Researchers develop a new framework to train AI in real-time

Researchers develop a new framework to train AI in real-time

A team of researchers from Duke University and the Army Research Laboratory has developed a new platform to help AI learn to perform complex tasks more like humans. This framework, nicknamed GUIDE, functions by allowing humans to observe and provide real-time feedback.

Even though artificial intelligence (AI) has made considerable advancements, it still relies on massive datasets and simulations, regardless of the application. This limits the adaptability of traditional feedback approaches, making it harder for AIs to make quick decisions based on limited information.

It remains a challenge for AI to handle tasks that require fast decision-making based on limited learning information,” says Boyuan Chen.

Existing training methods are often constrained by their reliance on extensive pre-existing datasets while also struggling with the limited adaptability of traditional feedback approaches. We aimed to bridge this gap by incorporating real-time continuous human feedback.

In a debut study, researchers used GUIDE to train AI to play hide-and-seek. This study game involves a central C-shaped barrier with two beetle-shaped red and green players. While computers control both of the players, GUIDE assists only the red player.

The field remains black in this game until the red player explores it. As the red player chases the green one and explores the area, the human trainer provides feedback on its searching strategy.

The experiment involved the participation of 50 adults with no background in prior training or specialized knowledge. Researchers found that just 10 minutes of feedback significantly improved the performance of AI. Compared to the current state-of-the-art human-guided learning methods, GUIDE demonstrated a 30% success rate.

GUIDE training gameGUIDE training game
GUIDE as a novel framework for real-time human-shaped agents enabling continuous feedback and continual improvements without human trainers. It also aims to understand how individual differences affect their guided agents’ performances.

Creating realistic 3D shapes using generative AI

This strong quantitative and qualitative evidence highlights the effectiveness of our approach. It shows how GUIDE can boost adaptability, helping AI to navigate and respond to complex, dynamic environments independently,” said Lingyu Zhang.

During their study, the team also found that human trainers are needed only for a short period. As these participants furnished their feedback, researchers created an AI simulation based on the insights. This will enable continued training of AI models.

One of GUIDE’s most intriguing capabilities is exploring individual differences among human trainers. A cognitive test of 50 individuals revealed certain capabilities that influence how effectively a person could guide an AI.

Addressing these highlights will enable AIs to be trained more intuitively, bridging the gap between human intuition and machine learning. This will also enable AI to have swift decision-making power in an environment with limited information.

As AI technologies become more prevalent, it’s crucial to design systems that are intuitive and accessible for everyday users. GUIDE paves the way for smarter, more responsive AI capable of functioning autonomously in dynamic and unpredictable environments,” says Chen.

Using AI to turn audio recordings into street-view images

Journal Reference:

  1. Zhang, L., Ji, Z., Waytowich, N. R., & Chen, B. (2024). GUIDE: Real-Time Human-Shaped Agents. ArXiv. DOI: 10.48550/arXiv.2410.15181


link