Remember meForgot password?
    Log in with Twitter

article imageGoogle AI created a 'child' more powerful than any human-made AI

By James Walker     Dec 4, 2017 in Technology
Google Brain researchers have announced an AI system created its own "child" AI more powerful than any neural network crafted by a human. The main AI is capable of developing its own specialised children to help it complete new tasks more rapidly.
Reinforcement learning
Google first unveiled its parent-and-child AI back in May. Recently, it returned to the project with a new challenge, tasking the controller AI with creating a descendant more powerful than anything a human could make. As detailed in Futurism, researchers on the Google Brain team used reinforcement learning to make the child more intelligent than any network they'd created in the past.
The "child" was tasked with intelligently identifying objects inside a video in real-time. It had to recognise and label several kinds of object in complex scenes as the video was playing. The child, called NASNet, would have its performance evaluated by its "parent" controller, the AutoML machine learning model. AutoML would analyse the results and then provide feedback to NASNet on how it could improve.
More versatile AI
After running the tests thousands of times, NASNet was so proficient at the task that it outperformed all other existing computer vision systems. In trials with two of the "most respected large-scale academic data sets," NASNet achieved an accuracy score 1.2 percent higher than any previous result, while also boasting 4 percent greater efficiency. By working together, the parent and child managed to outsmart human attempts to create more intelligent AI.
A sample of the training video used by NASNet
A sample of the training video used by NASNet
READ NEXT: China reasserts right to "cyber sovereignty"
The findings have significant implications for the future of machine learning. Beyond demonstrating the power of reinforcement learning, the research shows how automation might be the key to finding more accurate AI models. In the future, AI could be responsible for creating new AI, with humans left providing the input data and defining what the model should produce.
The parent-and-child approach to neural networks could make AI more adaptable and help it to learn quickly in pressurised situations. Autonomous vehicles, proactive assistance and medical robots could all benefit from AI tech capable of recognising objects and intentions more quickly, improving safety and the user experience.
AI evolution
The research also renews concerns that AI could one day escape humanity's grasp. While NASNet's unlikely to evolve into a killer machine overnight, self-generating AI does raise near-term issues. One concern is that biases in the "parent" could be passed on to generations of descendants, potentially excluding certain users or masking sources of truth in input datasets.
It's possible AI systems could create descendants so rapidly that humans can't keep up, either through legislation or in terms of technical support. Before a system's fully implemented, the AI at its core might have created a more powerful derivative with a new set of requirements.
Google's already invested in researching the ethical implications of AI through a working group formed by its DeepMind subsidiary. Several other major tech firms including Amazon and Facebook are also collaborating in a partnership that seeks to ensure AI will benefit society in an ethical manner. It's unclear yet whether rules will be needed to control the use of self-propagating AI.
More about Google, Ai, Artificial intelligence, machine learning, computer vision
Latest News
Top News