatically weeding out duplicate algorithms.
The researchers demonstrate in a preprint paper published last month on arXiv that the strategy may make errors using a variety of established machine learning approaches, including neural networks. While the answers are rudimentary in comparison to today’s most sophisticated algorithms, Le notes that the work serves as a proof of concept and that he is optimistic they can be scale up to produce considerably more complicated AIs.
Nonetheless, Joaquin Vanschoren, a computer scientist at Eindhoven University of Technology, believes the approach will take some time to catch up to the state-of-the-art. One way to improve the software, he argues, is to avoid having it start from zero and instead seed it with some of the tactics and approaches discovered by humans. “With learnt machine learning principles, we can prime the pump.”
That is something Le intends to address. Concentrating on tiny problems rather than complete algorithms, he adds, also has promise. On 6 April, his group released another paper on arXiv using a similar strategy to redesign a popular ready-made component that is utilise in a large number of neural networks.
However, Le believes that expanding the library’s mathematical operations and allocating additional computer resources to the programme could enable it to uncover totally new AI capabilities. “That is a direction in which we are extremely passionate,” he says. “To uncover something truly fundamental that will take people a long time to comprehend.”