What if artificial intelligence could learn and adapt without human help? Using evolutionary principles like "survival of the fittest," researchers have developed an innovative program to create AI algorithms from scratch that improve independently. This self-evolving AI can replicate what engineers have achieved over decades in less than a day, potentially transforming machine learning and pushing AI boundaries.
Creating complex algorithms, like neural networks, has long been a time-intensive process in AI research. Inspired by the human brain, neural networks power technologies like self-driving cars and speech recognition. Engineers painstakingly build smaller circuits, such as those identifying road signs for autonomous driving, and connect them into a coherent system over months. Although automation has streamlined parts of this work, human limitations and biases still shape AI’s architecture. Google AI researcher Quoc Le and his team developed an experimental solution to go beyond these limitations: AutoML-Zero, an AI that designs itself from scratch without predefined designs or constraints.
AutoML-Zero uses Darwinian evolution as its core principle. The system starts by generating a population of 100 candidate algorithms, combining basic mathematical operations randomly. These algorithms then tackle a simple task, like determining whether an image shows a cat or a truck.
The program evaluates each algorithm’s performance, selects the best, makes random adjustments (mutations), and reinserts these variations back into the population. Through repeated cycles, the algorithms evolve and improve autonomously. Unlike single-population systems, AutoML-Zero simultaneously manages thousands of populations, exploring tens of thousands of algorithms per second. This high-speed iteration helps the program discover efficient solutions, sharing successful algorithms among populations and removing duplicates to avoid dead ends.
AutoML-Zero remains in its early stages. While it can create viable algorithms, it has yet to reach the cutting-edge level of human-designed AI. Joaquin Vanschoren, a computer scientist at Eindhoven University of Technology, suggests “priming the pump” by seeding AutoML-Zero with foundational machine learning knowledge, thus reducing the learning curve. This hybrid approach of combining human insights with evolutionary algorithms could enable AutoML-Zero to make leaps toward even more sophisticated AI systems.
Le’s team has already explored this hybrid strategy. In another study, they showed that AutoML-Zero could apply its evolutionary approach to improve individual components within neural networks, potentially making impactful advancements incrementally.
One of the system’s most intriguing potentials lies in its capacity to discover entirely new AI capabilities. Le believes that expanding AutoML-Zero’s mathematical library and increasing computing resources could lead to groundbreaking discoveries in AI—a prospect he and his team are eager to explore.
Self-evolving AI could revolutionize fields that require continuous adaptation, such as healthcare, finance, and autonomous systems, where conditions frequently shift. The ability to adapt in real-time would enhance efficiency and responsiveness, fundamentally changing the landscape of these industries.
As self-evolving AI systems advance, ethical and societal considerations will grow, especially as they blur the line between human-generated and AI-generated knowledge. While evolutionary algorithms are not new, their increasing autonomy invites critical reflection on their potential impact on human lives and decisions.
AutoML-Zero and similar self-evolving systems are pioneering a new era in AI research. By integrating machine learning with evolutionary principles, researchers are accelerating the development of algorithms and enabling them to evolve autonomously. Although challenges remain, the future promises machines capable of driving the next wave of innovation and pushing technology toward new frontiers.