The rapid development of open-source large language models (LLMs) has been truly remarkable. However, the scaling law described in previous literature presents varying conclusions, which casts a dark cloud over scaling LLMs. We delve into the study of scaling laws and present our distinctive findings that facilitate scaling of large scale models in two commonly used open-source configurations, 7B and 67B. Guided by the scaling laws, we introduce DeepSeek LLM, a project dedicated to advancing open-source language models with a long-term perspective. To support the pre-training phase, we have developed a dataset that currently consists of 2 trillion tokens and is continuously expanding. We further conduct supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) on DeepSeek LLM Base models, resulting in the creation of DeepSeek Chat models. Our evaluation results demonstrate that DeepSeek LLM 67B surpasses LLaMA-2 70B on various benchmarks, particularly in the domains of code, mathematics, and reasoning. Furthermore, open-ended evaluations reveal that DeepSeek LLM 67B Chat exhibits superior performance compared to GPT-3.5.
In real-world reinforcement learning problems, the state information is often only partially observable, which breaks the basic assumption in Markov decision processes, and thus, leads to inferior performances. Partially Observable Markov Decision Processes have been introduced to explicitly take the issue into account for learning, exploration, and planning, but presenting significant computational and statistical challenges. To address these difficulties, we exploit the representation view, which leads to a coherent design framework for a practically tractable reinforcement learning algorithm upon partial observations. We provide a theoretical analysis for justifying the statistical efficiency of the proposed algorithm. We also empirically demonstrate the proposed algorithm can surpass state-of-the-art performance with partial observations across various benchmarks, therefore, pushing reliable reinforcement learning towards more practical applications.
Optimal control is notoriously difficult for stochastic nonlinear systems. Ren et al. introduced Spectral Dynamics Embedding for developing reinforcement learning methods for controlling an unknown system. It uses an infinite-dimensional feature to linearly represent the state-value function and exploits finite-dimensional truncation approximation for practical implementation. However, the finite-dimensional approximation properties in control have not been investigated even when the model is known. In this paper, we provide a tractable stochastic nonlinear control algorithm that exploits the nonlinear dynamics upon the finite-dimensional feature approximation, Spectral Dynamics Embedding Control (SDEC), with an in-depth theoretical analysis to characterize the approximation error induced by the finite-dimension truncation and statistical error induced by finite-sample approximation in both policy evaluation and policy optimization. We also empirically test the algorithm and compare the performance with Koopman-based methods and iLQR methods on the pendulum swingup problem.
Sliced Wasserstein (SW) distance suffers from redundant projections due to independent uniform random projecting directions. To partially overcome the issue, max K sliced Wasserstein (Max-K-SW) distance ($K\geq 1$), seeks the best discriminative orthogonal projecting directions. Despite being able to reduce the number of projections, the metricity of Max-K-SW cannot be guaranteed in practice due to the non-optimality of the optimization. Moreover, the orthogonality constraint is also computationally expensive and might not be effective. To address the problem, we introduce a new family of SW distances, named Markovian sliced Wasserstein (MSW) distance, which imposes a first-order Markov structure on projecting directions. We discuss various members of MSW by specifying the Markov structure including the prior distribution, the transition distribution, and the burning and thinning technique. Moreover, we investigate the theoretical properties of MSW including topological properties (metricity, weak convergence, and connection to other distances), statistical properties (sample complexity, and Monte Carlo estimation error), and computational properties (computational complexity and memory complexity). Finally, we compare MSW distances with previous SW variants in various applications such as gradient flows, color transfer, and deep generative modeling to demonstrate the favorable performance of MSW.
Deep latent variable models have achieved significant empirical successes in model-based reinforcement learning (RL) due to their expressiveness in modeling complex transition dynamics. On the other hand, it remains unclear theoretically and empirically how latent variable models may facilitate learning, planning, and exploration to improve the sample efficiency of RL. In this paper, we provide a representation view of the latent variable models for state-action value functions, which allows both tractable variational learning algorithm and effective implementation of the optimism/pessimism principle in the face of uncertainty for exploration. In particular, we propose a computationally efficient planning algorithm with UCB exploration by incorporating kernel embeddings of latent variable models. Theoretically, we establish the sample complexity of the proposed approach in the online and offline settings. Empirically, we demonstrate superior performance over current state-of-the-art algorithms across various benchmarks.
Recent advances in Transformer architecture have empowered its empirical success in various tasks across different domains. However, existing works mainly focus on improving the standard accuracy and computational cost, without considering the robustness of contaminated samples. Existing work has shown that the self-attention mechanism, which is the center of the Transformer architecture, can be viewed as a non-parametric estimator based on the well-known kernel density estimation (KDE). This motivates us to leverage the robust kernel density estimation (RKDE) in the self-attention mechanism, to alleviate the issue of the contamination of data by down-weighting the weight of bad samples in the estimation process. The modified self-attention mechanism can be incorporated into different Transformer variants. Empirical results on language modeling and image classification tasks demonstrate the effectiveness of this approach.
Sliced Wasserstein (SW) distance has been widely used in different application scenarios since it can be scaled to a large number of supports without suffering from the curse of dimensionality. The value of sliced Wasserstein distance is the average of transportation cost between one-dimensional representations (projections) of original measures that are obtained by Radon Transform (RT). Despite its efficiency in the number of supports, estimating the sliced Wasserstein requires a relatively large number of projections in high-dimensional settings. Therefore, for applications where the number of supports is relatively small compared with the dimension, e.g., several deep learning applications where the mini-batch approaches are utilized, the complexities from matrix multiplication of Radon Transform become the main computational bottleneck. To address this issue, we propose to derive projections by linearly and randomly combining a smaller number of projections which are named bottleneck projections. We explain the usage of these projections by introducing Hierarchical Radon Transform (HRT) which is constructed by applying Radon Transform variants recursively. We then formulate the approach into a new metric between measures, named Hierarchical Sliced Wasserstein (HSW) distance. By proving the injectivity of HRT, we derive the metricity of HSW. Moreover, we investigate the theoretical properties of HSW including its connection to SW variants and its computational and sample complexities. Finally, we compare the computational cost and generative quality of HSW with the conventional SW on the task of deep generative modeling using various benchmark datasets including CIFAR10, CelebA, and Tiny ImageNet.
Representation learning often plays a critical role in reinforcement learning by managing the curse of dimensionality. A representative class of algorithms exploits a spectral decomposition of the stochastic transition dynamics to construct representations that enjoy strong theoretical properties in an idealized setting. However, current spectral methods suffer from limited applicability because they are constructed for state-only aggregation and derived from a policy-dependent transition kernel, without considering the issue of exploration. To address these issues, we propose an alternative spectral method, Spectral Decomposition Representation (SPEDER), that extracts a state-action abstraction from the dynamics without inducing spurious dependence on the data collection policy, while also balancing the exploration-versus-exploitation trade-off during learning. A theoretical analysis establishes the sample efficiency of the proposed algorithm in both the online and offline settings. In addition, an experimental investigation demonstrates superior performance over current state-of-the-art algorithms across several benchmarks.
It is common to address the curse of dimensionality in Markov decision processes (MDPs) by exploiting low-rank representations. This motivates much of the recent theoretical study on linear MDPs. However, most approaches require a given representation under unrealistic assumptions about the normalization of the decomposition or introduce unresolved computational challenges in practice. Instead, we consider an alternative definition of linear MDPs that automatically ensures normalization while allowing efficient representation learning via contrastive estimation. The framework also admits confidence-adjusted index algorithms, enabling an efficient and principled approach to incorporating optimism or pessimism in the face of uncertainty. To the best of our knowledge, this provides the first practical representation learning method for linear MDPs that achieves both strong theoretical guarantees and empirical performance. Theoretically, we prove that the proposed algorithm is sample efficient in both the online and offline settings. Empirically, we demonstrate superior performance over existing state-of-the-art model-based and model-free algorithms on several benchmarks.
* ICML 2022. The first two authors contribute equally