Mixture of Experts

Mixture of Experts (MoE) models have emerged as a primary solution for reducing the computational cost of Large Language Models. In this work, we analyze their scaling properties, highlighting certain arbitrary assumptions present in the existing literature. In particular, we introduce a new hyperparameter, granularity, the modification of which allows for the optimal adjustment of the size of experts. Subsequently, we present scaling laws for fine-grained MoE, taking into account the number of training tokens, model size, and granularity. Using these scaling laws, we derive the optimal training configuration for a given computational budget. Furthermore, in contrast with previous works, we demonstrate that the gap in efficiency between dense and MoE models grows as we scale up the model size and training budget.

Mixture-of-Experts (MoE) is a powerful technique for enhancing the performance of neural networks while decoupling computational complexity from the number of parameters. However, despite this, scaling the number of experts requires adding more GPUs. In addition, the load imbalance in token load across experts causes unnecessary computation or straggler problems. We present ES-MoE, a novel method for efficient scaling MoE training. It offloads expert parameters to host memory and leverages pipelined expert processing to overlap GPU-CPU communication with GPU computation. It dynamically balances token loads across GPUs, improving computational efficiency. ES-MoE accelerates MoE training on a limited number of GPUs without degradation in model performance. We validate our approach on GPT-based MoE models, demonstrating 67$\times$ better scalability and up to 17.5$\times$ better throughput over existing frameworks.

Classification models are expected to perform equally well for different classes, yet in practice, there are often large gaps in their performance. This issue of class bias is widely studied in cases of datasets with sample imbalance, but is relatively overlooked in balanced datasets. In this work, we introduce the concept of spectral imbalance in features as a potential source for class disparities and study the connections between spectral imbalance and class bias in both theory and practice. To build the connection between spectral imbalance and class gap, we develop a theoretical framework for studying class disparities and derive exact expressions for the per-class error in a high-dimensional mixture model setting. We then study this phenomenon in 11 different state-of-the-art pre-trained encoders, and show how our proposed framework can be used to compare the quality of encoders, as well as evaluate and combine data augmentation strategies to mitigate the issue. Our work sheds light on the class-dependent effects of learning, and provides new insights into how state-of-the-art pre-trained features may have unknown biases that can be diagnosed through their spectra.

Recently, there has been a widespread proliferation of "expert" language models that are specialized to a specific task or domain through parameter-efficient fine-tuning. How can we recycle large collections of expert language models to improve zero-shot generalization to unseen tasks? In this work, we propose $\textbf{P}$ost-$\textbf{H}$oc $\textbf{A}$daptive $\textbf{T}$okenwise $\textbf{G}$ating $\textbf{O}$ver an $\textbf{O}$cean of $\textbf{S}$pecialized $\textbf{E}$xperts (**PHATGOOSE**), which learns to route among specialized modules that were produced through parameter-efficient fine-tuning. Unlike past methods that learn to route among specialized models, PHATGOOSE explores the possibility that zero-shot generalization will be improved if different experts can be adaptively chosen for each token and at each layer in the model. Crucially, our method is *post-hoc* - it does not require simultaneous access to the datasets used to create the specialized models and only requires a modest amount of additional compute after each expert model is trained. In experiments covering a range of specialized model collections and zero-shot generalization benchmarks, we find that PHATGOOSE outperforms past methods for post-hoc routing and, in some cases, outperforms explicit multitask training (which requires simultaneous data access). To better understand the routing strategy learned by PHATGOOSE, we perform qualitative experiments to validate that PHATGOOSE's performance stems from its ability to make adaptive per-token and per-module expert choices.

Learning with expert advice and multi-armed bandit are two classic online decision problems which differ on how the information is observed in each round of the game. We study a family of problems interpolating the two. For a vector $\mathbf{m}=(m_1,\dots,m_K)\in \mathbb N^K$, an instance of $\mathbf m$-MAB indicates that the arms are partitioned into $K$ groups and the $i$-th group contains $m_i$ arms. Once an arm is pulled, the losses of all arms in the same group are observed. We prove tight minimax regret bounds for $\mathbf m$-MAB and design an optimal PAC algorithm for its pure exploration version, $\mathbf m$-BAI, where the goal is to identify the arm with minimum loss with as few rounds as possible. We show that the minimax regret of $\mathbf m$-MAB is $\Theta\left(\sqrt{T\sum_{k=1}^K\log (m_k+1)}\right)$ and the minimum number of pulls for an $(\varepsilon,0.05)$-PAC algorithm of $\mathbf m$-BAI is $\Theta\left(\frac{1}{\varepsilon^2}\cdot \sum_{k=1}^K\log (m_k+1)\right)$. Both our upper bounds and lower bounds for $\mathbf m$-MAB can be extended to a more general setting, namely the bandit with graph feedback, in terms of the *clique cover* and related graph parameters. As consequences, we obtained tight minimax regret bounds for several families of feedback graphs.

We propose VecKM, a local point cloud geometry encoder that is descriptive and efficient to compute. VecKM leverages a unique approach by vectorizing a kernel mixture to represent the local point cloud. Such representation's descriptiveness is supported by two theorems that validate its ability to reconstruct and preserve the similarity of the local shape. Unlike existing encoders downsampling the local point cloud, VecKM constructs the local geometry encoding using all neighboring points, producing a more descriptive encoding. Moreover, VecKM is efficient to compute and scalable to large point cloud inputs: VecKM reduces the memory cost from $(n^2+nKd)$ to $(nd+np)$; and reduces the major runtime cost from computing $nK$ MLPs to $n$ MLPs, where $n$ is the size of the point cloud, $K$ is the neighborhood size, $d$ is the encoding dimension, and $p$ is a marginal factor. The efficiency is due to VecKM's unique factorizable property that eliminates the need of explicitly grouping points into neighbors. In the normal estimation task, VecKM demonstrates not only 100x faster inference speed but also highest accuracy and strongest robustness. In classification and segmentation tasks, integrating VecKM as a preprocessing module achieves consistently better performance than the PointNet, PointNet++, and point transformer baselines, and runs consistently faster by up to 10 times.

Merging various task-specific Transformer-based vision models trained on different tasks into a single unified model can execute all the tasks concurrently. Previous methods, exemplified by task arithmetic, have been proven to be both effective and scalable. Existing methods have primarily focused on seeking a static optimal solution within the original model parameter space. A notable challenge is mitigating the interference between parameters of different models, which can substantially deteriorate performance. In this paper, we propose to merge most of the parameters while upscaling the MLP of the Transformer layers to a weight-ensembling mixture of experts (MoE) module, which can dynamically integrate shared and task-specific knowledge based on the input, thereby providing a more flexible solution that can adapt to the specific needs of each instance. Our key insight is that by identifying and separating shared knowledge and task-specific knowledge, and then dynamically integrating them, we can mitigate the parameter interference problem to a great extent. We conduct the conventional multi-task model merging experiments and evaluate the generalization and robustness of our method. The results demonstrate the effectiveness of our method and provide a comprehensive understanding of our method. The code is available at https://github.com/tanganke/weight-ensembling_MoE

Learning to solve vehicle routing problems (VRPs) has garnered much attention. However, most neural solvers are only structured and trained independently on a specific problem, making them less generic and practical. In this paper, we aim to develop a unified neural solver that can cope with a range of VRP variants simultaneously. Specifically, we propose a multi-task vehicle routing solver with mixture-of-experts (MVMoE), which greatly enhances the model capacity without a proportional increase in computation. We further develop a hierarchical gating mechanism for the MVMoE, delivering a good trade-off between empirical performance and computational complexity. Experimentally, our method significantly promotes zero-shot generalization performance on 10 unseen VRP variants, and showcases decent results on the few-shot setting and real-world benchmark instances. We further conduct extensive studies on the effect of MoE configurations in solving VRPs, and observe the superiority of hierarchical gating when facing out-of-distribution data. The source code is available at: https://github.com/RoyalSkye/Routing-MVMoE.

Variational inference (VI) is a popular approach in Bayesian inference, that looks for the best approximation of the posterior distribution within a parametric family, minimizing a loss that is (typically) the reverse Kullback-Leibler (KL) divergence. Despite its empirical success, the theoretical properties of VI have only recently received attention, and is restricted to the Gaussian case. This research paper aims to contribute to the theoretical study of VI in the non-Gaussian case by investigating the setting of Mixture of Gaussians with fixed covariance. In this view, VI over this specific family can be casted as the minimization of a Mollified relative entropy, i.e. the KL between the convolution (with respect to a Gaussian kernel) of an atomic measure supported on Diracs, where the support of the atomic measure correspond to the localization of the Gaussian components, and the target distribution. Hence, solving variational inference is equivalent to optimizing the positions of the Diracs (the particles), which can be done through gradient descent and takes the form of an interacting particle system. We study two sources of error in variational inference in this context. The first is an optimization result that is a descent lemma establishing that the algorithm decreases the objective at each iteration. The second is an approximation error that upper bounds the mollified relative entropy between an optimal finite mixture and the target distribution.

Large Language Models (LLMs) exhibit strong generalization capabilities to novel tasks when prompted with language instructions and in-context demos. Since this ability sensitively depends on the quality of prompts, various methods have been explored to automate the instruction design. While these methods demonstrated promising results, they also restricted the searched prompt to one instruction. Such simplification significantly limits their capacity, as a single demo-free instruction might not be able to cover the entire complex problem space of the targeted task. To alleviate this issue, we adopt the Mixture-of-Expert paradigm and divide the problem space into a set of sub-regions; Each sub-region is governed by a specialized expert, equipped with both an instruction and a set of demos. A two-phase process is developed to construct the specialized expert for each region: (1) demo assignment: Inspired by the theoretical connection between in-context learning and kernel regression, we group demos into experts based on their semantic similarity; (2) instruction assignment: A region-based joint search of an instruction per expert complements the demos assigned to it, yielding a synergistic effect. The resulting method, codenamed Mixture-of-Prompts (MoP), achieves an average win rate of 81% against prior arts across several major benchmarks.

This paper explores test-agnostic long-tail recognition, a challenging long-tail task where the test label distributions are unknown and arbitrarily imbalanced. We argue that the variation in these distributions can be broken down hierarchically into global and local levels. The global ones reflect a broad range of diversity, while the local ones typically arise from milder changes, often focused On a particular neighbor. Traditional methods predominantly use a Mixture-of-Expert (MoE) approach, targeting a few fixed test label distributions that exhibit substantial global variations. However, the local variations are left unconsidered. To address this issue, we propose a new MoE strategy, $\mathsf{DirMixE}$, which assigns experts to different Dirichlet meta-distributions of the label distribution, each targeting a specific aspect of local variations. Additionally, the diversity among these Dirichlet meta-distributions inherently captures global variations. This dual-level approach also leads to a more stable objective function, allowing us to sample different test distributions better to quantify the mean and variance of performance outcomes. Theoretically, we show that our proposed objective benefits from enhanced generalization by virtue of the variance-based regularization. Comprehensive experiments across multiple benchmarks confirm the effectiveness of $\mathsf{DirMixE}$.

Dense-to-sparse gating mixture of experts (MoE) has recently become an effective alternative to a well-known sparse MoE. Rather than fixing the number of activated experts as in the latter model, which could limit the investigation of potential experts, the former model utilizes the temperature to control the softmax weight distribution and the sparsity of the MoE during training in order to stabilize the expert specialization. Nevertheless, while there are previous attempts to theoretically comprehend the sparse MoE, a comprehensive analysis of the dense-to-sparse gating MoE has remained elusive. Therefore, we aim to explore the impacts of the dense-to-sparse gate on the maximum likelihood estimation under the Gaussian MoE in this paper. We demonstrate that due to interactions between the temperature and other model parameters via some partial differential equations, the convergence rates of parameter estimations are slower than any polynomial rates, and could be as slow as $\mathcal{O}(1/\log(n))$, where $n$ denotes the sample size. To address this issue, we propose using a novel activation dense-to-sparse gate, which routes the output of a linear layer to an activation function before delivering them to the softmax function. By imposing linearly independence conditions on the activation function and its derivatives, we show that the parameter estimation rates are significantly improved to polynomial rates. Finally, we conduct a simulation study to empirically validate our theoretical results.

The recent rapid progress in (self) supervised learning models is in large part predicted by empirical scaling laws: a model's performance scales proportionally to its size. Analogous scaling laws remain elusive for reinforcement learning domains, however, where increasing the parameter count of a model often hurts its final performance. In this paper, we demonstrate that incorporating Mixture-of-Expert (MoE) modules, and in particular Soft MoEs (Puigcerver et al., 2023), into value-based networks results in more parameter-scalable models, evidenced by substantial performance increases across a variety of training regimes and model sizes. This work thus provides strong empirical evidence towards developing scaling laws for reinforcement learning.

This paper considers the parameter learning and data clustering problem for MLR with multiple sub-models and arbitrary mixing weights. To deal with the data streaming case, we propose an online learning algorithm to estimate the unknown parameters. By utilizing Ljung's ODE method, we establish the almost sure convergence results of this MLR problem without the traditional i.i.d. assumption on the input data for the first time. Based on the convergence property and using the classical stochastic Lyapunov function method, we also obtain the convergence rate analysis of the proposed algorithm for the first time. In addition, the data clustering can asymptotically achieve the same performance as the case with known parameters. Future work will consider how to relax the asymptotically stationary and ergodic assumption on the input data, and how to design algorithms with global convergence performance for the MLR problem.

Recent successes suggest that parameter-efficient fine-tuning of foundation models is becoming the state-of-the-art method for transfer learning in vision, gradually replacing the rich literature of alternatives such as meta-learning. In trying to harness the best of both worlds, meta-tuning introduces a subsequent optimization stage of foundation models but has so far only shown limited success and crucially tends to underperform on out-of-distribution (OOD) tasks. In this paper, we introduce Sparse MetA-Tuning (SMAT), a method inspired by sparse mixture-of-experts approaches and trained to isolate subsets of pre-trained parameters automatically for meta-tuning on each task. SMAT successfully overcomes OOD sensitivity and delivers on the promise of enhancing the transfer abilities of vision foundation models beyond parameter-efficient finetuning. We establish new state-of-the-art results on a challenging combination of Meta-Dataset augmented with additional OOD tasks in both zero-shot and gradient-based adaptation settings. In addition, we provide a thorough analysis of the superiority of learned over hand-designed sparsity patterns for sparse expert methods and the pivotal importance of the sparsity level in balancing between in-distribution and out-of-distribution generalization. Our code and models are publicly available.

Training a unified model to take multiple targets into account is a trend towards artificial general intelligence. However, how to efficiently mitigate the training conflicts among heterogeneous data collected from different domains or tasks remains under-explored. In this study, we explore to leverage Mixture of Low-rank Adapters (MoLA) to mitigate conflicts in heterogeneous data training, which requires to jointly train the multiple low-rank adapters and their shared backbone. Specifically, we introduce two variants of MoLA, namely, MoLA-Grad and MoLA-Router, to respectively handle the target-aware and target-agnostic scenarios during inference. The former uses task identifiers to assign personalized low-rank adapters to each task, disentangling task-specific knowledge towards their adapters, thereby mitigating heterogeneity conflicts. The latter uses a novel Task-wise Decorrelation (TwD) loss to intervene the router to learn oriented weight combinations of adapters to homogeneous tasks, achieving similar effects. We conduct comprehensive experiments to verify the superiority of MoLA over previous state-of-the-art methods and present in-depth analysis on its working mechanism. Source code is available at: https://github.com/MediaBrain-SJTU/MoLA

Diffusion models benefit from instillation of task-specific information into the score function to steer the sample generation towards desired properties. Such information is coined as guidance. For example, in text-to-image synthesis, text input is encoded as guidance to generate semantically aligned images. Proper guidance inputs are closely tied with the performance of diffusion models. A common observation is that strong guidance promotes a tight alignment to the task-specific information, while reduces the diversity of the generated samples. In this paper, we provide the first theoretical study towards the influence of guidance on diffusion models in the context of Gaussian mixture models. Under mild conditions, we prove that incorporating diffusion guidance not only boosts prediction confidence but also diminishes distribution diversity, leading to a reduction in the differential entropy of the output distribution. Our analysis covers the widely used DDPM and DDIM sampling schemes, and leverages comparison inequalities in differential equations as well as the Fokker-Planck equation that characterizes the evolution of probability density function, which may be of independent theoretical interest.

Mixture-of-Experts (MoE) has gained increasing popularity as a promising framework for scaling up large language models (LLMs). However, the reliability assessment of MoE lags behind its surging applications. Moreover, when transferred to new domains such as in fine-tuning MoE models sometimes underperform their dense counterparts. Motivated by the research gap and counter-intuitive phenomenon, we propose $\texttt{MoE-RBench}$, the first comprehensive assessment of SMoE reliability from three aspects: $\textit{(i)}$ safety and hallucination, $\textit{(ii)}$ resilience to adversarial attacks, and $\textit{(iii)}$ out-of-distribution robustness. Extensive models and datasets are tested to compare the MoE to dense networks from these reliability dimensions. Our empirical observations suggest that with appropriate hyperparameters, training recipes, and inference techniques, we can build the MoE model more reliably than the dense LLM. In particular, we find that the robustness of SMoE is sensitive to the basic training settings. We hope that this study can provide deeper insights into how to adapt the pre-trained MoE model to other tasks with higher-generation security, quality, and stability. Codes are available at https://github.com/UNITES-Lab/MoE-RBench.

Pre-trained Transformers inherently possess the characteristic of sparse activation, where only a small fraction of the neurons are activated for each token. While sparse activation has been explored through post-training methods, its potential in pre-training remains untapped. In this work, we first study how activation properties change during pre-training. Our examination reveals that Transformers exhibit sparse activation throughout the majority of the pre-training process while the activation correlation keeps evolving as training progresses. Leveraging this observation, we propose Switchable Sparse-Dense Learning (SSD). SSD adaptively switches between the Mixtures-of-Experts (MoE) based sparse training and the conventional dense training during the pre-training process, leveraging the efficiency of sparse training and avoiding the static activation correlation of sparse training. Compared to dense training, SSD achieves comparable performance with identical model size and reduces pre-training costs. Moreover, the models trained with SSD can be directly used as MoE models for sparse inference and achieve the same performance as dense models with up to $2\times$ faster inference speed. Codes are available at https://github.com/thunlp/moefication.

The size of spatial receptive fields, from the early 3$\times$3 convolutions in VGGNet to the recent 7$\times$7 convolutions in ConvNeXt, has always played a critical role in architecture design. In this paper, we propose a Mixture of Receptive Fields (MoRF) instead of using a single receptive field. MoRF contains the combinations of multiple receptive fields with different sizes, e.g., convolutions with different kernel sizes, which can be regarded as experts. Such an approach serves two functions: one is to select the appropriate receptive field according to the input, and the other is to expand the network capacity. Furthermore, we also introduce two types of routing mechanisms, hard routing and soft routing to automatically select the appropriate receptive field experts. In the inference stage, the selected receptive field experts are merged via re-parameterization to maintain a similar inference speed compared to the single receptive field. To demonstrate the effectiveness of MoRF, we integrate the MoRF concept into multiple architectures, e.g., ResNet and ConvNeXt. Extensive experiments show that our approach outperforms the baselines in image classification, object detection, and segmentation tasks without significantly increasing the inference time.