Two are better than one: Context window extension with multi-grained self-injection

The limited context window of contemporary large language models (LLMs) remains a huge barrier to their broader application across various domains. While continual pre-training on long-context data is a straightforward and effective solution, it incurs substantial costs in terms of data acquisition and computational resources. To alleviate this issue, we propose SharedLLM, a novel approach grounded in the design philosophy of multi-grained context compression and query-aware information retrieval. SharedLLM is composed of two short-context LLMs such as LLaMA-2, termed upper model and lower model. The lower model functions as a compressor while the upper model acts as a decoder. The upper model receives compressed, multi-grained context information from the lower model and performs context-aware modeling on the running text. Information transfer between the compressor and decoder occurs only at the lowest layers to refrain from long forward paths in the lower model and redundant cross-attention modules in the upper model. Based on this architecture, we introduce a specialized tree-style data structure to efficiently encode, store and retrieve multi-grained contextual information for text chunks. This structure, combined with a search algorithm, enables rapid encoding and retrieval of relevant information from various levels of the tree based on the input query. This entire process, wherein the sender and receiver are derived from the same LLM layer, is referred to as self-injection.

PEDRO: Parameter-Efficient Fine-tuning with Prompt DEpenDent Representation MOdification

Due to their substantial sizes, large language models (LLMs) are typically deployed within a single-backbone multi-tenant framework. In this setup, a single instance of an LLM backbone must cater to multiple users or tasks through the application of various parameter-efficient fine-tuning (PEFT) models. Despite the availability of numerous effective PEFT techniques such as LoRA, there remains a need for a PEFT approach that achieves both high efficiency during inference and competitive performance on downstream tasks. In this research, we introduce a new and straightforward PEFT methodology named \underline{P}rompt D\underline{E}pen\underline{D}ent \underline{R}epresentation M\underline{O}dification (PEDRO). The proposed method involves integrating a lightweight vector generator into each Transformer layer, which generates vectors contingent upon the input prompts. These vectors then modify the hidden representations created by the LLM through a dot product operation, thereby influencing the semantic output and generated content of the model. Extensive experimentation across a variety of tasks indicates that: (a) PEDRO surpasses recent PEFT benchmarks when using a similar number of tunable parameters. (b) Under the single-backbone multi-tenant deployment model, PEDRO exhibits superior efficiency compared to LoRA, indicating significant industrial potential.

Towards Faster Graph Partitioning via Pre-training and Inductive Inference

Graph partitioning (GP) is a classic problem that divides the node set of a graph into densely-connected blocks. Following the IEEE HPEC Graph Challenge and recent advances in pre-training techniques (e.g., large-language models), we propose PR-GPT (Pre-trained & Refined Graph ParTitioning) based on a novel pre-training & refinement paradigm. We first conduct the offline pre-training of a deep graph learning (DGL) model on small synthetic graphs with various topology properties. By using the inductive inference of DGL, one can directly generalize the pre-trained model (with frozen model parameters) to large graphs and derive feasible GP results. We also use the derived partition as a good initialization of an efficient GP method (e.g., InfoMap) to further refine the quality of partitioning. In this setting, the online generalization and refinement of PR-GPT can not only benefit from the transfer ability regarding quality but also ensure high inference efficiency without re-training. Based on a mechanism of reducing the scale of a graph to be processed by the refinement method, PR-GPT also has the potential to support streaming GP. Experiments on the Graph Challenge benchmark demonstrate that PR-GPT can ensure faster GP on large-scale graphs without significant quality degradation, compared with running a refinement method from scratch. We will make our code public at https://github.com/KuroginQin/PRGPT.

Riemann-based Multi-scale Attention Reasoning Network for Text-3D Retrieval

Due to the challenges in acquiring paired Text-3D data and the inherent irregularity of 3D data structures, combined representation learning of 3D point clouds and text remains unexplored. In this paper, we propose a novel Riemann-based Multi-scale Attention Reasoning Network (RMARN) for text-3D retrieval. Specifically, the extracted text and point cloud features are refined by their respective Adaptive Feature Refiner (AFR). Furthermore, we introduce the innovative Riemann Local Similarity (RLS) module and the Global Pooling Similarity (GPS) module. However, as 3D point cloud data and text data often possess complex geometric structures in high-dimensional space, the proposed RLS employs a novel Riemann Attention Mechanism to reflect the intrinsic geometric relationships of the data. Without explicitly defining the manifold, RMARN learns the manifold parameters to better represent the distances between text-point cloud samples. To address the challenges of lacking paired text-3D data, we have created the large-scale Text-3D Retrieval dataset T3DR-HIT, which comprises over 3,380 pairs of text and point cloud data. T3DR-HIT contains coarse-grained indoor 3D scenes and fine-grained Chinese artifact scenes, consisting of 1,380 and over 2,000 text-3D pairs, respectively. Experiments on our custom datasets demonstrate the superior performance of the proposed method. Our code and proposed datasets are available at \url{https://github.com/liwrui/RMARN}.

Exploiting Fine-Grained Prototype Distribution for Boosting Unsupervised Class Incremental Learning

The dynamic nature of open-world scenarios has attracted more attention to class incremental learning (CIL). However, existing CIL methods typically presume the availability of complete ground-truth labels throughout the training process, an assumption rarely met in practical applications. Consequently, this paper explores a more challenging problem of unsupervised class incremental learning (UCIL). The essence of addressing this problem lies in effectively capturing comprehensive feature representations and discovering unknown novel classes. To achieve this, we first model the knowledge of class distribution by exploiting fine-grained prototypes. Subsequently, a granularity alignment technique is introduced to enhance the unsupervised class discovery. Additionally, we proposed a strategy to minimize overlap between novel and existing classes, thereby preserving historical knowledge and mitigating the phenomenon of catastrophic forgetting. Extensive experiments on the five datasets demonstrate that our approach significantly outperforms current state-of-the-art methods, indicating the effectiveness of the proposed method.

A Mean Field Ansatz for Zero-Shot Weight Transfer

The pre-training cost of large language models (LLMs) is prohibitive. One cutting-edge approach to reduce the cost is zero-shot weight transfer, also known as model growth for some cases, which magically transfers the weights trained in a small model to a large model. However, there are still some theoretical mysteries behind the weight transfer. In this paper, inspired by prior applications of mean field theory to neural network dynamics, we introduce a mean field ansatz to provide a theoretical explanation for weight transfer. Specifically, we propose the row-column (RC) ansatz under the mean field point of view, which describes the measure structure of the weights in the neural network (NN) and admits a close measure dynamic. Thus, the weights of different sizes NN admit a common distribution under proper assumptions, and weight transfer methods can be viewed as sampling methods. We empirically validate the RC ansatz by exploring simple MLP examples and LLMs such as GPT-3 and Llama-3.1. We show the mean-field point of view is adequate under suitable assumptions which can provide theoretical support for zero-shot weight transfer.

FuXi Weather: An end-to-end machine learning weather data assimilation and forecasting system

Operational numerical weather prediction systems consist of three fundamental components: the global observing system for data collection, data assimilation for generating initial conditions, and the forecasting model to predict future weather conditions. While NWP have undergone a quiet revolution, with forecast skills progressively improving over the past few decades, their advancement has slowed due to challenges such as high computational costs and the complexities associated with assimilating an increasing volume of observational data and managing finer spatial grids. Advances in machine learning offer an alternative path towards more efficient and accurate weather forecasts. The rise of machine learning based weather forecasting models has also spurred the development of machine learning based DA models or even purely machine learning based weather forecasting systems. This paper introduces FuXi Weather, an end-to-end machine learning based weather forecasting system. FuXi Weather employs specialized data preprocessing and multi-modal data fusion techniques to integrate information from diverse sources under all-sky conditions, including microwave sounders from 3 polar-orbiting satellites and radio occultation data from Global Navigation Satellite System. Operating on a 6-hourly DA and forecasting cycle, FuXi Weather independently generates robust and accurate 10-day global weather forecasts at a spatial resolution of 0.25\textdegree. It surpasses the European Centre for Medium-range Weather Forecasts high-resolution forecasts in terms of predictability, extending the skillful forecast lead times for several key weather variables such as the geopotential height at 500 hPa from 9.25 days to 9.5 days. The system's high computational efficiency and robust performance, even with limited observations, demonstrates its potential as a promising alternative to traditional NWP systems.

Multi-scale Restoration of Missing Data in Optical Time-series Images with Masked Spatial-Temporal Attention Network

Due to factors such as thick cloud cover and sensor limitations, remote sensing images often suffer from significant missing data, resulting in incomplete time-series information. Existing methods for imputing missing values in remote sensing images do not fully exploit spatio-temporal auxiliary information, leading to limited accuracy in restoration. Therefore, this paper proposes a novel deep learning-based approach called MS2TAN (Multi-scale Masked Spatial-Temporal Attention Network), for reconstructing time-series remote sensing images. Firstly, we introduce an efficient spatio-temporal feature extractor based on Masked Spatial-Temporal Attention (MSTA), to obtain high-quality representations of the spatio-temporal neighborhood features in the missing regions. Secondly, a Multi-scale Restoration Network consisting of the MSTA-based Feature Extractors, is employed to progressively refine the missing values by exploring spatio-temporal neighborhood features at different scales. Thirdly, we propose a ``Pixel-Structure-Perception'' Multi-Objective Joint Optimization method to enhance the visual effects of the reconstruction results from multiple perspectives and preserve more texture structures. Furthermore, the proposed method reconstructs missing values in all input temporal phases in parallel (i.e., Multi-In Multi-Out), achieving higher processing efficiency. Finally, experimental evaluations on two typical missing data restoration tasks across multiple research areas demonstrate that the proposed method outperforms state-of-the-art methods with an improvement of 0.40dB/1.17dB in mean peak signal-to-noise ratio (mPSNR) and 3.77/9.41 thousandths in mean structural similarity (mSSIM), while exhibiting stronger texture and structural consistency.

Retrieval Meets Reasoning: Dynamic In-Context Editing for Long-Text Understanding

Current Large Language Models (LLMs) face inherent limitations due to their pre-defined context lengths, which impede their capacity for multi-hop reasoning within extensive textual contexts. While existing techniques like Retrieval-Augmented Generation (RAG) have attempted to bridge this gap by sourcing external information, they fall short when direct answers are not readily available. We introduce a novel approach that re-imagines information retrieval through dynamic in-context editing, inspired by recent breakthroughs in knowledge editing. By treating lengthy contexts as malleable external knowledge, our method interactively gathers and integrates relevant information, thereby enabling LLMs to perform sophisticated reasoning steps. Experimental results demonstrate that our method effectively empowers context-limited LLMs, such as Llama2, to engage in multi-hop reasoning with improved performance, which outperforms state-of-the-art context window extrapolation methods and even compares favorably to more advanced commercial long-context models. Our interactive method not only enhances reasoning capabilities but also mitigates the associated training and computational costs, making it a pragmatic solution for enhancing LLMs' reasoning within expansive contexts.

Beyond Scaling Laws: Understanding Transformer Performance with Associative Memory

Increasing the size of a Transformer model does not always lead to enhanced performance. This phenomenon cannot be explained by the empirical scaling laws. Furthermore, improved generalization ability occurs as the model memorizes the training samples. We present a theoretical framework that sheds light on the memorization process and performance dynamics of transformer-based language models. We model the behavior of Transformers with associative memories using Hopfield networks, such that each transformer block effectively conducts an approximate nearest-neighbor search. Based on this, we design an energy function analogous to that in the modern continuous Hopfield network which provides an insightful explanation for the attention mechanism. Using the majorization-minimization technique, we construct a global energy function that captures the layered architecture of the Transformer. Under specific conditions, we show that the minimum achievable cross-entropy loss is bounded from below by a constant approximately equal to 1. We substantiate our theoretical results by conducting experiments with GPT-2 on various data sizes, as well as training vanilla Transformers on a dataset of 2M tokens.