CompCog: Computational, distributed accounts of human memory: improving cognitive models

点击翻译按钮获取中文摘要

Memory is among the most impressive aspects of human cognition, allowing us to learn new words or new ideas from just a few examples. However, the scientific understanding of how this learning occurs is limited. This research project focuses on how learning occurs in the context of memory for language. Within the human mind, there is something like a dictionary that tells people what words mean (semantics) and how words are combined to make grammatical sentences (syntax). How does the mind learn this dictionary from experience with a language? Computer simulations can help science better understand this learning process. This scientific understanding can, in turn, help teach languages in the classroom and aid in the early detection of language deficits, whether it be developmental deficits in children, or age-related deficits in adults. Furthermore, improving the ability of computers to simulate language learning processes can also lead to the development of better technology such as machine translation, web search, and virtual assistants. This project considers how a better understanding of language learning can help us avoid common pitfalls of memory connected to the use of language. For example, humans easily over-generalize and judge a "book by its cover", associating certain occupations or personality traits with a gender. If we know how people come up with associations between words and concepts, we can also detect and prevent prejudices in language to help ensure that artificial intelligence applications, such as web search, do not produce prejudiced results. The project supports an interdisciplinary and diverse team of researchers and students at Penn State, attracting college students to engage with research in cognitive science and artificial intelligence.In this project, the researchers are designing a new model of human memory, the Hierarchical Holographic Model. This computational model helps explain certain aspects of how words and languages are learned. The model draws on the successes of artificial intelligence and deep neural networks, and applies these insights to psychology. With this model, the researchers investigate the question of whether human memory has the ability to detect arbitrarily indirect associations between concepts. The model uses a recursive learning process, building on previously learned knowledge to acquire new knowledge, which allows the model to learn arbitrarily indirect and abstract relationships between words. The researchers consider evidence that sensitivity to abstract relations between words improves the ability of the computer model to learn syntax, such as parts-of-speech, and to use words appropriately to construct grammatical sentences. This work will be assessed against human language data and competing computational models. The success of the computational model should provide evidence that (1) language acquisition depends on indirect associations, and (2) human memory must be able to form indirect associations to facilitate it.