The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining "abstract knowledge". As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named $\text{Memory}^3$, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.
Parameter regularization or allocation methods are effective in overcoming catastrophic forgetting in lifelong learning. However, they solve all tasks in a sequence uniformly and ignore the differences in the learning difficulty of different tasks. So parameter regularization methods face significant forgetting when learning a new task very different from learned tasks, and parameter allocation methods face unnecessary parameter overhead when learning simple tasks. In this paper, we propose the Parameter Allocation & Regularization (PAR), which adaptively select an appropriate strategy for each task from parameter allocation and regularization based on its learning difficulty. A task is easy for a model that has learned tasks related to it and vice versa. We propose a divergence estimation method based on the Nearest-Prototype distance to measure the task relatedness using only features of the new task. Moreover, we propose a time-efficient relatedness-aware sampling-based architecture search strategy to reduce the parameter overhead for allocation. Experimental results on multiple benchmarks demonstrate that, compared with SOTAs, our method is scalable and significantly reduces the model's redundancy while improving the model's performance. Further qualitative analysis indicates that PAR obtains reasonable task-relatedness.
In this report, we (SuperHelix team) present our solution to KDD Cup 2021-PCQM4M-LSC, a large-scale quantum chemistry dataset on predicting HOMO-LUMO gap of molecules. Our solution, Lite Geometry Enhanced Molecular representation learning (LiteGEM) achieves a mean absolute error (MAE) of 0.1204 on the test set with the help of deep graph neural networks and various self-supervised learning tasks. The code of the framework can be found in https://github.com/PaddlePaddle/PaddleHelix/tree/dev/competition/kddcup2021-PCQM4M-LSC/.
Retrieval-Augmented Generation (RAG) is a technique that enhances the capabilities of large language models (LLMs) by incorporating external knowledge sources. This method addresses common LLM limitations, including outdated information and the tendency to produce inaccurate "hallucinated" content. However, the evaluation of RAG systems is challenging, as existing benchmarks are limited in scope and diversity. Most of the current benchmarks predominantly assess question-answering applications, overlooking the broader spectrum of situations where RAG could prove advantageous. Moreover, they only evaluate the performance of the LLM component of the RAG pipeline in the experiments, and neglect the influence of the retrieval component and the external knowledge database. To address these issues, this paper constructs a large-scale and more comprehensive benchmark, and evaluates all the components of RAG systems in various RAG application scenarios. Specifically, we have categorized the range of RAG applications into four distinct types-Create, Read, Update, and Delete (CRUD), each representing a unique use case. "Create" refers to scenarios requiring the generation of original, varied content. "Read" involves responding to intricate questions in knowledge-intensive situations. "Update" focuses on revising and rectifying inaccuracies or inconsistencies in pre-existing texts. "Delete" pertains to the task of summarizing extensive texts into more concise forms. For each of these CRUD categories, we have developed comprehensive datasets to evaluate the performance of RAG systems. We also analyze the effects of various components of the RAG system, such as the retriever, the context length, the knowledge base construction, and the LLM. Finally, we provide useful insights for optimizing the RAG technology for different scenarios.
In-context Learning (ICL) is one of the key methods for enhancing the performance of large language models on specific tasks by providing a set of few-shot examples. However, the ICL capability of different types of models shows significant variation due to factors such as model architecture, volume of learning data, and the size of parameters. Generally, the larger the model's parameter size and the more extensive the learning data, the stronger its ICL capability. In this paper, we propose a method SLEICL that involves learning from examples using strong language models and then summarizing and transferring these learned skills to weak language models for inference and application. This ensures the stability and effectiveness of ICL. Compared to directly enabling weak language models to learn from prompt examples, SLEICL reduces the difficulty of ICL for these models. Our experiments, conducted on up to eight datasets with five language models, demonstrate that weak language models achieve consistent improvement over their own zero-shot or few-shot capabilities using the SLEICL method. Some weak language models even surpass the performance of GPT4-1106-preview (zero-shot) with the aid of SLEICL.
Graph neural network (GNN) and label propagation algorithm (LPA) are both message passing algorithms, which have achieved superior performance in semi-supervised classification. GNN performs feature propagation by a neural network to make predictions, while LPA uses label propagation across graph adjacency matrix to get results. However, there is still no effective way to directly combine these two kinds of algorithms. To address this issue, we propose a novel Unified Message Passaging Model (UniMP) that can incorporate feature and label propagation at both training and inference time. First, UniMP adopts a Graph Transformer network, taking feature embedding and label embedding as input information for propagation. Second, to train the network without overfitting in self-loop input label information, UniMP introduces a masked label prediction strategy, in which some percentage of input label information are masked at random, and then predicted. UniMP conceptually unifies feature propagation and label propagation and is empirically powerful. It obtains new state-of-the-art semi-supervised classification results in Open Graph Benchmark (OGB).
Lifelong learning remains an open problem. One of its main difficulties is catastrophic forgetting. Many dynamic expansion approaches have been proposed to address this problem, but they all use homogeneous models of predefined structure for all tasks. The common original model and expansion structures ignore the requirement of different model structures on different tasks, which leads to a less compact model for multiple tasks and causes the model size to increase rapidly as the number of tasks increases. Moreover, they can not perform best on all tasks. To solve those problems, in this paper, we propose a new lifelong learning framework named Searchable Extension Units (SEU) by introducing Neural Architecture Search into lifelong learning, which breaks down the need for a predefined original model and searches for specific extension units for different tasks, without compromising the performance of the model on different tasks. Our approach can obtain a much more compact model without catastrophic forgetting. The experimental results on the PMNIST, the split CIFAR10 dataset, the split CIFAR100 dataset, and the Mixture dataset empirically prove that our method can achieve higher accuracy with much smaller model, whose size is about 25-33 percentage of that of the state-of-the-art methods.