Dense ReLU Neural Networks for Temporal-spatial Model

In this paper, we focus on fully connected deep neural networks utilizing the Rectified Linear Unit (ReLU) activation function for nonparametric estimation. We derive non-asymptotic bounds that lead to convergence rates, addressing both temporal and spatial dependence in the observed measurements. By accounting for dependencies across time and space, our models better reflect the complexities of real-world data, enhancing both predictive performance and theoretical robustness. We also tackle the curse of dimensionality by modeling the data on a manifold, exploring the intrinsic dimensionality of high-dimensional data. We broaden existing theoretical findings of temporal-spatial analysis by applying them to neural networks in more general contexts and demonstrate that our proof techniques are effective for models with short-range dependence. Our empirical simulations across various synthetic response functions underscore the superior performance of our method, outperforming established approaches in the existing literature. These findings provide valuable insights into the strong capabilities of dense neural networks for temporal-spatial modeling across a broad range of function classes.

Statistical Guarantees for Lifelong Reinforcement Learning using PAC-Bayesian Theory

Lifelong reinforcement learning (RL) has been developed as a paradigm for extending single-task RL to more realistic, dynamic settings. In lifelong RL, the "life" of an RL agent is modeled as a stream of tasks drawn from a task distribution. We propose EPIC (\underline{E}mpirical \underline{P}AC-Bayes that \underline{I}mproves \underline{C}ontinuously), a novel algorithm designed for lifelong RL using PAC-Bayes theory. EPIC learns a shared policy distribution, referred to as the \textit{world policy}, which enables rapid adaptation to new tasks while retaining valuable knowledge from previous experiences. Our theoretical analysis establishes a relationship between the algorithm's generalization performance and the number of prior tasks preserved in memory. We also derive the sample complexity of EPIC in terms of RL regret. Extensive experiments on a variety of environments demonstrate that EPIC significantly outperforms existing methods in lifelong RL, offering both theoretical guarantees and practical efficacy through the use of the world policy.

SDP4Bit: Toward 4-bit Communication Quantization in Sharded Data Parallelism for LLM Training

Recent years have witnessed a clear trend towards language models with an ever-increasing number of parameters, as well as the growing training overhead and memory usage. Distributed training, particularly through Sharded Data Parallelism (ShardedDP) which partitions optimizer states among workers, has emerged as a crucial technique to mitigate training time and memory usage. Yet, a major challenge in the scalability of ShardedDP is the intensive communication of weights and gradients. While compression techniques can alleviate this issue, they often result in worse accuracy. Driven by this limitation, we propose SDP4Bit (Toward 4Bit Communication Quantization in Sharded Data Parallelism for LLM Training), which effectively reduces the communication of weights and gradients to nearly 4 bits via two novel techniques: quantization on weight differences, and two-level gradient smooth quantization. Furthermore, SDP4Bit presents an algorithm-system co-design with runtime optimization to minimize the computation overhead of compression. In addition to the theoretical guarantees of convergence, we empirically evaluate the accuracy of SDP4Bit on the pre-training of GPT models with up to 6.7 billion parameters, and the results demonstrate a negligible impact on training loss. Furthermore, speed experiments show that SDP4Bit achieves up to 4.08$\times$ speedup in end-to-end throughput on a scale of 128 GPUs.

DODT: Enhanced Online Decision Transformer Learning through Dreamer's Actor-Critic Trajectory Forecasting

Advancements in reinforcement learning have led to the development of sophisticated models capable of learning complex decision-making tasks. However, efficiently integrating world models with decision transformers remains a challenge. In this paper, we introduce a novel approach that combines the Dreamer algorithm's ability to generate anticipatory trajectories with the adaptive learning strengths of the Online Decision Transformer. Our methodology enables parallel training where Dreamer-produced trajectories enhance the contextual decision-making of the transformer, creating a bidirectional enhancement loop. We empirically demonstrate the efficacy of our approach on a suite of challenging benchmarks, achieving notable improvements in sample efficiency and reward maximization over existing methods. Our results indicate that the proposed integrated framework not only accelerates learning but also showcases robustness in diverse and dynamic scenarios, marking a significant step forward in model-based reinforcement learning.

MoE-Pruner: Pruning Mixture-of-Experts Large Language Model using the Hints from Its Router

Mixture-of-Experts (MoE) architectures face challenges such as high memory consumption and redundancy in experts. Pruning MoE can reduce network weights while maintaining model performance. Motivated by the recent observation of emergent large magnitude features in Large Language Models (LLM) and MoE routing policy, we propose MoE-Pruner, a method that prunes weights with the smallest magnitudes multiplied by the corresponding input activations and router weights, on each output neuron. Our pruning method is one-shot, requiring no retraining or weight updates. We evaluate our method on Mixtral-8x7B and Mixtral-8x22B across multiple language benchmarks. Experimental results show that our pruning method significantly outperforms state-of-the-art LLM pruning methods. Furthermore, our pruned MoE models can benefit from a pretrained teacher model through expert-wise knowledge distillation, improving performance post-pruning. Experimental results demonstrate that the Mixtral-8x7B model with 50% sparsity maintains 99% of the performance of the original model after the expert-wise knowledge distillation.

Just say what you want: only-prompting self-rewarding online preference optimization

We address the challenge of online Reinforcement Learning from Human Feedback (RLHF) with a focus on self-rewarding alignment methods. In online RLHF, obtaining feedback requires interaction with the environment, which can be costly when using additional reward models or the GPT-4 API. Current self-rewarding approaches rely heavily on the discriminator's judgment capabilities, which are effective for large-scale models but challenging to transfer to smaller ones. To address these limitations, we propose a novel, only-prompting self-rewarding online algorithm that generates preference datasets without relying on judgment capabilities. Additionally, we employ fine-grained arithmetic control over the optimality gap between positive and negative examples, generating more hard negatives in the later stages of training to help the model better capture subtle human preferences. Finally, we conduct extensive experiments on two base models, Mistral-7B and Mistral-Instruct-7B, which significantly bootstrap the performance of the reference model, achieving 34.5% in the Length-controlled Win Rates of AlpacaEval 2.0.

Visual SLAM with 3D Gaussian Primitives and Depth Priors Enabling Novel View Synthesis

Conventional geometry-based SLAM systems lack dense 3D reconstruction capabilities since their data association usually relies on feature correspondences. Additionally, learning-based SLAM systems often fall short in terms of real-time performance and accuracy. Balancing real-time performance with dense 3D reconstruction capabilities is a challenging problem. In this paper, we propose a real-time RGB-D SLAM system that incorporates a novel view synthesis technique, 3D Gaussian Splatting, for 3D scene representation and pose estimation. This technique leverages the real-time rendering performance of 3D Gaussian Splatting with rasterization and allows for differentiable optimization in real time through CUDA implementation. We also enable mesh reconstruction from 3D Gaussians for explicit dense 3D reconstruction. To estimate accurate camera poses, we utilize a rotation-translation decoupled strategy with inverse optimization. This involves iteratively updating both in several iterations through gradient-based optimization. This process includes differentiably rendering RGB, depth, and silhouette maps and updating the camera parameters to minimize a combined loss of photometric loss, depth geometry loss, and visibility loss, given the existing 3D Gaussian map. However, 3D Gaussian Splatting (3DGS) struggles to accurately represent surfaces due to the multi-view inconsistency of 3D Gaussians, which can lead to reduced accuracy in both camera pose estimation and scene reconstruction. To address this, we utilize depth priors as additional regularization to enforce geometric constraints, thereby improving the accuracy of both pose estimation and 3D reconstruction. We also provide extensive experimental results on public benchmark datasets to demonstrate the effectiveness of our proposed methods in terms of pose accuracy, geometric accuracy, and rendering performance.

Let the Code LLM Edit Itself When You Edit the Code

In this work, we investigate a typical scenario in code generation where a developer edits existing code in real time and requests a code assistant, e.g., a large language model, to re-predict the next token or next line on the fly. Naively, the LLM needs to re-encode the entire KV cache to provide an accurate prediction. However, this process is computationally expensive, especially when the sequence length is long. Simply encoding the edited subsequence and integrating it to the original KV cache meets the temporal confusion problem, leading to significantly worse performance. We address this efficiency and accuracy trade-off by introducing \underline{\textbf{Positional \textbf{I}ntegrity \textbf{E}ncoding} (PIE). Building upon the rotary positional encoding, PIE first removes the rotary matrices in the Key cache that introduce temporal confusion and then reapplies the correct rotary matrices. This process ensures that positional relationships between tokens are correct and requires only a single round of matrix multiplication. We validate the effectiveness of PIE through extensive experiments on the RepoBench-C-8k dataset, utilizing DeepSeek-Coder models with 1.3B, 6.7B, and 33B parameters. Our evaluation includes three real-world coding tasks: code insertion, code deletion, and multi-place code editing. Results demonstrate that PIE reduces computational overhead by over 85% compared to the standard full recomputation approach across all model sizes and tasks while well approximating the model performance.

Interference Cancellation Based Neural Receiver for Superimposed Pilot in Multi-Layer Transmission

In this paper, an interference cancellation based neural receiver for superimposed pilot (SIP) in multi-layer transmission is proposed, where the data and pilot are non-orthogonally superimposed in the same time-frequency resource. Specifically, to deal with the intra-layer and inter-layer interference of SIP under multi-layer transmission, the interference cancellation with superimposed symbol aided channel estimation is leveraged in the neural receiver, accompanied by the pre-design of pilot code-division orthogonal mechanism at transmitter. In addition, to address the complexity issue for inter-vendor collaboration and the generalization problem in practical deployments, respectively, this paper also provides a fixed SIP (F-SIP) design based on constant pilot power ratio and scalable mechanisms for different modulation and coding schemes (MCSs) and transmission layers. Simulation results demonstrate the superiority of the proposed schemes on the performance of block error rate and throughput compared with existing counterparts.

ETA-INIT: Enhancing the Translation Accuracy for Stereo Visual-Inertial SLAM Initialization

As the current initialization method in the state-of-the-art Stereo Visual-Inertial SLAM framework, ORB-SLAM3 has limitations. Its success depends on the performance of the pure stereo SLAM system and is based on the underlying assumption that pure visual SLAM can accurately estimate the camera trajectory, which is essential for inertial parameter estimation. Meanwhile, the further improved initialization method for ORB-SLAM3, known as Stereo-NEC, is time-consuming due to applying keypoint tracking to estimate gyroscope bias with normal epipolar constraints. To address the limitations of previous methods, this paper proposes a method aimed at enhancing translation accuracy during the initialization stage. The fundamental concept of our method is to improve the translation estimate with a 3 Degree-of-Freedom (DoF) Bundle Adjustment (BA), independently, while the rotation estimate is fixed, instead of using ORB-SLAM3's 6-DoF BA. Additionally, the rotation estimate will be updated by considering IMU measurements and gyroscope bias, unlike ORB-SLAM3's rotation, which is directly obtained from stereo visual odometry and may yield inferior results when operating in challenging scenarios. We also conduct extensive evaluations on the public benchmark, the EuRoC dataset, demonstrating that our method excels in accuracy.