Mingyang Yi

Shuchen Xue, Mingyang Yi, Weijian Luo et al.

Diffusion Probabilistic Models (DPMs) have achieved considerable success in generation tasks. As sampling from DPMs is equivalent to solving diffusion SDE or ODE which is time-consuming, numerous fast sampling methods built upon improved differential equation solvers are proposed. The majority of such techniques consider solving the diffusion ODE due to its superior efficiency. However, stochastic sampling could offer additional advantages in generating diverse and high-quality data. In this work, we engage in a comprehensive analysis of stochastic sampling from two aspects: variance-controlled diffusion SDE and linear multi-step SDE solver. Based on our analysis, we propose \textit{SA-Solver}, which is an improved efficient stochastic Adams method for solving diffusion SDE to generate data with high quality. Our experiments show that \textit{SA-Solver} achieves: 1) improved or comparable performance compared with the existing state-of-the-art (SOTA) sampling methods for few-step sampling; 2) SOTA FID on substantial benchmark datasets under a suitable number of function evaluations (NFEs). Code is available at https://github.com/scxue/SA-Solver.

Jiaxing Wang, Deping Xiang, Jin Xu et al.

The capabilities of large language models (LLMs) significantly depend on training data drawn from various domains. Optimizing domain-specific mixture ratios can be modeled as a bi-level optimization problem, which we simplify into a single-level penalized form and solve with twin networks: a proxy model trained on primary data and a dynamically updated reference model trained with additional data. Our proposed method, Twin Networks for bi-level DatA mixturE optiMization (TANDEM), measures the data efficacy through the difference between the twin models and up-weights domains that benefit more from the additional data. TANDEM provides theoretical guarantees and wider applicability, compared to prior approaches. Furthermore, our bi-level perspective suggests new settings to study domain reweighting such as data-restricted scenarios and supervised fine-tuning, where optimized mixture ratios significantly improve the performance. Extensive experiments validate TANDEM's effectiveness in all scenarios.

Mingyang Yi, Ruoyu Wang, Jiachen Sun et al.

Recently, generalization on out-of-distribution (OOD) data with correlation shift has attracted great attentions. The correlation shift is caused by the spurious attributes that correlate to the class label, as the correlation between them may vary in training and test data. For such a problem, we show that given the class label, the models that are conditionally independent of spurious attributes are OOD generalizable. Based on this, a metric Conditional Spurious Variation (CSV) which controls the OOD generalization error, is proposed to measure such conditional independence. To improve the OOD generalization, we regularize the training process with the proposed CSV. Under mild assumptions, our training objective can be formulated as a nonconvex-concave mini-max problem. An algorithm with a provable convergence rate is proposed to solve the problem. Extensive empirical results verify our algorithm's efficacy in improving OOD generalization.

Ruoyu Wang, Mingyang Yi, Zhitang Chen et al.

In real-world applications, it is important and desirable to learn a model that performs well on out-of-distribution (OOD) data. Recently, causality has become a powerful tool to tackle the OOD generalization problem, with the idea resting on the causal mechanism that is invariant across domains of interest. To leverage the generally unknown causal mechanism, existing works assume a linear form of causal feature or require sufficiently many and diverse training domains, which are usually restrictive in practice. In this work, we obviate these assumptions and tackle the OOD problem without explicitly recovering the causal feature. Our approach is based on transformations that modify the non-causal feature but leave the causal part unchanged, which can be either obtained from prior knowledge or learned from the training data in the multi-domain scenario. Under the setting of invariant causal mechanism, we theoretically show that if all such transformations are available, then we can learn a minimax optimal model across the domains using only single domain data. Noticing that knowing a complete set of these causal invariant transformations may be impractical, we further show that it suffices to know only a subset of these transformations. Based on the theoretical findings, a regularized training procedure is proposed to improve the OOD generalization capability. Extensive experimental results on both synthetic and real datasets verify the effectiveness of the proposed algorithm, even with only a few causal invariant transformations.

Yining Zhang, Mingyang Yi, Chen Wang et al.

High-performance GPU kernels are essential for efficient LLM deployment, yet optimizing them remains expertise-intensive. Recent LLM-based code generation makes automatic GPU operator generation promising, but operator optimization remains a hardware-aware search problem. Existing LLM-based methods face a granularity mismatch: coarse hints are reusable but hard to execute, whereas detailed memories are actionable but enlarge the search space and obscure optimization bottlenecks. The key challenge is therefore to organize optimization experience at an appropriate granularity. To address this issue, this paper proposes HTAM (Hierarchical Transition-Attended Memory), a coarse-to-fine framework for LLM-based operator optimization. HTAM builds a two-level Hierarchical Transition Graph (HTG) to organize coarse global directions, detailed local strategies, and transition experience between optimization steps. During each evolution step, HTAM selects a global direction from the current state and recent optimization history, retrieves the corresponding local strategy memory, and uses it to guide concrete CUDA code generation. Experiments on the full KernelBench suite demonstrate that HTAM consistently improves correctness, fast-solution rate, and speedup over LLM-based baselines, while backend and Robust-KBench studies indicate transferable benefits from structured memory.

Haoyang Jiang, Zekun Wang, Mingyang Yi et al.

The Diffusion Probabilistic Model (DPM) achieves remarkable performance in image generation, while its increasing parameter size and computational overhead hinder its deployment in practical applications. To improve this, the existing literature focuses on obtaining a smaller model with a fixed architecture through model compression. However, in practice, DPMs usually need to be deployed on various devices with different resource constraints, which leads to multiple compression processes, incurring significant overhead for repeated training. To obviate this, we propose a once-for-all (OFA) compression framework for DPMs that yields different subnetworks with various computations in a one-shot training manner. The existing OFA framework typically involves massive subnetworks with different parameter sizes, while such a huge candidate space slows the optimization. Thus, we propose to restrict the candidate subnetworks with a certain set of parameter sizes, where each size corresponds to a specific subnetwork. Specifically, to construct each subnetwork with a given size, we gradually allocate the maintained channels by their importance. Furthermore, we propose a reweighting strategy to balance the optimization process of different subnetworks. Experimental results show that our approach can produce compressed DPMs for various sizes with significantly lower training overhead while achieving satisfactory performance.

Haoyang Jiang, Mingyang Yi, Shaolei Zhang et al.

Recently, detecting AI-generated images produced by diffusion-based models has attracted increasing attention due to their potential threat to safety. Among existing approaches, reconstruction-based methods have emerged as a prominent paradigm for this task. However, we find that such methods exhibit severe security vulnerabilities to adversarial perturbations; that is, by adding imperceptible adversarial perturbations to input images, the detection accuracy of classifiers collapses to near zero. To verify this threat, we present a systematic evaluation of the adversarial robustness of three representative detectors across four diverse generative backbone models. First, we construct adversarial attacks in white-box scenarios, which degrade the performance of all well-trained detectors. Moreover, we find that these attacks demonstrate transferability; specifically, attacks crafted against one detector can be transferred to others, indicating that adversarial attacks on detectors can also be constructed in a black-box setting. Finally, we assess common countermeasures and find that standard defense methods against adversarial attacks provide limited mitigation. We attribute these failures to the low signal-to-noise ratio (SNR) of attacked samples as perceived by the detectors. Overall, our results reveal fundamental security limitations of reconstruction-based detectors and highlight the need to rethink existing detection strategies.

Siqi Luo, Haoran Yang, Yi Xin et al.

Large pre-trained models achieve remarkable performance in vision tasks but are impractical for fine-tuning due to high computational and storage costs. Parameter-Efficient Fine-Tuning (PEFT) methods mitigate this issue by updating only a subset of parameters; however, most existing approaches are task-agnostic, failing to fully exploit task-specific adaptations, which leads to suboptimal efficiency and performance. To address this limitation, we propose Task-Relevant Parameter and Token Selection (TR-PTS), a task-driven framework that enhances both computational efficiency and accuracy. Specifically, we introduce Task-Relevant Parameter Selection, which utilizes the Fisher Information Matrix (FIM) to identify and fine-tune only the most informative parameters in a layer-wise manner, while keeping the remaining parameters frozen. Simultaneously, Task-Relevant Token Selection dynamically preserves the most informative tokens and merges redundant ones, reducing computational overhead. By jointly optimizing parameters and tokens, TR-PTS enables the model to concentrate on task-discriminative information. We evaluate TR-PTS on benchmark, including FGVC and VTAB-1k, where it achieves state-of-the-art performance, surpassing full fine-tuning by 3.40% and 10.35%, respectively. The code are available at https://github.com/synbol/TR-PTS.

Zhaochun Li, Mingyang Yi, Yue Wang et al.

Reinforcement Learning from Human Feedback (RLHF) and its variants have emerged as the dominant approaches for aligning Large Language Models with human intent. While empirically effective, the theoretical generalization properties of these methods in high-dimensional settings remain to be explored. To this end, we build the generalization theory on RLHF of LLMs under the linear reward model, through the framework of algorithmic stability. In contrast to the existing works built upon the consistency of maximum likelihood estimations on reward model, our analysis is presented under an end-to-end learning framework, which is consistent with practice. Concretely, we prove that under a key \textbf{feature coverage} condition, the empirical optima of policy model have a generalization bound of order $\mathcal{O}(n^{-\frac{1}{2}})$. Moreover, the results can be extrapolated to parameters obtained by gradient-based learning algorithms, i.e., Gradient Ascent (GA) and Stochastic Gradient Ascent (SGA). Thus, we argue that our results provide new theoretical evidence for the empirically observed generalization of LLMs after RLHF.

Zekun Wang, Mingyang Yi, Shuchen Xue et al.

Diffusion Probabilistic Models (DPMs) have achieved significant success in generative tasks. However, their training and sampling processes suffer from the issue of distribution mismatch. During the denoising process, the input data distributions differ between the training and inference stages, potentially leading to inaccurate data generation. To obviate this, we analyze the training objective of DPMs and theoretically demonstrate that this mismatch can be alleviated through Distributionally Robust Optimization (DRO), which is equivalent to performing robustness-driven Adversarial Training (AT) on DPMs. Furthermore, for the recently proposed Consistency Model (CM), which distills the inference process of the DPM, we prove that its training objective also encounters the mismatch issue. Fortunately, this issue can be mitigated by AT as well. Based on these insights, we propose to conduct efficient AT on both DPM and CM. Finally, extensive empirical studies validate the effectiveness of AT in diffusion-based models. The code is available at https://github.com/kugwzk/AT_Diff.

Mingyang Yi, Aoxue Li, Yi Xin et al.

Recently, the strong latent Diffusion Probabilistic Model (DPM) has been applied to high-quality Text-to-Image (T2I) generation (e.g., Stable Diffusion), by injecting the encoded target text prompt into the gradually denoised diffusion image generator. Despite the success of DPM in practice, the mechanism behind it remains to be explored. To fill this blank, we begin by examining the intermediate statuses during the gradual denoising generation process in DPM. The empirical observations indicate, the shape of image is reconstructed after the first few denoising steps, and then the image is filled with details (e.g., texture). The phenomenon is because the low-frequency signal (shape relevant) of the noisy image is not corrupted until the final stage in the forward process (initial stage of generation) of adding noise in DPM. Inspired by the observations, we proceed to explore the influence of each token in the text prompt during the two stages. After a series of experiments of T2I generations conditioned on a set of text prompts. We conclude that in the earlier generation stage, the image is mostly decided by the special token [\texttt{EOS}] in the text prompt, and the information in the text prompt is already conveyed in this stage. After that, the diffusion model completes the details of generated images by information from themselves. Finally, we propose to apply this observation to accelerate the process of T2I generation by properly removing text guidance, which finally accelerates the sampling up to 25\%+.

Xiuyu Li, Jinkai Zhang, Mingyang Yi et al.

Reinforcement Learning (RL) post-training alignment for language models is effective, but also costly and unstable in practice, owing to its complicated training process. To address this, we propose a training-free inference method to sample directly from the optimal RL policy. The transition probability applied to Masked Language Modeling (MLM) consists of a reference policy model and an energy term. Based on this, our algorithm, Energy-Guided Test-Time Scaling (ETS), estimates the key energy term via online Monte Carlo, with a provable convergence rate. Moreover, to ensure practical efficiency, ETS leverages modern acceleration frameworks alongside tailored importance sampling estimators, substantially reducing inference latency while provably preserving sampling quality. Experiments on MLM (including autoregressive models and diffusion language models) across reasoning, coding, and science benchmarks show that our ETS consistently improves generation quality, validating its effectiveness and design.

Yuxin Zhang, Meihao Fan, Ju Fan et al.

Recent advances in large language models (LLMs) have significantly improved performance on the Text-to-SQL task by leveraging their powerful reasoning capabilities. To enhance accuracy during the reasoning process, external Process Reward Models (PRMs) can be introduced during training and inference to provide fine-grained supervision. However, if misused, PRMs may distort the reasoning trajectory and lead to suboptimal or incorrect SQL generation. To address this challenge, we propose Reward-SQL, a framework that systematically explores how to incorporate PRMs into the Text-to-SQL reasoning process effectively. Our approach follows a "cold start, then PRM supervision" paradigm. Specifically, we first train the model to decompose SQL queries into structured stepwise reasoning chains using common table expressions (Chain-of-CTEs), establishing a strong and interpretable reasoning baseline. Then, we investigate four strategies for integrating PRMs, and find that combining PRM as an online training signal (e.g.,GRPO) with PRM-guided inference (e.g., best-of-N sampling) yields the best results. Empirically, on the BIRD benchmark, Reward-SQL enables models supervised by PRM (7B) to achieve a 13.1% performance gain across various guidance strategies. Notably, our GRPO-aligned policy model based on Qwen2.5-Coder-7B-Instruct achieves 68.9% accuracy on the BIRD development set, outperforming all baseline methods under the same model size. These results demonstrate the effectiveness of Reward-SQL in leveraging reward-based supervision for Text-to-SQL reasoning.

Xuerui Su, Yue Wang, Jinhua Zhu et al.

With the rapid development of Large Language Models (LLMs), numerous Reinforcement Learning from Human Feedback (RLHF) algorithms have been introduced to improve model safety and alignment with human preferences. These algorithms can be divided into two main frameworks based on whether they require an explicit reward (or value) function for training: actor-critic-based Proximal Policy Optimization (PPO) and alignment-based Direct Preference Optimization (DPO). The mismatch between DPO and PPO, such as DPO's use of a classification loss driven by human-preferred data, has raised confusion about whether DPO should be classified as a Reinforcement Learning (RL) algorithm. To address these ambiguities, we focus on three key aspects related to DPO, RL, and other RLHF algorithms: (1) the construction of the loss function; (2) the target distribution at which the algorithm converges; (3) the impact of key components within the loss function. Specifically, we first establish a unified framework named UDRRA connecting these algorithms based on the construction of their loss functions. Next, we uncover their target policy distributions within this framework. Finally, we investigate the critical components of DPO to understand their impact on the convergence rate. Our work provides a deeper understanding of the relationship between DPO, RL, and other RLHF algorithms, offering new insights for improving existing algorithms.

Aoxue Li, Mingyang Yi, Zhenguo Li

Recently, text-to-image (T2I) editing has been greatly pushed forward by applying diffusion models. Despite the visual promise of the generated images, inconsistencies with the expected textual prompt remain prevalent. This paper aims to systematically improve the text-guided image editing techniques based on diffusion models, by addressing their limitations. Notably, the common idea in diffusion-based editing firstly reconstructs the source image via inversion techniques e.g., DDIM Inversion. Then following a fusion process that carefully integrates the source intermediate (hidden) states (obtained by inversion) with the ones of the target image. Unfortunately, such a standard pipeline fails in many cases due to the interference of texture retention and the new characters creation in some regions. To mitigate this, we incorporate human annotation as an external knowledge to confine editing within a ``Mask-informed'' region. Then we carefully Fuse the edited image with the source image and a constructed intermediate image within the model's Self-Attention module. Extensive empirical results demonstrate the proposed ``MaSaFusion'' significantly improves the existing T2I editing techniques.

Yu Li, Mingyang Yi, Xiuyu Li et al.

Agentic Reinforcement Learning (ARL) focuses on training large language models (LLMs) to interleave reasoning with external tool execution to solve complex tasks. Most existing ARL methods train a single shared model parameters to support both reasoning and tool use behaviors, implicitly assuming that joint training leads to improved overall agent performance. Despite its widespread adoption, this assumption has rarely been examined empirically. In this paper, we systematically investigate this assumption by introducing a Linear Effect Attribution System(LEAS), which provides quantitative evidence of interference between reasoning and tool-use behaviors. Through an in-depth analysis, we show that these two capabilities often induce misaligned gradient directions, leading to training interference that undermines the effectiveness of joint optimization and challenges the prevailing ARL paradigm. To address this issue, we propose Disentangled Action Reasoning Tuning(DART), a simple and efficient framework that explicitly decouples parameter updates for reasoning and tool-use via separate low-rank adaptation modules. Experimental results show that DART consistently outperforms baseline methods with averaged 6.35 percent improvements and achieves performance comparable to multi-agent systems that explicitly separate tool-use and reasoning using a single model.

Bowei Zhu, Shaojie Li, Mingyang Yi et al.

Prior work (Klochkov $\&$ Zhivotovskiy, 2021) establishes at most $O\left(\log (n)/n\right)$ excess risk bounds via algorithmic stability for strongly-convex learners with high probability. We show that under the similar common assumptions -- - Polyak-Lojasiewicz condition, smoothness, and Lipschitz continous for losses -- - rates of $O\left(\log^2(n)/n^2\right)$ are at most achievable. To our knowledge, our analysis also provides the tightest high-probability bounds for gradient-based generalization gaps in nonconvex settings.

Mingyang Yi, Bohan Wang

Recently, optimization on the Riemannian manifold have provided valuable insights to the optimization community. In this regard, extending these methods to to the Wasserstein space is of particular interest, since optimization on Wasserstein space is closely connected to practical sampling processes. Generally, the standard (continuous) optimization method on Wasserstein space is Riemannian gradient flow (i.e., Langevin dynamics when minimizing KL divergence). In this paper, we aim to enrich the family of continuous optimization methods in the Wasserstein space, by extending the gradient flow on it into the stochastic gradient descent (SGD) flow and stochastic variance reduction gradient (SVRG) flow. By leveraging the property of Wasserstein space, we construct stochastic differential equations (SDEs) to approximate the corresponding discrete Euclidean dynamics of the desired Riemannian stochastic methods. Then, we obtain the flows in Wasserstein space by Fokker-Planck equation. Finally, we establish convergence rates of the proposed stochastic flows, which align with those known in the Euclidean setting.

Mingyang Yi, Jiacheng Sun, Zhenguo Li

The diffusion probabilistic generative models are widely used to generate high-quality data. Though they can synthetic data that does not exist in the training set, the rationale behind such generalization is still unexplored. In this paper, we formally define the generalization of the generative model, which is measured by the mutual information between the generated data and the training set. The definition originates from the intuition that the model which generates data with less correlation to the training set exhibits better generalization ability. Meanwhile, we show that for the empirical optimal diffusion model, the data generated by a deterministic sampler are all highly related to the training set, thus poor generalization. This result contradicts the observation of the trained diffusion model's (approximating empirical optima) extrapolation ability (generating unseen data). To understand this contradiction, we empirically verify the difference between the sufficiently trained diffusion model and the empirical optima. We found, though obtained through sufficient training, there still exists a slight difference between them, which is critical to making the diffusion model generalizable. Moreover, we propose another training objective whose empirical optimal solution has no potential generalization problem. We empirically show that the proposed training objective returns a similar model to the original one, which further verifies the generalization ability of the trained diffusion model.

Weiran Huang, Mingyang Yi, Xuyang Zhao et al.

Recently, self-supervised learning has attracted great attention, since it only requires unlabeled data for model training. Contrastive learning is one popular method for self-supervised learning and has achieved promising empirical performance. However, the theoretical understanding of its generalization ability is still limited. To this end, we define a kind of $(σ,δ)$-measure to mathematically quantify the data augmentation, and then provide an upper bound of the downstream classification error rate based on the measure. It reveals that the generalization ability of contrastive self-supervised learning is related to three key factors: alignment of positive samples, divergence of class centers, and concentration of augmented data. The first two factors are properties of learned representations, while the third one is determined by pre-defined data augmentation. We further investigate two canonical contrastive losses, InfoNCE and cross-correlation, to show how they provably achieve the first two factors. Moreover, we conduct experiments to study the third factor, and observe a strong correlation between downstream performance and the concentration of augmented data.

Mingyang Yi, Lu Hou, Jiacheng Sun et al.

Recently, learning a model that generalizes well on out-of-distribution (OOD) data has attracted great attention in the machine learning community. In this paper, after defining OOD generalization via Wasserstein distance, we theoretically show that a model robust to input perturbation generalizes well on OOD data. Inspired by previous findings that adversarial training helps improve input-robustness, we theoretically show that adversarially trained models have converged excess risk on OOD data, and empirically verify it on both image classification and natural language understanding tasks. Besides, in the paradigm of first pre-training and then fine-tuning, we theoretically show that a pre-trained model that is more robust to input perturbation provides a better initialization for generalization on downstream OOD data. Empirically, after fine-tuning, this better-initialized model from adversarial pre-training also has better OOD generalization.

Mingyang Yi, Lu Hou, Lifeng Shang et al.

Data augmentation is an effective technique to improve the generalization of deep neural networks. However, previous data augmentation methods usually treat the augmented samples equally without considering their individual impacts on the model. To address this, for the augmented samples from the same training example, we propose to assign different weights to them. We construct the maximal expected loss which is the supremum over any reweighted loss on augmented samples. Inspired by adversarial training, we minimize this maximal expected loss (MMEL) and obtain a simple and interpretable closed-form solution: more attention should be paid to augmented samples with large loss values (i.e., harder examples). Minimizing this maximal expected loss enables the model to perform well under any reweighting strategy. The proposed method can generally be applied on top of any data augmentation methods. Experiments are conducted on both natural language understanding tasks with token-level data augmentation, and image classification tasks with commonly-used image augmentation techniques like random crop and horizontal flip. Empirical results show that the proposed method improves the generalization performance of the model.

Mingyang Yi, Huishuai Zhang, Wei Chen et al.

It is arguably believed that flatter minima can generalize better. However, it has been pointed out that the usual definitions of sharpness, which consider either the maxima or the integral of loss over a $δ$ ball of parameters around minima, cannot give consistent measurement for scale invariant neural networks, e.g., networks with batch normalization layer. In this paper, we first propose a measure of sharpness, BN-Sharpness, which gives consistent value for equivalent networks under BN. It achieves the property of scale invariance by connecting the integral diameter with the scale of parameter. Then we present a computation-efficient way to calculate the BN-sharpness approximately i.e., one dimensional integral along the "sharpest" direction. Furthermore, we use the BN-sharpness to regularize the training and design an algorithm to minimize the new regularized objective. Our algorithm achieves considerably better performance than vanilla SGD over various experiment settings.

Mingyang Yi

Batch normalization (BN) has become a critical component across diverse deep neural networks. The network with BN is invariant to positively linear re-scale transformation, which makes there exist infinite functionally equivalent networks with different scales of weights. However, optimizing these equivalent networks with the first-order method such as stochastic gradient descent will obtain a series of iterates converging to different local optima owing to their different gradients across training. To obviate this, we propose a quotient manifold \emph{PSI manifold}, in which all the equivalent weights of the network with BN are regarded as the same element. Next, we construct gradient descent and stochastic gradient descent on the proposed PSI manifold to train the network with BN. The two algorithms guarantee that every group of equivalent weights (caused by positively re-scaling) converge to the equivalent optima. Besides that, we give convergence rates of the proposed algorithms on the PSI manifold. The results show that our methods accelerate training compared with the algorithms on the Euclidean weight space. Finally, empirical results verify that our algorithms consistently improve the existing methods in both convergence rate and generalization ability under various experimental settings.

Mingyang Yi, Ruoyu Wang, Zhi-Ming Ma

We establish upper bounds for the expected excess risk of models trained by proper iterative algorithms which approximate the local minima. Unlike the results built upon the strong globally strongly convexity or global growth conditions e.g., PL-inequality, we only require the population risk to be \emph{locally} strongly convex around its local minima. Concretely, our bound under convex problems is of order $\tilde{\cO}(1/n)$. For non-convex problems with $d$ model parameters such that $d/n$ is smaller than a threshold independent of $n$, the order of $\tilde{\cO}(1/n)$ can be maintained if the empirical risk has no spurious local minima with high probability. Moreover, the bound for non-convex problem becomes $\tilde{\cO}(1/\sqrt{n})$ without such assumption. Our results are derived via algorithmic stability and characterization of the empirical risk's landscape. Compared with the existing algorithmic stability based results, our bounds are dimensional insensitive and without restrictions on the algorithm's implementation, learning rate, and the number of iterations. Our bounds underscore that with locally strongly convex population risk, the models trained by any proper iterative algorithm can generalize well, even for non-convex problems, and $d$ is large.

Huishuai Zhang, Da Yu, Mingyang Yi et al.

We study the stability and convergence of training deep ResNets with gradient descent. Specifically, we show that the parametric branch in the residual block should be scaled down by a factor $τ=O(1/\sqrt{L})$ to guarantee stable forward/backward process, where $L$ is the number of residual blocks. Moreover, we establish a converse result that the forward process is unbounded when $τ>L^{-\frac{1}{2}+c}$, for any positive constant $c$. The above two results together establish a sharp value of the scaling factor in determining the stability of deep ResNet. Based on the stability result, we further show that gradient descent finds the global minima if the ResNet is properly over-parameterized, which significantly improves over the previous work with a much larger range of $τ$ that admits global convergence. Moreover, we show that the convergence rate is independent of the depth, theoretically justifying the advantage of ResNet over vanilla feedforward network. Empirically, with such a factor $τ$, one can train deep ResNet without normalization layer. Moreover, for ResNets with normalization layer, adding such a factor $τ$ also stabilizes the training and obtains significant performance gain for deep ResNet.

Mingyang Yi, Qi Meng, Wei Chen et al.

It was empirically confirmed by Keskar et al.\cite{SharpMinima} that flatter minima generalize better. However, for the popular ReLU network, sharp minimum can also generalize well \cite{SharpMinimacan}. The conclusion demonstrates that the existing definitions of flatness fail to account for the complex geometry of ReLU neural networks because they can't cover the Positively Scale-Invariant (PSI) property of ReLU network. In this paper, we formalize the PSI causes problem of existing definitions of flatness and propose a new description of flatness - \emph{PSI-flatness}. PSI-flatness is defined on the values of basis paths \cite{GSGD} instead of weights. Values of basis paths have been shown to be the PSI-variables and can sufficiently represent the ReLU neural networks which ensure the PSI property of PSI-flatness. Then we study the relation between PSI-flatness and generalization theoretically and empirically. First, we formulate a generalization bound based on PSI-flatness which shows generalization error decreasing with the ratio between the largest basis path value and the smallest basis path value. That is to say, the minimum with balanced values of basis paths will more likely to be flatter and generalize better. Finally. we visualize the PSI-flatness of loss surface around two learned models which indicates the minimum with smaller PSI-flatness can indeed generalize better.