Click on the blue text
Follow us
AI TIME welcomes every AI enthusiast to join!
With the rise of general artificial intelligence (AGI), what changes will the graph field usher in? How will graph data processing, graph neural networks, graph reasoning and recommendation, graph databases and knowledge graphs cope with the challenges of the new era? These issues are worthy of our in-depth discussion and thinking.
On August 16, 2023, the 17th AI TIME PhD Debate event "In the new era of AI, where should graph research go?" specially invited Liu Qian, a researcher at Singapore's Sea AI Lab, Zhu Jing, a doctoral student at the University of Michigan, Ann Arbor, and Mo Luo Linhao, a doctoral student at Nash University, Wang Heng, an undergraduate student at Xi'an Jiaotong University, and Mao Haitao, a doctoral student at Michigan State University. The five guests focused on the latest research results in the field of graph, discussed innovative application cases, and jointly explored the development direction and prospects of graph in the new era of AI.
How can the Graph field help the existing AGI development?
Liu Gan: So far, the graph field has not developed a universal model similar to Bert, because it lacks a common basic unit, and the basic unit has different meanings in different scenarios, making it difficult to unify. Finding information that can be shared between different graphs and conducting research on transfer learning on graphs are issues worth exploring, and are also where graphs may be connected to AGI. But it is difficult to train a GNN network model with a complex structure. In the field of NLP, there are two methods. The first is that the transformer can be trained in parallel. The second is that after the transformer has deepened the number of training layers, although there is instability in the training, there is still a chance to learn good features. At present, it is generally believed that scaling up is a shortcut to faster artificial intelligence. If the GNN network does not have the characteristics of scaling up, the training of the model and the understanding of the graph structure will still be a big bottleneck.
Mao Haitao: In large models, text information alone is unstable. Graph can supplement additional structural information, adding a structural mode to the text modality, making the generation of large models more robust. Although the structural information of graph can make up for the shortcomings of text information, how to extract effective information in the graph structure is currently a problem.
Luo Linhao: The recent research hotspot is whether there is a transformer-like structure to understand the graph structure. The core of this series of research is how to better amplify the structural information in the knowledge graph to help the language model improve the performance. The improvement of performance includes two aspects: on the one hand, it is to improve the performance and understanding of the graph structure of the language model; on the other hand, it is how to extract the structural information of the knowledge graph to help the language model enhance certain specific capabilities. In addition, the structure of related reasoning questions such as common sense question and answer is not a single path, and the graph structure cannot be directly understood in the form of a language model. GNN needs to be integrated into it, and the structured information is extracted and then injected into the language model to improve its complex reasoning capabilities.
Wang Heng: There will be some robot accounts in social networks. If you only use the existing language model to detect the accounts, it is easy to have deviations. However, using the graph structure to assist the existing AGI can enhance the robustness of the model.
Zhu Jing: The knowledge graph is more of a strong regularization method, because the information in it is largely correct, but the information output by the large language model may not be completely correct. Its understanding of the structure It is soft regularized. In the future, both soft regularization and strong regularization should be able to coexist at the same time.
What impact will the existing AGI development have on the Graph field?
Luo Linhao: The large language model itself has some shortcomings. Its representation of knowledge is implicit, which can cause "illusion" problems and generate information that is contrary to the facts. Moreover, we have no way of knowing the reasoning process of the large language model. The structured representation of knowledge graphs shows advantages. Its knowledge storage is structured and highly accurate, which can help large language models improve their capabilities. Of course, large language models also have certain advantages. They have strong text understanding capabilities and good generalization. These characteristics can be used to improve the incompleteness of knowledge graphs, thereby using LLM to solve some challenges faced by knowledge graphs.
Liu Gan: If it were in the Bert era, Bert could provide a good representation of graph nodes; if there was a very intelligent model, then using it in the graph field would have unexpected capabilities. Applying knowledge graphs to LLM requires maintaining structured information without being able to change the structure of the language model itself, which is a difficult problem. Knowledge graphs are used in fine-tuning and require repeated training, so we consider updating the knowledge stored in the language model with new knowledge. The knowledge spectrum replaces existing knowledge, but the difficulty in adding new knowledge is how to find the corresponding storage location and add it to the model weights in an approximately lossless manner.
Zhu Jing: AGI can enhance graph’s feature extraction and zero-sample learning capabilities. At the same time, it will also play an important role in the fields of graph reasoning and recommendation. Through in-depth understanding and analysis of graph-structured data, AGI can provide more accurate and personalized reasoning and recommendation results, helping to improve the user experience and effects of search engines, advertising recommendations, smart assistants and other applications.
Wang Heng: The emergence of AGI may promote the development of graph data processing technology. It can better understand and process complex graph structure data, thereby providing more efficient and accurate graph analysis and graph computing capabilities. This will drive better results in areas such as social network analysis, recommendation systems, bioinformatics, and more.
Mao Haitao: Knowledge graph is a graph-structured data model used to represent and organize knowledge. AGI has the potential to promote innovation in graph databases and knowledge graphs. The development of AGI may provide better methods and tools to build and maintain large-scale graph databases and knowledge graphs, thereby improving the efficiency and accuracy of knowledge management and intelligent retrieval.
Does Graph need its own large model? What are the possible pathways?
Zhu Jing: The most difficult tasks for graph are node classification and link prediction. For node classification, the node depends on the construction of the graph on the one hand and the position of the label on the model on the other hand; for link prediction, structured understanding tasks need to be solved. So the biggest problem in graph is how to define a task and translate it into understandable natural language.
Luo Linhao: The structural information in the knowledge graph is more of an expansion of factual information. Our understanding of the structure of the knowledge graph is essentially the understanding of text information. From the field of knowledge graph, it is possible to be unified by large models , but how to adjust the size of the knowledge graph to integrate it into a large model will be a greater challenge. For the graph field, each graph itself is a domain or model, and its feature representation dimensions may be different. The graph model may not be unified like the language model. More possibilities are based on a In the form of transfer learning, using large models under graphs of different modes requires improvements and alignment operations.
Liu Qian: Today there are many large models in many fields, but they involve some complex problems and require skills in many fields to solve. The same situation exists in graph. If we make different basic graph models for different fields, it will be difficult to show surprising characteristics.
Wang Heng: Large language models have achieved remarkable results in fields such as natural language processing and computer vision, but whether similar large models are also needed in the graph field remains to be further studied and discussed. For example, graph reasoning involves discovering hidden patterns and relationships from graphs, while graph recommendation involves extracting useful information from graphs to make personalized recommendations. Large models may help to better understand and utilize complex relationships in graphs. Thereby improving the accuracy and effectiveness of reasoning and recommendations.
Mao Haitao: The final form of graph will not be unified by large models. It may have larger models to solve specific problems in certain specific fields such as social networks, citizen networks, friend networks, etc.
Does Graph require higher intelligence? What issues should the Graph field be concerned about in the future?
Mao Haitao: It must require higher intelligence, but it is uncertain whether it can be presented in the form of LLM. First of all, graph itself is very difficult, and secondly, it also has some problems. I think intelligent reasoning can be compared to machine translation, but the biggest challenge at present is to solve the gap between symbolic language and natural language. For example, natural language is ambiguous, while symbolic language is accurate; natural language is short, And symbolic language is complex and so on.
Wang Heng: Large language models are used in new tasks with implicit graph structures. A very important potential issue is whether large language models can use natural language to perform graph-based reasoning. By constructing the NLGraph data set, we simulated 8 graph reasoning problems of different difficulties and found that large language models have basic graph reasoning capabilities, and advanced prompting methods have mixed effects on graph reasoning. Context learning is difficult to work on difficult graph reasoning problems, and large language models are very vulnerable to spurious correlations in graph reasoning problems. We also made some attempts in the later period, such as proposing prompting techniques to improve graph reasoning capabilities. Experimental results show that they are effective on simple tasks, but the effect is not significant on complex tasks.
Luo Linhao: How to perform some reasoning operations or path searches on the knowledge graph or graph structure is an issue worthy of attention. With the constraints of the graph structure, reasoning on the large language model is transformed into reasoning on the graph, making the entire model more controllable, which is also a key point. In the future, whether graph will also have multi-modal requirements and integrate multi-modal information with graph information may also be an issue that graph needs to consider to achieve higher-level intelligence.
Liu Gan: Graph has its own language, such as sparkle. Many programming languages focus on multi-effect reasoning and numerical calculations. Numerical calculations in graph are very difficult. Regarding the characteristics of different languages in sparkle, SQL and language models, I think the biggest difference is that before the emergence of language models, people generally believed that common sense was an independent problem. After its emergence, we found that it was difficult to express in symbolic language before. The problem can be easily solved by language models. Therefore, designing a form that can express complex logic more clearly is a better research direction, and it is also a direction worth exploring in code generation and natural language reasoning.
Zhu Jing: Graph will definitely have higher intelligence in the future. Multi-model graph is a direction that can be studied in the future. If we combine multi-modal information with the structural information of graph, we may be able to solve more complex problems. The task of reasoning.
Organized by: Chen Yan
Reviewers: Liu Qian, Zhu Jing, Luo Linhao, Wang Heng, Mao Haitao
Recommended articles from past issues
Remember to follow us! New knowledge every day!
About AI TIME
AI TIME originated in 2019, aiming to carry forward the spirit of scientific speculation, invite people from all walks of life to explore the essential issues of artificial intelligence theory, algorithms and scenario applications, strengthen the collision of ideas, and connect global AI scholars, industry experts and enthusiasts, hoping to In the form of a debate, we explore the contradiction between artificial intelligence and the future of mankind, and explore the future of the field of artificial intelligence.
So far, AI TIME has invited more than 1,300 speakers at home and abroad, held more than 600 events, and been watched by more than 6 million people.
I know you
look in
oh
~
Click to read the original text to view the replay!