In this short note, we give the convergence analysis of the policy in the recent famous policy mirror descent (PMD). We mainly consider the unregularized setting following [11] with generalized Bregman divergence. The difference is that we directly give the convergence rates of policy under generalized Bregman divergence. Our results are inspired by the convergence of value function in previous works and are an extension study of policy mirror descent. Though some results have already appeared in previous work, we further discover a large body of Bregman divergences could give finite-step convergence to an optimal policy, such as the classical Euclidean distance.

Mixture-of-Experts (MoE) architectures have recently gained increasing popularity within the domain of large language models (LLMs) due to their ability to significantly reduce training and inference overhead. However, MoE architectures face challenges, such as significant disparities in the number of tokens assigned to each expert and a tendency toward homogenization among experts, which adversely affects the model's semantic generation capabilities. In this paper, we introduce LocMoE+, a refined version of the low-overhead LocMoE, incorporating the following enhancements: (1) Quantification and definition of the affinity between experts and tokens. (2) Implementation of a global-level adaptive routing strategy to rearrange tokens based on their affinity scores. (3) Reestimation of the lower bound for expert capacity, which has been shown to progressively decrease as the token feature distribution evolves. Experimental results demonstrate that, without compromising model convergence or efficacy, the number of tokens each expert processes can be reduced by over 60%. Combined with communication optimizations, this leads to an average improvement in training efficiency ranging from 5.4% to 46.6%. After fine-tuning, LocMoE+ exhibits a performance improvement of 9.7% to 14.1% across the GDAD, C-Eval, and TeleQnA datasets.

We study finite-sum distributed optimization problems involving a master node and $n-1$ local nodes under the popular $\delta$-similarity and $\mu$-strong convexity conditions. We propose two new algorithms, SVRS and AccSVRS, motivated by previous works. The non-accelerated SVRS method combines the techniques of gradient sliding and variance reduction and achieves a better communication complexity of $\tilde{\mathcal{O}}(n {+} \sqrt{n}\delta/\mu)$ compared to existing non-accelerated algorithms. Applying the framework proposed in Katyusha X, we also develop a directly accelerated version named AccSVRS with the $\tilde{\mathcal{O}}(n {+} n^{3/4}\sqrt{\delta/\mu})$ communication complexity. In contrast to existing results, our complexity bounds are entirely smoothness-free and exhibit superiority in ill-conditioned cases. Furthermore, we establish a nearly matched lower bound to verify the tightness of our AccSVRS method.

In this paper, we follow Eftekhari's work to give a non-local convergence analysis of deep linear networks. Specifically, we consider optimizing deep linear networks which have a layer with one neuron under quadratic loss. We describe the convergent point of trajectories with arbitrary starting point under gradient flow, including the paths which converge to one of the saddle points or the original point. We also show specific convergence rates of trajectories that converge to the global minimizer by stages. To achieve these results, this paper mainly extends the machinery in Eftekhari's work to provably identify the rank-stable set and the global minimizer convergent set. We also give specific examples to show the necessity of our definitions. Crucially, as far as we know, our results appear to be the first to give a non-local global analysis of linear neural networks from arbitrary initialized points, rather than the lazy training regime which has dominated the literature of neural networks, and restricted benign initialization in Eftekhari's work. We also note that extending our results to general linear networks without one hidden neuron assumption remains a challenging open problem.

We consider the fundamental problem of learning linear predictors (i.e., separable datasets with zero margin) using neural networks with gradient flow or gradient descent. Under the assumption of spherically symmetric data distribution, we show directional convergence guarantees with exact convergence rate for two-layer non-linear networks with only two hidden nodes, and (deep) linear networks. Moreover, our discovery is built on dynamic from the initialization without both initial loss and perfect classification constraint in contrast to previous works. We also point out and study the challenges in further strengthening and generalizing our results.

Deep neural networks have achieved remarkable empirical successes in the domains of computer vision, speech recognition, and natural language processing, sparking an interest in the theory behind their architectures and training. However, deep neural networks are often found to be highly overparameterized making them computationally expensive with large amounts of memory and computational power, which may take up to weeks on a modern multi-GPU server for large datasets such as ImageNet Deng et al. [7]. Hence, they are often unsuitable for smaller devices like embedded electronics, and there is a pressing demand for techniques to optimize models with reduced model size, faster inference and lower power consumption. Sparse networks, that is, neural networks in which a large subset of the model parameters are zero, have emerged as one of the leading approaches for reducing model parameter count. It has been shown empirically that deep neural networks can achieve state-of-the-art results under high levels of sparsity [17, 14, 22].

Quantized Neural Networks (QNNs) use low bit-width fixed-point numbers for representing weight parameters and activations, and are often used in real-world applications due to their saving of computation resources and reproducibility of results. Batch Normalization (BN) poses a challenge for QNNs for requiring floating points in reciprocal operations, and previous QNNs either require computing BN at high precision or revise BN to some variants in heuristic ways. In this work, we propose a novel method to quantize BN by converting an affine transformation of two floating points to a fixed-point operation with shared quantized scale, which is friendly for hardware acceleration and model deployment. We confirm that our method maintains same outputs through rigorous theoretical analysis and numerical analysis. Accuracy and efficiency of our quantization method are verified by experiments at layer level on CIFAR and ImageNet datasets. We also believe that our method is potentially useful in other problems involving quantization.

Numerous empirical evidence has corroborated that the noise plays a crucial rule in effective and efficient training of neural networks. The theory behind, however, is still largely unknown. This paper studies this fundamental problem through training a simple two-layer convolutional neural network model. Although training such a network requires solving a nonconvex optimization problem with a spurious local optimum and a global optimum, we prove that perturbed gradient descent and perturbed mini-batch stochastic gradient algorithms in conjunction with noise annealing is guaranteed to converge to a global optimum in polynomial time with arbitrary initialization. This implies that the noise enables the algorithm to efficiently escape from the spurious local optimum. Numerical experiments are provided to support our theory.

Stochastic variance-reduced gradient (SVRG) is a classical optimization method. Although it is theoretically proved to have better convergence performance than stochastic gradient descent (SGD), the generalization performance of SVRG remains open. In this paper we investigate the effects of some training techniques, mini-batching and learning rate decay, on the generalization performance of SVRG, and verify the generalization performance of Batch-SVRG (B-SVRG). In terms of the relationship between optimization and generalization, we believe that the average norm of gradients on each training sample as well as the norm of average gradient indicate how flat the landscape is and how well the model generalizes. Based on empirical observations of such metrics, we perform a sign switch on B-SVRG and derive a practical algorithm, BatchPlus-SVRG (BP-SVRG), which is numerically shown to enjoy better generalization performance than B-SVRG, even SGD in some scenarios of deep neural networks.