Teaching AI to "think" more like humans
Researchers in Hong Kong have made significant progress in developing AI models that can perform more like the human brain. Their work suggests that these models can be improved by going beyond simply predicting the next word in a sequence.
Current large language models are adept at predicting the next word in a sentence, but this approach may not fully capture how humans comprehend language. Humans understand the meaning of a sentence by considering the relationships between words and the overall context.
The research team, led by Professor Ping Li, Dean of the Hong Kong Polytechnic University, explored a technique called next sentence prediction (NSP) to train AI models. NSP challenges the model to predict the next sentence in a sequence, rather than just the next word. This approach, according to the study, resulted in a model that showed better alignment with brain activity in areas associated with language processing.
The improved model also performed better on tasks that required understanding the connection between sentences. This suggests that NSP can be a valuable tool for developing AI models with a deeper understanding of human language.
Overall, this research, which has been published in Science Advances, highlights the importance of moving beyond word prediction when designing AI models for language comprehension. Techniques like NSP can help create AI models that come closer to mimicking the way humans process and understand language. At the same time, this area of research could lead to new insights into higher-level language mechanisms in the brain.