• 강사 : 안성진 교수 (KAIST 전산학부)
• 일시 : 2024. 11. 4. (월) 16:00~17:30

The Language of Thought Hypothesis (LoTH) argues that human cognition is based on a structured, language-like system known as ‘Mentalese,’ where units similar to words are arranged in sentence-like structures to convey meaning. Neural language models gain substantial benefits from the compositional and symbolic structures within language data; however, how such structures can be learned from non-linguistic observations (e.g., images, videos, audio) remains unclear. This talk will present research on how neural networks might develop these language-like representations through unsupervised learning. Specifically, I emphasize the need for three essential characteristics: structural compositionality, symbolic representation, and generalization ability. Focusing on these attributes, I will outline potential directions for developing neural models of language-like thought, which could deepen our understanding of human cognitive processes and open new avenues for artificial intelligence.