The ‘Plasticity Problem’ and LLM’s Inability to Learn Continuously

AI models face significant learning limitations: Recent research reveals that deep learning AI models, including large language models, struggle to incorporate new information without complete retraining.

• A study published in Nature by scientists from the University of Alberta highlights a major flaw in AI models’ ability to learn continuously.
• Deep learning AI models, which find patterns in vast amounts of data, fail to function effectively in “continual learning settings” where new concepts are introduced to existing training.
• Attempting to teach an existing deep learning model new information often requires retraining it entirely from scratch.

The plasticity problem: When new information is introduced to existing AI models, their artificial neurons can lose their ability to learn, significantly impacting their functionality.

• Without complete retraining, artificial neurons in AI models can sink to a value of zero, resulting in a loss of “plasticity” or the ability to learn.
• Lead study author Shibhansh Dohare likens this to 90 percent of neurons in a human brain becoming dead, leaving insufficient capacity for learning.
• This plasticity loss creates a significant barrier between current AI models and the concept of artificial general intelligence (AGI), which aims to match human-level intelligence.

Financial implications for AI companies: The necessity for complete retraining presents a substantial financial obstacle for AI companies.

• Training advanced AI models is an expensive and resource-intensive process.
• For large language models using substantial portions of internet data, each retraining could cost millions of dollars in computation.
• This financial burden adds to the already high operational costs of AI companies.

Potential solutions and ongoing challenges: Researchers are exploring ways to address the plasticity problem, but a comprehensive solution remains elusive.

• The study authors developed an algorithm capable of randomly reviving certain damaged or “dead” AI neurons, showing some success in countering the plasticity issue.
• However, a practical and comprehensive solution for continual learning in AI models is still out of reach.
• Dohare emphasizes that solving the continual learning problem could significantly reduce the cost of training these models, making it a “billion-dollar question” for the AI industry.

Broader implications for AI development: The plasticity problem highlights the fundamental differences between artificial and human intelligence, raising questions about the future of AI advancement.

• This limitation underscores the challenges in creating truly adaptable and continuously learning AI systems.
• It also emphasizes the need for innovative approaches to AI model development that can more closely mimic human learning processes.
• As the AI industry continues to evolve, addressing this limitation could be crucial for the development of more advanced and cost-effective AI technologies.