Neural networks trained to minimize the logistic (a.k.a. cross-entropy) loss with gradient-based methods are observed to perform well in many supervised classification tasks. Towards understanding this phenomenon, we analyze the training and generalization behavior of infinitely wide two-layer neural networks with homogeneous activations. We show that the limits of the gradient flow on exponentially tailed losses can be fully characterized as a max-margin classifier in a certain non-Hilbertian space of functions. In presence of hidden low-dimensional structures, the resulting margin is independent of the ambiant dimension, which leads to strong generalization bounds. In contrast, training only the output layer implicitly solves a kernel support vector machine, which a priori does not enjoy such an adaptivity. Our analysis of training is non-quantitative in terms of running time but we prove computational guarantees in simplified settings by showing equivalences with online mirror descent. Finally, numerical experiments suggest that our analysis describes well the practical behavior of two-layer neural networks with ReLU activations and confirm the statistical benefits of this implicit bias.
translated by 谷歌翻译
过度分化的深网络的泛化神秘具有有动力的努力,了解梯度下降(GD)如何收敛到概括井的低损耗解决方案。现实生活中的神经网络从小随机值初始化,并以分类的“懒惰”或“懒惰”或“NTK”的训练训练,分析更成功,以及最近的结果序列(Lyu和Li ,2020年; Chizat和Bach,2020; Ji和Telgarsky,2020)提供了理论证据,即GD可以收敛到“Max-ramin”解决方案,其零损失可能呈现良好。但是,仅在某些环境中证明了余量的全球最优性,其中神经网络无限或呈指数级宽。目前的纸张能够为具有梯度流动训练的两层泄漏的Relu网,无论宽度如何,都能为具有梯度流动的双层泄漏的Relu网建立这种全局最优性。分析还为最近的经验研究结果(Kalimeris等,2019)给出了一些理论上的理由,就GD的所谓简单的偏见为线性或其他“简单”的解决方案,特别是在训练中。在悲观方面,该论文表明这种结果是脆弱的。简单的数据操作可以使梯度流量会聚到具有次优裕度的线性分类器。
translated by 谷歌翻译
In a series of recent theoretical works, it was shown that strongly overparameterized neural networks trained with gradient-based methods could converge exponentially fast to zero training loss, with their parameters hardly varying. In this work, we show that this "lazy training" phenomenon is not specific to overparameterized neural networks, and is due to a choice of scaling, often implicit, that makes the model behave as its linearization around the initialization, thus yielding a model equivalent to learning with positive-definite kernels. Through a theoretical analysis, we exhibit various situations where this phenomenon arises in non-convex optimization and we provide bounds on the distance between the lazy and linearized optimization paths. Our numerical experiments bring a critical note, as we observe that the performance of commonly used non-linear deep convolutional neural networks in computer vision degrades when trained in the lazy regime. This makes it unlikely that "lazy training" is behind the many successes of neural networks in difficult high dimensional tasks.
translated by 谷歌翻译
Many tasks in machine learning and signal processing can be solved by minimizing a convex function of a measure. This includes sparse spikes deconvolution or training a neural network with a single hidden layer. For these problems, we study a simple minimization method: the unknown measure is discretized into a mixture of particles and a continuous-time gradient descent is performed on their weights and positions. This is an idealization of the usual way to train neural networks with a large hidden layer. We show that, when initialized correctly and in the many-particle limit, this gradient flow, although non-convex, converges to global minimizers. The proof involves Wasserstein gradient flows, a by-product of optimal transport theory. Numerical experiments show that this asymptotic behavior is already at play for a reasonable number of particles, even in high dimension.
translated by 谷歌翻译
了解通过随机梯度下降(SGD)训练的神经网络的特性是深度学习理论的核心。在这项工作中,我们采取了平均场景,并考虑通过SGD培训的双层Relu网络,以实现一个非变量正则化回归问题。我们的主要结果是SGD偏向于简单的解决方案:在收敛时,Relu网络实现输入的分段线性图,以及“结”点的数量 - 即,Relu网络估计器的切线变化的点数 - 在两个连续的训练输入之间最多三个。特别地,随着网络的神经元的数量,通过梯度流的解决方案捕获SGD动力学,并且在收敛时,重量的分布方法接近相关的自由能量的独特最小化器,其具有GIBBS形式。我们的主要技术贡献在于分析了这一最小化器产生的估计器:我们表明其第二阶段在各地消失,除了代表“结”要点的一些特定地点。我们还提供了经验证据,即我们的理论预测的不同可能发生与数据点不同的位置的结。
translated by 谷歌翻译
We consider the idealized setting of gradient flow on the population risk for infinitely wide two-layer ReLU neural networks (without bias), and study the effect of symmetries on the learned parameters and predictors. We first describe a general class of symmetries which, when satisfied by the target function $f^*$ and the input distribution, are preserved by the dynamics. We then study more specific cases. When $f^*$ is odd, we show that the dynamics of the predictor reduces to that of a (non-linearly parameterized) linear predictor, and its exponential convergence can be guaranteed. When $f^*$ has a low-dimensional structure, we prove that the gradient flow PDE reduces to a lower-dimensional PDE. Furthermore, we present informal and numerical arguments that suggest that the input neurons align with the lower-dimensional structure of the problem.
translated by 谷歌翻译
在负面的感知问题中,我们给出了$ n $数据点$({\ boldsymbol x} _i,y_i)$,其中$ {\ boldsymbol x} _i $是$ d $ -densional vector和$ y_i \ in \ { + 1,-1 \} $是二进制标签。数据不是线性可分离的,因此我们满足自己的内容,以找到最大的线性分类器,具有最大的\ emph {否定}余量。换句话说,我们想找到一个单位常规矢量$ {\ boldsymbol \ theta} $,最大化$ \ min_ {i \ le n} y_i \ langle {\ boldsymbol \ theta},{\ boldsymbol x} _i \ rangle $ 。这是一个非凸优化问题(它相当于在Polytope中找到最大标准矢量),我们在两个随机模型下研究其典型属性。我们考虑比例渐近,其中$ n,d \ to \ idty $以$ n / d \ to \ delta $,并在最大边缘$ \ kappa _ {\ text {s}}(\ delta)上证明了上限和下限)$或 - 等效 - 在其逆函数$ \ delta _ {\ text {s}}(\ kappa)$。换句话说,$ \ delta _ {\ text {s}}(\ kappa)$是overparametization阈值:以$ n / d \ le \ delta _ {\ text {s}}(\ kappa) - \ varepsilon $一个分类器实现了消失的训练错误,具有高概率,而以$ n / d \ ge \ delta _ {\ text {s}}(\ kappa)+ \ varepsilon $。我们在$ \ delta _ {\ text {s}}(\ kappa)$匹配,以$ \ kappa \ to - \ idty $匹配。然后,我们分析了线性编程算法来查找解决方案,并表征相应的阈值$ \ delta _ {\ text {lin}}(\ kappa)$。我们观察插值阈值$ \ delta _ {\ text {s}}(\ kappa)$和线性编程阈值$ \ delta _ {\ text {lin {lin}}(\ kappa)$之间的差距,提出了行为的问题其他算法。
translated by 谷歌翻译
We consider the problem of estimating the optimal transport map between a (fixed) source distribution $P$ and an unknown target distribution $Q$, based on samples from $Q$. The estimation of such optimal transport maps has become increasingly relevant in modern statistical applications, such as generative modeling. At present, estimation rates are only known in a few settings (e.g. when $P$ and $Q$ have densities bounded above and below and when the transport map lies in a H\"older class), which are often not reflected in practice. We present a unified methodology for obtaining rates of estimation of optimal transport maps in general function spaces. Our assumptions are significantly weaker than those appearing in the literature: we require only that the source measure $P$ satisfies a Poincar\'e inequality and that the optimal map be the gradient of a smooth convex function that lies in a space whose metric entropy can be controlled. As a special case, we recover known estimation rates for bounded densities and H\"older transport maps, but also obtain nearly sharp results in many settings not covered by prior work. For example, we provide the first statistical rates of estimation when $P$ is the normal distribution and the transport map is given by an infinite-width shallow neural network.
translated by 谷歌翻译
We consider neural networks with a single hidden layer and non-decreasing positively homogeneous activation functions like the rectified linear units. By letting the number of hidden units grow unbounded and using classical non-Euclidean regularization tools on the output weights, they lead to a convex optimization problem and we provide a detailed theoretical analysis of their generalization performance, with a study of both the approximation and the estimation errors. We show in particular that they are adaptive to unknown underlying linear structures, such as the dependence on the projection of the input variables onto a low-dimensional subspace. Moreover, when using sparsity-inducing norms on the input weights, we show that high-dimensional non-linear variable selection may be achieved, without any strong assumption regarding the data and with a total number of variables potentially exponential in the number of observations. However, solving this convex optimization problem in infinite dimensions is only possible if the non-convex subproblem of addition of a new unit can be solved efficiently. We provide a simple geometric interpretation for our choice of activation functions and describe simple conditions for convex relaxations of the finite-dimensional non-convex subproblem to achieve the same generalization error bounds, even when constant-factor approximations cannot be found. We were not able to find strong enough convex relaxations to obtain provably polynomial-time algorithms and leave open the existence or non-existence of such tractable algorithms with non-exponential sample complexities.
translated by 谷歌翻译
过分分度化是没有凸起的关键因素,以解释神经网络的全局渐变(GD)的全局融合。除了研究良好的懒惰政权旁边,已经为浅网络开发了无限宽度(平均场)分析,使用凸优化技术。为了弥合懒惰和平均场制度之间的差距,我们研究残留的网络(RESNET),其中残留块具有线性参数化,同时仍然是非线性的。这种Resnets承认无限深度和宽度限制,在再现内核Hilbert空间(RKHS)中编码残差块。在这个限制中,我们证明了当地的Polyak-Lojasiewicz不等式。因此,每个关键点都是全球最小化器和GD的局部收敛结果,并检索懒惰的制度。与其他平均场研究相比,它在残留物的表达条件下适用于参数和非参数案。我们的分析导致实用和量化的配方:从通用RKHS开始,应用随机傅里叶特征来获得满足我们的表征条件的高概率的有限维参数化。
translated by 谷歌翻译
尽管他们的超大容量过度装备能力,但是由特定优化算法训练的深度神经网络倾向于概括到看不见的数据。最近,研究人员通过研究优化算法的隐式正则化效果来解释它。卓越的进展是工作(Lyu&Li,2019),其证明了梯度下降(GD)最大化了均匀深神经网络的余量。除GD外,诸如Adagrad,RMSProp和Adam之类的自适应算法由于其快速培训过程而流行。然而,仍然缺乏适应性优化算法的概括的理论保证。在本文中,我们研究了自适应优化算法的隐式正则化,当它们在均匀深神经网络上优化逻辑损失时。我们证明了在调节器(如亚当和RMSProp)中采用指数移动平均策略的自适应算法可以最大化神经网络的余量,而Adagrad直接在调节器中总和历史平方梯度。它表明了调节剂设计中指数移动平均策略的概括的优越性。从技术上讲,我们提供统一的框架,通过构建新的自适应梯度流量和代理余量来分析自适应优化算法的会聚方向。我们的实验可以很好地支持适应性优化算法的会聚方向的理论发现。
translated by 谷歌翻译
We consider the constrained sampling problem where the goal is to sample from a distribution $\pi(x)\propto e^{-f(x)}$ and $x$ is constrained on a convex body $\mathcal{C}\subset \mathbb{R}^d$. Motivated by penalty methods from optimization, we propose penalized Langevin Dynamics (PLD) and penalized Hamiltonian Monte Carlo (PHMC) that convert the constrained sampling problem into an unconstrained one by introducing a penalty function for constraint violations. When $f$ is smooth and the gradient is available, we show $\tilde{\mathcal{O}}(d/\varepsilon^{10})$ iteration complexity for PLD to sample the target up to an $\varepsilon$-error where the error is measured in terms of the total variation distance and $\tilde{\mathcal{O}}(\cdot)$ hides some logarithmic factors. For PHMC, we improve this result to $\tilde{\mathcal{O}}(\sqrt{d}/\varepsilon^{7})$ when the Hessian of $f$ is Lipschitz and the boundary of $\mathcal{C}$ is sufficiently smooth. To our knowledge, these are the first convergence rate results for Hamiltonian Monte Carlo methods in the constrained sampling setting that can handle non-convex $f$ and can provide guarantees with the best dimension dependency among existing methods with deterministic gradients. We then consider the setting where unbiased stochastic gradients are available. We propose PSGLD and PSGHMC that can handle stochastic gradients without Metropolis-Hasting correction steps. When $f$ is strongly convex and smooth, we obtain an iteration complexity of $\tilde{\mathcal{O}}(d/\varepsilon^{18})$ and $\tilde{\mathcal{O}}(d\sqrt{d}/\varepsilon^{39})$ respectively in the 2-Wasserstein distance. For the more general case, when $f$ is smooth and non-convex, we also provide finite-time performance bounds and iteration complexity results. Finally, we test our algorithms on Bayesian LASSO regression and Bayesian constrained deep learning problems.
translated by 谷歌翻译
这项工作确立了梯度流量(GF)和随机梯度下降(SGD)的低测试误差(SGD)在具有标准初始化的两层relu网络上,在三个方案中,关键的重量集很少旋转(自然要么是由于GF和SGD,要么是由于GF和SGD,或达到人为的约束),并利用边缘作为核心分析技术。第一个制度几乎是初始化的,特别是直到权重以$ \ mathcal {o}(\ sqrt m)$移动为止,其中$ m $表示网络宽度,这与$ \ mathcal {o}(O}(O}(O})形成鲜明对比) 1)神经切线内核(NTK)允许的重量运动;在这里显示,GF和SGD仅需要网络宽度和样本数量与NTK边缘成反比,此外,GF至少达到了NTK保证金本身,这足以建立避免距离范围目标的不良KKT点的逃脱,该点的距离逃脱了。而先前的工作只能确定不折扣但任意的边缘。第二个制度是神经塌陷(NC)设置,其中数据在于极度隔离的组中,样品复杂性尺度与组数。在这里,先前工作的贡献是对初始化的整个GF轨迹的分析。最后,如果内层的权重限制为仅在规范中变化并且无法旋转,则具有较大宽度的GF达到了全球最大边缘,并且其样品复杂度与它们的逆尺度相比;这与先前的工作相反,后者需要无限的宽度和一个棘手的双收敛假设。作为纯粹的技术贡献,这项工作开发了各种潜在功能和其他工具,希望有助于未来的工作。
translated by 谷歌翻译
了解随机梯度下降(SGD)的隐式偏见是深度学习的关键挑战之一,尤其是对于过度透明的模型,损失功能的局部最小化$ l $可以形成多种多样的模型。从直觉上讲,SGD $ \ eta $的学习率很小,SGD跟踪梯度下降(GD),直到它接近这种歧管为止,梯度噪声阻止了进一步的收敛。在这样的政权中,Blanc等人。 (2020)证明,带有标签噪声的SGD局部降低了常规术语,损失的清晰度,$ \ mathrm {tr} [\ nabla^2 l] $。当前的论文通过调整Katzenberger(1991)的想法提供了一个总体框架。它原则上允许使用随机微分方程(SDE)描述参数的限制动力学的SGD围绕此歧管的正规化效应(即“隐式偏见”)的正则化效应,这是由损失共同确定的功能和噪声协方差。这产生了一些新的结果:(1)与Blanc等人的局部分析相比,对$ \ eta^{ - 2} $ steps有效的隐性偏差进行了全局分析。 (2020)仅适用于$ \ eta^{ - 1.6} $ steps和(2)允许任意噪声协方差。作为一个应用程序,我们以任意大的初始化显示,标签噪声SGD始终可以逃脱内核制度,并且仅需要$ o(\ kappa \ ln d)$样本用于学习$ \ kappa $ -sparse $ -sparse yroverparame parametrized linearized Linear Modal in $ \ Mathbb {r}^d $(Woodworth等,2020),而GD在内核制度中初始化的GD需要$ \ omega(d)$样本。该上限是最小值的最佳,并改善了先前的$ \ tilde {o}(\ kappa^2)$上限(Haochen等,2020)。
translated by 谷歌翻译
通过在线规范相关性分析的问题,我们提出了\ emph {随机缩放梯度下降}(SSGD)算法,以最小化通用riemannian歧管上的随机功能的期望。 SSGD概括了投影随机梯度下降的思想,允许使用缩放的随机梯度而不是随机梯度。在特殊情况下,球形约束的特殊情况,在广义特征向量问题中产生的,我们建立了$ \ sqrt {1 / t} $的令人反感的有限样本,并表明该速率最佳最佳,直至具有积极的积极因素相关参数。在渐近方面,一种新的轨迹平均争论使我们能够实现局部渐近常态,其速率与鲁普特 - Polyak-Quaditsky平均的速率匹配。我们将这些想法携带在一个在线规范相关分析,从事文献中的第一次获得了最佳的一次性尺度算法,其具有局部渐近融合到正常性的最佳一次性尺度算法。还提供了用于合成数据的规范相关分析的数值研究。
translated by 谷歌翻译
我们检查了在未注册的逻辑回归问题上的梯度下降,并在线性可分离数据集上具有均匀的线性预测指标。我们显示了预测变量收敛到最大边缘(硬边缘SVM)解决方案的方向。结果还推广到其他单调的损失函数,在无穷大时降低了损失功能,多级问题,并在某个受限的环境中训练在深网中的重量层。此外,我们表明这种融合非常慢,只有在损失本身的融合中的对数。这可以有助于解释即使训练错误为零,并且训练损失非常小,并且正如我们所显示的,即使验证损失增加了,也可以继续优化逻辑或跨透明度损失的好处。我们的方法还可以帮助理解隐式正则化n更复杂的模型以及其他优化方法。
translated by 谷歌翻译
尽管有许多有吸引力的财产,但内核方法受到维度的诅咒受到严重影响。例如,在$ \ mathbb {r} ^ d $的内部产品内核的情况下,再现内核希尔伯特空间(RKHS)规范对于依赖于小方向子集(RIDGE函数)的功能往往非常大。相应地,使用内核方法难以学习这样的功能。这种观察结果有动力研究内核方法的概括,由此rkhs规范 - 它等同于加权$ \ ell_2 $ norm - 被加权函数$ \ ell_p $ norm替换,我们将其称为$ \ mathcal {f} _p $ norm。不幸的是,这些方法的陶油是不清楚的。内核技巧不可用,最大限度地减少这些规范要求解决无限维凸面问题。我们将随机特征近似于这些规范,表明,对于$ p> 1 $,近似于原始学习问题所需的随机功能的数量是由样本大小的多项式的上限。因此,使用$ \ mathcal {f} _p $ norms在这些情况下是易行的。我们介绍了一种基于双重均匀浓度的证明技术,这可以对超分子化模型的研究更广泛。对于$ p = 1 $,我们对随机功能的保证近似分解。我们证明了使用$ \ mathcal {f} _1 $ norm的学习是在随机减少的$ \ mathsf {np} $ - 基于噪音的半个空间问题的问题。
translated by 谷歌翻译
在本文中,我们重新审视了私人经验风险最小化(DP-erm)和差异私有随机凸优化(DP-SCO)的问题。我们表明,来自统计物理学(Langevin Exfusion(LD))的经过良好研究的连续时间算法同时为DP-SCO和DP-SCO提供了最佳的隐私/实用性权衡,$ \ epsilon $ -DP和$ $ \ epsilon $ -DP和$ (\ epsilon,\ delta)$ - dp均用于凸和强烈凸损失函数。我们为LD提供新的时间和尺寸独立统一稳定性,并使用我们为$ \ epsilon $ -DP提供相应的最佳超额人口风险保证。 $ \ epsilon $ -DP的DP-SCO保证的一个重要属性是,它们将非私人最佳界限匹配为$ \ epsilon \与\ infty $。在此过程中,我们提供了各种技术工具,这些工具可能引起独立的关注:i)在两个相邻数据集上运行损失功能时,一个新的r \'enyi Divergence绑定了LD,ii)最后一个过多的经验风险范围迭代LD,类似于Shamir和Zhang的嘈杂随机梯度下降(SGD)和iii)的LD,对LD进行了两期多余的风险分析,其中第一阶段是当扩散在任何合理意义上都没有在任何合理意义上融合到固定分布时,在第二阶段扩散已收敛到吉布斯分布的变体。我们的普遍性结果至关重要地依赖于LD的动力学。当它融合到固定分布时,我们获得了$ \ epsilon $ -DP的最佳界限。当它仅在很短的时间内运行$ \ propto 1/p $时,我们在$(\ epsilon,\ delta)$ -DP下获得最佳界限。在这里,$ p $是模型空间的维度。
translated by 谷歌翻译
非线性自适应控制理论中的一个关键假设是系统的不确定性可以在一组已知基本函数的线性跨度中表示。虽然该假设导致有效的算法,但它将应用限制为非常特定的系统类别。我们介绍一种新的非参数自适应算法,其在参数上学习无限尺寸密度,以取消再现内核希尔伯特空间中的未知干扰。令人惊讶的是,所产生的控制输入承认,尽管其底层无限尺寸结构,但是尽管它的潜在无限尺寸结构实现了其实施的分析表达。虽然这种自适应输入具有丰富和富有敏感性的 - 例如,传统的线性参数化 - 其计算复杂性随时间线性增长,使其比其参数对应力相对较高。利用随机傅里叶特征的理论,我们提供了一种有效的随机实现,该实现恢复了经典参数方法的复杂性,同时可透明地保留非参数输入的表征性。特别地,我们的显式范围仅取决于系统的基础参数,允许我们所提出的算法有效地缩放到高维系统。作为该方法的说明,我们展示了随机近似算法学习由牛顿重力交互的十点批量组成的60维系统的预测模型的能力。
translated by 谷歌翻译
变性推理(VI)为基于传统的采样方法提供了一种吸引人的替代方法,用于实施贝叶斯推断,因为其概念性的简单性,统计准确性和计算可扩展性。然而,常见的变分近似方案(例如平均场(MF)近似)需要某些共轭结构以促进有效的计算,这可能会增加不必要的限制对可行的先验分布家族,并对变异近似族对差异进行进一步的限制。在这项工作中,我们开发了一个通用计算框架,用于实施MF-VI VIA WASSERSTEIN梯度流(WGF),这是概率度量空间上的梯度流。当专门针对贝叶斯潜在变量模型时,我们将分析基于时间消化的WGF交替最小化方案的算法收敛,用于实现MF近似。特别是,所提出的算法类似于EM算法的分布版本,包括更新潜在变量变异分布的E step以及在参数的变异分布上进行最陡峭下降的m step。我们的理论分析依赖于概率度量空间中的最佳运输理论和细分微积分。我们证明了时间限制的WGF的指数收敛性,以最大程度地减少普通大地测量学严格的凸度的通用物镜功能。我们还提供了通过使用时间限制的WGF的固定点方程从MF近似获得的变异分布的指数收缩的新证明。我们将方法和理论应用于两个经典的贝叶斯潜在变量模型,即高斯混合模型和回归模型的混合物。还进行了数值实验,以补充这两个模型下的理论发现。
translated by 谷歌翻译