Computer scientists working for a high-tech division of Google are testing how machine learning algorithms can be created from scratch, then evolve naturally, based on simple math.
Experts behind Google’s AutoML suite of artificial intelligence tools have now showcased fresh research which suggests the existing software could potentially be updated to “automatically discover” completely unknown algorithms while also reducing human bias during the data input process.
According to ScienceMag, the software, known as AutoML-Zero, resembles the process of evolution, with code improving every generation with little human interaction.
Machine learning tools are “trained” to find patterns in vast amounts of data while automating such processes and constantly being refined based on past experience.
But researchers say this comes with drawbacks that AutoML-Zero aims to fix. Namely, the introduction of bias.
“Human-designed components bias the search results in favor of human-designed algorithms, possibly reducing the innovation potential of AutoML,” their team’s paper states. “Innovation is also limited by having fewer options: you cannot discover what you cannot search for.”
The analysis, which was published last month on arXiv, is titled “Evolving Machine Learning Algorithms From Scratch” and is credited to a team working for Google Brain division.
“The nice thing about this kind of AI is that it can be left to its own devices without any pre-defined parameters, and is able to plug away 24/7 working on developing new algorithms,” Ray Walsh, a computer expert and digital researcher at ProPrivacy, told Newsweek.
As noted by ScienceMag, AutoML-Zero is designed to create a population of 100 “candidate algorithms” by combining basic random math, then testing the results on simple tasks such as image differentiation. The best performing algorithms then “evolve” by randomly changing their code.
The results—which will be variants of the most successful algorithms—then get added to the general population, as older and less successful algorithms get left behind, and the process continues to repeat. The network grows significantly, in turn giving the system more natural algorithms to work with.
Haran Jackson, the chief technology officer (CTO) at Techspert, who has a PhD in Computing from the University of Cambridge, told Newsweek that AutoML tools are typically used to “identify and extract” the most useful features from datasets—and this approach is a welcome development.
“As exciting as AutoML is, it is restricted to finding top-performing algorithms out of the, admittedly large, assortment of algorithms that we already know of,” he said.
“There is a sense amongst many members of the community that the most impressive feats of artificial intelligence will only be achieved with the invention of new algorithms that are fundamentally different to those that we as a species have so far devised.
“This is what makes the aforementioned paper so interesting. It presents a method by which we can automatically construct and test completely novel machine learning algorithms.”
Jackson, too, said the approach taken was similar to the facts of evolution first proposed by Charles Darwin, noting how the Google team was able to induce “mutations” into the set of algorithms.
“The mutated algorithms that did a better job of solving real-world problems were kept alive, with the poorly-performing ones being discarded,” he elaborated.
“This was done repeatedly, until a set of high-performing algorithms was found. One intriguing aspect of the study is that this process ‘rediscovered’ some of the neural network algorithms that we already know and use. It’s extremely exciting to see if it can turn up any algorithms that we haven’t even thought of yet, the impact of which to our daily lives may be enormous.” Google has been contacted for comment.
The development of AutoML was previously praised by Alphabet’s CEO Sundar Pichai, who said it had been used to improve an algorithm that could detect the spread of breast cancer to adjacent lymph nodes. “It’s inspiring to see how AI is starting to bear fruit,” he wrote in a 2018 blog post.
The Google Brain team members who collaborated on the paper said the concepts in the most recent research were a solid starting point, but stressed that the project is far from over.
“Starting from empty component functions and using only basic mathematical operations, we evolved linear regressors, neural networks, gradient descent… multiplicative interactions. These results are promising, but there is still much work to be done,” the scientists’ preprint paper noted.
Walsh told Newsweek: “The developers of AutoML-Zero believe they have produced a system that has the ability to output algorithms human developers may never have thought of.
“According to the developers, due to its lack of human intervention AutoML-Zero has the potential to produce algorithms that are more free from human biases. This theoretically could result in cutting-edge algorithms that businesses could rely on to improve their efficiency.
“However, it is worth bearing in mind that for the time being the AI is still proof of concept and it will be some time before it is able to output the complex kinds of algorithms currently in use. On the other hand, the research [demonstrates how] the future of AI may be algorithms produced by other machines.”