Victor Gong

UCSC Professor Yi Zhang talks about how AI lanaguage models comprehend human language. Research club held the speaker event last Friday for those interested in computer science and artificial intelligence.

University of California, Santa Cruz Professor Yi Zhang presented her award-winning research in Natural Language Processing, Multimodality and Information Retrieval in Chris Spenner’s Room last Friday from 3:15 to 4:15 p.m. 

The session began with Research Club Officer Linda Zeng (10) introducing Dr. Zhang’s background. Linda met Dr. Zhang in the California State Summer School for Mathematics and Science program, where she was the main professor teaching Cluster Eleven, a class focused on artificial intelligence. 

“The way she frames explaining machine learning to people who aren’t familiar with it is really good,” Linda said. “First as a mathematical function and then as a human brain — I’ll definitely use some of her ideas when I try to explain my research to other people.”

To begin the presentation, Dr. Zhang dove into the intricacies behind AI language models, types used by ChatGPT and other conversation bots. Before responding, a language model must first understand natural human language. The program converts sentences into structured representations and deals with variations in structure as well as ambiguities, such as similar sentences with different meanings.

Dr. Zhang elaborated on the natural language generation portion of the language model, outlining the two main types: retrieval-based and generation-based. While retrieval-based models compare inputs with a database of predefined responses, generation-based models can dynamically generate responses, either through filling the gaps of a template or through “sequence-to-sequence”-based machine learning, where the model learns how to respond to inputs through training. 

Focusing on the adaptive capabilities of AI, Dr. Zhang explained how “computational” neurons drive machine learning. Imitating the function of biological neurons, computational neurons process input-output pairs and learn to recognize patterns. Given enough data, a network of neurons can formulate a function to respond to any prompt. Dr. Zhang compared the function to how humans determine a table is a table.

“Large-scale language models can be used to write stories, write poems, answer questions, translate languages and even more,” Dr. Zhang said. “Imagine if you have read all the books in the world and all the documents on the Internet, what will you be? Those neural networks have done that, and they trained their model that way.”

Dr. Zhang also mentioned the possible impacts of AI models on society, including on employment, education and entertainment. Although these AI programs may offer assistance, she also addressed concerns surrounding AI seizing jobs and creative roles in the future. Daniel Wu (10), who attended the lecture, comments on the impacts of AI.

“I don’t think AI is going to take over society, and we’re all going to die, but I think a lot of jobs will be replaced,” Daniel said. “The necessity of work is going to shift, and a lot of labor is going to become doable by AI.”

Transitioning to research advice, Dr. Zhang cautioned the audience about the intense critique of research journal editing committees. She recommended that they revise and resubmit papers if they did not get accepted, stressing that many highly regarded papers in the scientific community underwent multiple rounds of rejection before finally being published. 

The presentation ended with a Q&A segment where Dr. Zhang responded to questions about her experience in research labs and requests for advice on submitting to conferences. The audience also inquired into the possible uses and limitations of AI. 

“There are certain things that AI cannot replace,” Dr. Zhang said. “Initially, people think that creativity will not be replaced, but it will. I think it’ll be like 60% done by AI and 40% done by humans. In the field, we call it ‘co-pilot’ instead of ‘pilot’, because in the end, AI still makes mistakes.”