How Artificial General Intelligence Might Learn Like a Human

Christopher Kanan discusses how training AI resembles raising a child, advocating for learning methods inspired by neuroscience. He outlines the distinctions between AGI and ANI, emphasizing AI’s current learning limits and the need for regulations to manage its integration into society. Despite progress, achieving AGI remains a significant challenge, as current architectures are not equipped for broader, human-like reasoning.

Computer scientist Christopher Kanan likens training artificial intelligence (AI) systems to raising a child. Many AI researchers now draw inspiration from how children learn through curiosity and exploration. Kanan emphasizes that ideas from neuroscience and child development could help tackle issues faced in today’s AI algorithms. He warns, however, that while AI systems are advancing, it’s critical to implement safety measures early in development, rather than at the end, stating, “It shouldn’t be the last step, otherwise we can unleash a monster.”

Artificial General Intelligence (AGI) compares to Artificial Narrow Intelligence (ANI) in its goal to replicate human-like understanding and reasoning across various tasks. Kanan notes that AGI has yet to be realized, while ANI excels in specific tasks like image recognition and game strategy. AGI’s advancement remains a core aim of research, and it’s essential to derive inspirations from neuroscience to facilitate continual learning in AI, much like children do.

AI systems learn primarily via deep neural networks, leveraging data to boost their capabilities. Since 2014, deep learning has trained systems to dissect vast amounts of human-annotated data. It enables AI applications to thrive in fields like computer vision and natural language processing. Large language models (LLMs), like GPT-4, learn from an immense corpus of human writing, capable of predicting text patterns without explicit guidance, taking an astonishingly long time for humans to read all that information.

AI shines in handling human languages and performs exceptionally well in various tests. Kanan observes that models such as GPT-4 excel in translating languages, writing essays, and other linguistically demanding tasks, often achieving high scores on standardized tests. They can even provide emotional insights in scenarios, acting as co-researchers to scientists generating novel hypotheses and drafting proposals across disciplines.

Despite their impressive capabilities, generative AIs fall short in human-like self-awareness and reasoning abilities. Kanan points out that they can “hallucinate,” producing incorrect but believable information. Their knowledge is static after training, lacking the adaptability and ongoing learning abilities seen in humans. This deficiency in self-awareness prevents them from effectively navigating real-world uncertainties and limits their interaction capabilities.

The advent of generative AI has sparked discussions about its impact on the workforce and the pressing need for regulatory measures. Kanan highlights that AI significantly alters white-collar jobs by enhancing productivity, potentially leading to significant workforce reductions. However, certain roles requiring human skills are less likely to be overtaken by AI. Regarding existential risks associated with advanced AI, Kanan suggests genuine threats arise from misapplications of AI rather than the technology itself.

While many researchers, including AI pioneers, believe AGI is obtainable, Kanan outlines limitations in current LLM architectures. He notes these AIs are bound to language, contrasting with human thought, which encompasses abstract reasoning and visual imagination. Current models are seen as inadequate for achieving broader intellect akin to humans, marking a significant hurdle in the quest for AGI.

Kanan, an associate professor at the University of Rochester, specializes in AI, continual learning, and brain-inspired algorithms. His research efforts strive to bridge gaps between AI systems and human-like understanding, representing a beacon of hope in the evolving landscape of artificial intelligence.

Christopher Kanan’s insights into AI training highlight the importance of mimicking human learning experiences. While AI systems show remarkable promise, particularly in language and task-specific capabilities, they still lack the self-awareness and reasoning found in humans. The quest for Artificial General Intelligence involves overcoming significant challenges, as current frameworks and models are insufficient. Thus, while the potential for AGI exists, truly replicating human-like intelligence requires further exploration and innovative approaches.

Original Source: www.rochester.edu