We investigate gradient descent training of wide neural networks and the corresponding implicit bias in function space. For univariate regression, we show that the solution of training a width-$n$ shallow ReLU network is within $n^{- 1/2}$ of the function which fits the training data and whose difference from the initial function has the smallest 2-norm of the second derivative weighted by a curvature penalty that depends on the probability distribution that is used to initialize the network parameters. We compute the curvature penalty function explicitly for various common initialization procedures. For instance, asymmetric initialization with a uniform distribution yields a constant curvature penalty, and thence the solution function is the natural cubic spline interpolation of the training data. \hj{For stochastic gradient descent we obtain the same implicit bias result.} We obtain a similar result for different activation functions. For multivariate regression we show an analogous result, whereby the second derivative is replaced by the Radon transform of a fractional Laplacian. For initialization schemes that yield a constant penalty function, the solutions are polyharmonic splines. Moreover, we show that the training trajectories are captured by trajectories of smoothing splines with decreasing regularization strength.
相關內容
Recent works in learning-integrated optimization have shown promise in settings where the optimization problem is only partially observed or where general-purpose optimizers perform poorly without expert tuning. By learning an optimizer $\mathbf{g}$ to tackle these challenging problems with $f$ as the objective, the optimization process can be substantially accelerated by leveraging past experience. The optimizer can be trained with supervision from known optimal solutions or implicitly by optimizing the compound function $f\circ \mathbf{g}$. The implicit approach may not require optimal solutions as labels and is capable of handling problem uncertainty; however, it is slow to train and deploy due to frequent calls to optimizer $\mathbf{g}$ during both training and testing. The training is further challenged by sparse gradients of $\mathbf{g}$, especially for combinatorial solvers. To address these challenges, we propose using a smooth and learnable Landscape Surrogate $M$ as a replacement for $f\circ \mathbf{g}$. This surrogate, learnable by neural networks, can be computed faster than the solver $\mathbf{g}$, provides dense and smooth gradients during training, can generalize to unseen optimization problems, and is efficiently learned via alternating optimization. We test our approach on both synthetic problems, including shortest path and multidimensional knapsack, and real-world problems such as portfolio optimization, achieving comparable or superior objective values compared to state-of-the-art baselines while reducing the number of calls to $\mathbf{g}$. Notably, our approach outperforms existing methods for computationally expensive high-dimensional problems.
This paper presents RAYEN, a framework to impose hard convex constraints on the output or latent variable of a neural network. RAYEN guarantees that, for any input or any weights of the network, the constraints are satisfied at all times. Compared to other approaches, RAYEN does not perform a computationally-expensive orthogonal projection step onto the feasible set, does not rely on soft constraints (which do not guarantee the satisfaction of the constraints at test time), does not use conservative approximations of the feasible set, and does not perform a potentially slow inner gradient descent correction to enforce the constraints. RAYEN supports any combination of linear, convex quadratic, second-order cone (SOC), and linear matrix inequality (LMI) constraints, achieving a very small computational overhead compared to unconstrained networks. For example, it is able to impose 1K quadratic constraints on a 1K-dimensional variable with an overhead of less than 8 ms, and an LMI constraint with 300x300 dense matrices on a 10K-dimensional variable in less than 12 ms. When used in neural networks that approximate the solution of constrained optimization problems, RAYEN achieves computation times between 20 and 7468 times faster than state-of-the-art algorithms, while guaranteeing the satisfaction of the constraints at all times and obtaining a cost very close to the optimal one.
We consider gradient-related methods for low-rank matrix optimization with a smooth cost function. The methods operate on single factors of the low-rank factorization and share aspects of both alternating and Riemannian optimization. Two possible choices for the search directions based on Gauss-Southwell type selection rules are compared: one using the gradient of a factorized non-convex formulation, the other using the Riemannian gradient. While both methods provide gradient convergence guarantees that are similar to the unconstrained case, numerical experiments on a quadratic cost function indicate that the version based on the Riemannian gradient is significantly more robust with respect to small singular values and the condition number of the cost function. As a side result of our approach, we also obtain new convergence results for the alternating least squares method.
The ability to remove features from the input of machine learning models is very important to understand and interpret model predictions. However, this is non-trivial for vision models since masking out parts of the input image typically causes large distribution shifts. This is because the baseline color used for masking (typically grey or black) is out of distribution. Furthermore, the shape of the mask itself can contain unwanted signals which can be used by the model for its predictions. Recently, there has been some progress in mitigating this issue (called missingness bias) in image masking for vision transformers. In this work, we propose a new masking method for CNNs we call layer masking in which the missingness bias caused by masking is reduced to a large extent. Intuitively, layer masking applies a mask to intermediate activation maps so that the model only processes the unmasked input. We show that our method (i) is able to eliminate or minimize the influence of the mask shape or color on the output of the model, and (ii) is much better than replacing the masked region by black or grey for input perturbation based interpretability techniques like LIME. Thus, layer masking is much less affected by missingness bias than other masking strategies. We also demonstrate how the shape of the mask may leak information about the class, thus affecting estimates of model reliance on class-relevant features derived from input masking. Furthermore, we discuss the role of data augmentation techniques for tackling this problem, and argue that they are not sufficient for preventing model reliance on mask shape. The code for this project is publicly available at //github.com/SriramB-98/layer_masking
In this article, we propose two kinds of neural networks inspired by power method and inverse power method to solve linear eigenvalue problems. These neural networks share similar ideas with traditional methods, in which the differential operator is realized by automatic differentiation. The eigenfunction of the eigenvalue problem is learned by the neural network and the iterative algorithms are implemented by optimizing the specially defined loss function. The largest positive eigenvalue, smallest eigenvalue and interior eigenvalues with the given prior knowledge can be solved efficiently. We examine the applicability and accuracy of our methods in the numerical experiments in one dimension, two dimensions and higher dimensions. Numerical results show that accurate eigenvalue and eigenfunction approximations can be obtained by our methods.
We study the bias of Stochastic Gradient Descent (SGD) to learn low-rank weight matrices when training deep ReLU neural networks. Our results show that training neural networks with mini-batch SGD and weight decay causes a bias towards rank minimization over the weight matrices. Specifically, we show, both theoretically and empirically, that this bias is more pronounced when using smaller batch sizes, higher learning rates, or increased weight decay. Additionally, we predict and observe empirically that weight decay is necessary to achieve this bias. In addition, we show that in the presence of intermediate neural collapse, the learned weights are particularly low-rank. Unlike previous literature, our analysis does not rely on assumptions about the data, convergence, or optimality of the weight matrices. Furthermore, it applies to a wide range of neural network architectures of any width or depth. Finally, we empirically investigate the connection between this bias and generalization, finding that it has a marginal effect on generalization.
Despite widespread adoption in practice, guarantees for the LASSO and Group LASSO are strikingly lacking in settings beyond statistical problems, and these algorithms are usually considered to be a heuristic in the context of sparse convex optimization on deterministic inputs. We give the first recovery guarantees for the Group LASSO for sparse convex optimization with vector-valued features. We show that if a sufficiently large Group LASSO regularization is applied when minimizing a strictly convex function $l$, then the minimizer is a sparse vector supported on vector-valued features with the largest $\ell_2$ norm of the gradient. Thus, repeating this procedure selects the same set of features as the Orthogonal Matching Pursuit algorithm, which admits recovery guarantees for any function $l$ with restricted strong convexity and smoothness via weak submodularity arguments. This answers open questions of Tibshirani et al. and Yasuda et al. Our result is the first to theoretically explain the empirical success of the Group LASSO for convex functions under general input instances assuming only restricted strong convexity and smoothness. Our result also generalizes provable guarantees for the Sequential Attention algorithm, which is a feature selection algorithm inspired by the attention mechanism proposed by Yasuda et al. As an application of our result, we give new results for the column subset selection problem, which is well-studied when the loss is the Frobenius norm or other entrywise matrix losses. We give the first result for general loss functions for this problem that requires only restricted strong convexity and smoothness.
This paper studies online convex optimization with stochastic constraints. We propose a variant of the drift-plus-penalty algorithm that guarantees $O(\sqrt{T})$ expected regret and zero constraint violation, after a fixed number of iterations, which improves the vanilla drift-plus-penalty method with $O(\sqrt{T})$ constraint violation. Our algorithm is oblivious to the length of the time horizon $T$, in contrast to the vanilla drift-plus-penalty method. This is based on our novel drift lemma that provides time-varying bounds on the virtual queue drift and, as a result, leads to time-varying bounds on the expected virtual queue length. Moreover, we extend our framework to stochastic-constrained online convex optimization under two-point bandit feedback. We show that by adapting our algorithmic framework to the bandit feedback setting, we may still achieve $O(\sqrt{T})$ expected regret and zero constraint violation, improving upon the previous work for the case of identical constraint functions. Numerical results demonstrate our theoretical results.
Understanding the implicit regularization imposed by neural network architectures and gradient based optimization methods is a key challenge in deep learning and AI. In this work we provide sharp results for the implicit regularization imposed by the gradient flow of Diagonal Linear Networks (DLNs) in the over-parameterized regression setting and, potentially surprisingly, link this to the phenomenon of phase transitions in generalized hardness of approximation (GHA). GHA generalizes the phenomenon of hardness of approximation from computer science to, among others, continuous and robust optimization. It is well-known that the $\ell^1$-norm of the gradient flow of DLNs with tiny initialization converges to the objective function of basis pursuit. We improve upon these results by showing that the gradient flow of DLNs with tiny initialization approximates minimizers of the basis pursuit optimization problem (as opposed to just the objective function), and we obtain new and sharp convergence bounds w.r.t.\ the initialization size. Non-sharpness of our results would imply that the GHA phenomenon would not occur for the basis pursuit optimization problem -- which is a contradiction -- thus implying sharpness. Moreover, we characterize $\textit{which}$ $\ell_1$ minimizer of the basis pursuit problem is chosen by the gradient flow whenever the minimizer is not unique. Interestingly, this depends on the depth of the DLN.
Residual networks (ResNets) have displayed impressive results in pattern recognition and, recently, have garnered considerable theoretical interest due to a perceived link with neural ordinary differential equations (neural ODEs). This link relies on the convergence of network weights to a smooth function as the number of layers increases. We investigate the properties of weights trained by stochastic gradient descent and their scaling with network depth through detailed numerical experiments. We observe the existence of scaling regimes markedly different from those assumed in neural ODE literature. Depending on certain features of the network architecture, such as the smoothness of the activation function, one may obtain an alternative ODE limit, a stochastic differential equation or neither of these. These findings cast doubts on the validity of the neural ODE model as an adequate asymptotic description of deep ResNets and point to an alternative class of differential equations as a better description of the deep network limit.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191