Large-scale language models such as GPT-4 and PaLM 2 have evolved to become a key part of modern AI systems, revolutionizing our understanding of natural language processing and transforming many fields. Despite significant advances in comprehension and ability to generate appropriate responses in context, LLM still has some shortcomings. The fact that multi-turn interactions with language models generate a large number of tokens well beyond LLM’s input token limit is one of the key issues. For example, GPT-4 is limited to 32,000 tokens. LLMs should maintain contextual information during encounters and generate responses according to the information gathered.
However, simply concatenating all the contextual information and stuffing it into the LLM can easily exceed the processing power of the LLM and accumulate errors, causing the model to lose track of the conversation and reduce the accuracy of the response. Several neural memory mechanisms have been investigated to overcome the token input limitation problem of LLM. The memory component acts as a storage and retrieval system for relevant information from previous interactions. However, scaling LLM with traditional neural memory typically makes storing, retrieving, and manipulating historical information in memory difficult, especially for tasks that require complex multi-hop inference.
There are two main reasons, they all rely on vector similarity calculations, so historical data is not kept in a structured form and not manipulated symbolically, which leads to errors and the accumulation of errors. can cause Researchers from Tsinghua University, Beijing Academy of Artificial Intelligence, and Zhejiang University advocate using databases as an innovative symbolic memory for LLM to solve the above problems. ChatDB is the name of the entire framework. Figure 1 below shows the two parts that make up ChatDB: the LLM controller and its memory. Read and write operations to memory are controlled by the LLM controller. The LLM controller can be any widely used LLM.
An LLM’s memory, which can be symbolic, non-symbolic, or a hybrid of the two, is responsible for tracking past and distributing data so that the LLM can react to human input when necessary. I’m here. ChatDB focuses on leveraging the database as a symbolic memory, allowing historical data to be stored systematically through metaphorical language, or SQL command execution. LLM created these SQL statements. Databases can be used as symbolic memory in situations where historical data needs to be accurately recorded, updated, queried, deleted, and analyzed. For example, a store manager needs to track daily sales. Therefore, it is inappropriate to use matrices or plaintext as memory.
However, using the database as external symbolic memory is very appropriate. Databases use SQL commands to perform precise actions such as inserting, deleting, updating, and selecting data. As a result, using the database as external symbolic memory ensures accuracy and efficiency in managing and manipulating historical data, improving LLM’s performance in situations requiring highly accurate and long-term data capture and processing. significantly improved. The ChatDB framework proposes a memory chaining strategy to make better use of external symbolic memory to further enhance LLM’s reasoning power.
User input is transformed via memory chain technology into a series of intermediate memory manipulation stages to produce the desired output. Complex problems are divided into several memory manipulation stages using memory chain technology, greatly reducing the difficulty of problem solving. Each intermediate step in ChatDB requires one or more SQL statements. The LLM space has benefited greatly from ChatDB. First, it suggests adding the database to LLM as external symbolic memory. It enables organized archiving of historical data and enables symbolic and complex data manipulation using SQL statements.
Second, memory can be manipulated effectively by using memory chaining techniques to transform user input into multi-step intermediate memory operations. This makes ChatDB more efficient, allowing it to manage complex multi-table database transactions more accurately and stably. Finally, their study shows that adding symbolic memory to LLM improves multi-hop reasoning skills, reduces error accumulation, and makes his ChatDB outperform his ChatGPT on synthetic datasets. It shows what you can do.
please check out paper and plan. don’t forget to join 23,000+ ML SubReddit, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email me. Asif@marktechpost.com
🚀 Check out 100’s of AI Tools at the AI Tools Club
Aneesh Tickoo is a consulting intern at MarktechPost. He is currently pursuing his Bachelor of Science in Data Science and Artificial Intelligence from the Indian Institute of Technology (IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is in image processing and he is passionate about building solutions around it. He loves connecting with people and collaborating on interesting projects.
