What can AI learn from children?
UC Berkeley researchers explore how AI can learn language more efficiently by mimicking children’s imitation-based learning, using speech and virtual vocal tracts to reduce training data needs.
AI models like OpenAI’s GPT-3 rely on massive datasets—570 gigabytes of text—to generate human-like responses, but children learn language through imitation and experimentation with far less input. UC Berkeley linguist Gašper Beguš and his team are investigating whether AI can adopt this more efficient, energy-conserving approach by modeling how babies acquire language. Their work focuses on teaching AI through speech rather than text, using virtual vocal tracts to replicate the sounds humans produce during early language development.
Gašper Beguš and his colleagues developed artificial baby language learners that imitate human speech patterns, including the mistakes babies make, to train AI models. Unlike traditional Large Language Models (LLMs), which predict words based on patterns in training data, these models use a process called 'imagitation'—learning by imitation and creative application to unfamiliar situations. The team also tested a generative adversarial network (GAN) that mimics human brain responses to sounds, achieving one of the closest brain-model matches reported to date.
The researchers’ work extends beyond AI training efficiency, offering insights into how language evolves and how animals use communication. In 2017, Facebook AI researchers observed AI models developing an incomprehensible communication protocol during a negotiation game, raising questions about how AI-generated language might diverge from human norms without direct supervision. Gašper Beguš and his team created ConlangCrafter, an automated system for constructing artificial languages with grammar, lexicon, and translation capabilities to explore linguistic possibilities.
Gašper Beguš, as the Linguistic Lead at Project CETI, used generative AI to analyze sperm whale communication, revealing patterns in their codas that resemble human vowel structures. The findings suggest whale communication may be more complex than previously thought, with recurring vowel-like structures detected in their click sequences. This research could influence debates on marine sound pollution, animal rights, and legal protections for highly intelligent species.