The “popular” ChatGPT does not yet have an “autonomous mind” ——An interview with Professor Jin Yaochu, an expert in artificial intelligence and a member of the European Academy of Sciences | The world trend of technological innovation
Science and Technology Daily reporter Li Shan
Recently, focusing on the hot topic ChatGPT, a reporter from Science and Technology Daily interviewed Jin Yaochu, academician of the European Academy of Sciences and Humboldt Professor of Artificial Intelligence at Bielefeld University in Germany. Professor Jin is concurrently the Chair Professor of “Computational Intelligence” in the Department of Computing at the University of Surrey, the President-designate of the IEEE Computational Intelligence Society, and the editor-in-chief of the “IEEE Transactions on Cognitive and Developmental Systems”.
Professor Jin Yaochu from Bielefeld University in Germany has long been engaged in theoretical research and engineering applications of computational intelligence, evolutionary computing, machine learning, computational neuroscience and other interdisciplinary subjects. He has published more than 500 academic papers, and the papers have been cited more than 37,400 times. From 2019 to 2022, he has been rated as “Global Highly Cited Scientist” by Web of Science for four consecutive years.
Photos provided by interviewees
A major advance in generative models
Talking about the core of ChatGPT, Jin Yaochu said that the technology represented by ChatGPT is a major progress made in the third wave of artificial intelligence technology. It is based on the generative pre-trained Transformer model and uses a “self-attention mechanism”. Compared with the traditional discriminative model, the biggest feature of the generative model is that it directly learns the characteristics of the training data itself, so it can more effectively learn large-capacity samples, especially massive language and visual information.
Traditional language processing models generally can only find the relationship between words based on their adjacent words before and after, so they have great limitations. After Transformer introduces the self-attention mechanism, it can flexibly find not only a certain sentence, but also the relationship between words between different sentences, or even the entire article or different articles, and the learning ability is greatly enhanced. .
Turned out not a day’s work
Regarding the emergence of ChatGPT, Jin Yaochu emphasized that this is only the feeling of the public. From the perspective of scientific research, ChatGPT has also evolved step by step. According to the paper published by OpenAI, ChatGPT took a long time to continuously train and fine-tune with various learning methods. The answer given by the generative model is not the training data given to it in advance, but generated by the model, so it is difficult to guarantee 100% correctness. ChatGPT uses a lot of the latest learning techniques in the training process. For example, unsupervised learning, supervised learning, multi-task learning, few-shot learning, self-supervised learning, etc.
Additionally, it incorporates reinforcement learning based on human feedback for human-aligned, empathetic capabilities. It selects a best answer by human evaluation, that is, scoring the answers it generates. This process is actually a reinforcement learning process, that is, fine-tuning the original model. Other AI models may not use a lot of human feedback to make the model produce answers that are more in line with human expectations like ChatGPT, and the corresponding experience will be different.
Does not yet have an autonomous “mind”
Jin Yaochu said that as a text-to-text dialogue system, the strongest point of ChatGPT is the full learning of natural language, and the generated text is more “fluent” and in line with human experience. This is a major technological innovation. But it is worth noting that there are many generative machine learning models, such as from text to voice information, or text to image information, the performance of ChatGPT may not be so good in these aspects. This is one of its limitations.
On the other hand, the current ChatGPT training data is as of 2021, and it is impossible to answer it accurately without training. In other words, current models are not equipped to create new knowledge. ChatGPT uses inference when answering questions. It’s not a concept of searching a database for an answer. Generative models are about associating an answer from the question itself. This answer is something that the model “learns to digest”, not the original data. Strictly speaking, the model itself does not even know what the answer really means, nor can it guarantee that the answer is correct.
ChatGPT is not “universal”
Jin Yaochu believes that ChatGPT is still far from a real professional application. Now everyone feels very novel and is trying to play. Some people even think that ChatGPT can do everything and can be quickly used in many applications such as autonomous driving. This is actually a misunderstanding. The point of self-driving is not a text-to-text problem, purely generative models are not enough. It requires other discriminative machine learning models, especially for real-time scene recognition.
Take the hotly discussed medical service as an example. Although ChatGPT has achieved good results in the US Medical Licensing Examination, will people actually use ChatGPT to replace doctors for diagnosis and treatment? At least not yet, because there is no guarantee that its answers are 100% correct. Some people say that ChatGPT is sometimes serious nonsense, such a description is more pertinent, and people cannot ignore this risk.
Looking forward to more breakthroughs in the future
Jin Yaochu said that this wave of artificial intelligence may not die down as soon as the first and second waves of artificial intelligence. It will continue to develop, and in the next three to five years, there may be some surprising breakthroughs in the field of artificial intelligence. Many new ideas or technologies based on ChatGPT may emerge, and on this basis, find really good applications in subdivided fields.
As for the combination of artificial intelligence and robots, Jin Yaochu emphasized that, in a sense, this has always been one of the focuses of artificial intelligence research. Now that ChatGPT talks so smoothly with people, if it is combined with digital virtual human technology, it may be difficult for ordinary people to distinguish whether the person answering the question on the computer screen is a real person or a digital person. Of course, the future development process will involve many essential issues of artificial intelligence, including trust, responsibility, ethics, and legal issues. Optimistically, the influence of artificial intelligence applications such as ChatGPT has expanded rapidly, which in turn will promote the resolution of the above problems to a certain extent.