1 Introduction
The wellknown work of Zhang et al. (2017) highlighted intriguing experimental phenomena about deep net training – specifically, optimization and generalization – and asked whether theory could explain them. They showed that sufficiently powerful nets (with vastly more parameters than number of training samples) can attain zero training error, regardless of whether the data is properly labeled or randomly labeled. Obviously, training with randomly labeled data cannot generalize, whereas training with properly labeled data generalizes. See Figure 2 replicating some of these results.
Recent papers have begun to provide explanations, showing that gradient descent can allow an overparametrized multilayer net to attain arbitrarily low training error on fairly generic datasets (Du et al., 2018a, c; Li & Liang, 2018; AllenZhu et al., 2018b; Zou et al., 2018), provided the amount of overparametrization is a high polynomial of the relevant parameters (i.e. vastly more than the overparametrization in Zhang et al. (2017)). Under further assumptions it can also be shown that the trained net generalizes (AllenZhu et al., 2018a). But some issues were not addressed in these papers, and the goal of the current paper is to address them.
First, the experiments in Zhang et al. (2017) show that though the nets attain zero training error on even random data, the convergence rate is much slower. See Figure 1.
Question 1.
Why do true labels give faster convergence rate than random labels for gradient descent?
The above papers do not answer this question, since their proof of convergence does not distinguish between good and random labels.
The next issue is about generalization: clearly, some property of properly labeled data controls generalization, but what? Classical measures used in generalization theory such as VCdimension and Rademacher complexity are much too pessimistic. A line of research proposed normbased (e.g. Bartlett et al. (2017a)) and compressionbased bounds (Arora et al., 2018). But the sample complexity upper bounds obtained are still far too weak. Furthermore they rely on some property of the trained net that is revealed/computed at the end of training. There is no property of data alone that determine upfront whether the trained net will generalize. A recent paper (AllenZhu et al., 2018a) assumed that there exists an underlying (unknown) neural network that achieves low error on the data distribution, and the amount of data available is quite a bit more than the minimum number of samples needed to learn this underlying neural net. Under this condition, the overparametrized net (which has way more parameters) can learn in a way that generalizes. However, it is hard to verify from data whether this assumption is satisfied, even after the larger net has finished training.^{1}^{1}1In Section 2, we discuss the related works in more details. Thus the assumption is in some sense unverifiable.
Question 2.
Is there an easily verifiable complexity measure that can differentiate true labels and random labels?
Without explicit regularization, to attack this problem, one must resort to algorithmdependent generalization analysis. One such line of work established that firstorder methods can automatically find minimumnorm/maximummargin solutions that fit the data in the settings of logistic regression, deep linear networks, and symmetric matrix factorization
(Soudry et al., 2018; Gunasekar et al., 2018a, b; Ji & Telgarsky, 2018; Li et al., 2018b). However, how to extend these results to nonlinear neural networks remains unclear (Wei et al., 2018). Another line of algorithmdependent analysis of generalization (Hardt et al., 2015; Mou et al., 2017; Chen et al., 2018)used stability of specific optimization algorithms that satisfy certain generic properties like convexity, smoothness, etc. However, as the number of epochs becomes large, these generalization bounds are vacuous.
Our results.
We give a new analysis that provides answers to Questions 1 and 2
for overparameterized twolayer neural networks with ReLU activation trained by gradient descent (GD), when the number of neurons in the hidden layer is sufficiently large. In this setting,
Du et al. (2018c) have proved that GD with random initialization can achieve zero training error for any nondegenerate data. We give a more refined analysis of the trajectory of GD which enables us to provide answers to Questions 1 and 2. In particular:
In Section 4
, using the trajectory of the network predictions on the training data during optimization, we accurately estimate the magnitude of training loss in each iteration. Our key finding is that the number of iterations needed to achieve a target accuracy depends on the projections of data labels on the eigenvectors of a certain Gram matrix to be defined in Equation (
3). On MNIST and CIFAR datasets, we find that such projections are significantly different for true labels and random labels, and as a result we are able to answer Question 1. 
In Section 5, we give a generalization bound for the solution found by GD, based on accurate estimates of how much the network parameters can move during optimization (in suitable norms). Our generalization bound depends on a datadependent complexity measure (c.f. Equation (10)), and notably, is completely independent of the number of hidden units in the network. Again, we test this complexity measure on MNIST and CIFAR, and find that the complexity measures for true and random labels are significantly different, which thus answers Question 2.
Notice that because zero training error is achieved by the solution found by GD, a generalization bound is an upper bound on the error on the data distribution (test error). We also remark that our generalization bound is valid for any data labels – it does not require the existence of a small groundtruth network as in (AllenZhu et al., 2018a). Moreover, our bound can be efficiently computed for any data labels.

In Section 6, we further study what kind of functions can be provably learned by twolayer ReLU networks trained by GD. Combining the optimization and generalization results, we uncover a broad class of learnable functions, including linear functions, twolayer neural networks with polynomial activation or cosine activation, etc. Our requirement on the smoothness of learnable functions is weaker than that in (AllenZhu et al., 2018a).
Finally, we note that the intriguing generalization phenomena in deep learning were observed in kernel methods as well
Belkin et al. (2018). The analysis in the current paper is also related to a kernel from the ReLU activation (c.f. Equation (3)).2 Related Work
In this section we survey previous works on optimization and generalization aspects of neural networks.
Optimization.
Many papers tried to characterize geometric landscapes of objective functions (Safran & Shamir, 2017; Zhou & Liang, 2017; Freeman & Bruna, 2016; Hardt & Ma, 2016; Nguyen & Hein, 2017; Kawaguchi, 2016; Venturi et al., 2018; Soudry & Carmon, 2016; Du & Lee, 2018; Soltanolkotabi et al., 2018; Haeffele & Vidal, 2015). The hope is to leverage recent advance in firstorder algorithms (Ge et al., 2015; Lee et al., 2016; Jin et al., 2017) which showed that if the landscape satisfies (1) all local minima are global and (2) all saddle points are strict (i.e., there exists a negative curvature), then firstorder methods can escape all saddle points and find a global minimum. Unfortunately, these desired properties do not hold even for simple nonlinear shallow neural networks (Yun et al., 2018) or 3layer linear neural networks (Kawaguchi, 2016).
Another approach is to directly analyze trajectory of the optimization method and to show convergence to global minimum. A series of papers made strong assumptions on input distribution as well as realizability of labels, and showed global convergence of (stochastic) gradient descent for some shallow neural networks
(Tian, 2017; Soltanolkotabi, 2017; Brutzkus & Globerson, 2017; Du et al., 2017a, b; Li & Yuan, 2017). Some local convergence results have also been proved (Zhong et al., 2017; Zhang et al., 2018). However, these assumptions are not satisfied in practice.For twolayer neural networks, a line of papers used mean field analysis to establish that for infinitely wide neural networks, the empirical distribution of the neural network parameters can be described as a Wasserstein gradient flow (Mei et al., 2018; Chizat & Bach, 2018a; Sirignano & Spiliopoulos, 2018; Rotskoff & VandenEijnden, 2018; Wei et al., 2018). However, it is unclear whether this framework can explain the behavior of firstorder methods on finitesize neural networks.
Recent breakthroughs were made in understanding optimization of overparameterized neural networks through the trajectorybased approach. They proved global polynomial time convergence of (stochastic) gradient descent on nonlinear neural networks for minimizing empirical risk. Their proof techniques can be roughly classified into two categories.
Li & Liang (2018); AllenZhu et al. (2018b); Zou et al. (2018) analyzed the trajectory of parameters and showed that on the trajectory, the objective function satisfies certain gradient dominance property. On the other hand, Du et al. (2018a, c) analyzed the trajectory of network predictions on training samples and showed that it enjoys a stronglyconvexlike property.Generalization.
It is well known that the VCdimension of neural networks is at least linear in the number of parameters (Bartlett et al., 2017b), and therefore classical VC theory cannot explain the generalization ability of modern neural networks with more parameters than training samples. Researchers have proposed normbased generalization bounds (Bartlett & Mendelson, 2002; Bartlett et al., 2017a; Neyshabur et al., 2015, 2017, 2019; Konstantinos et al., 2017; Golowich et al., 2017; Li et al., 2018a) and compressionbased bounds (Arora et al., 2018). Dziugaite & Roy (2017); Zhou et al. (2019)
used the PACBayes approach to compute nonvacuous generalization bounds for MNIST and ImageNet, respectively. All these bounds are
posterior in nature – they depend on certain properties of the trained neural networks. Therefore, one has to finish training a neural network to know whether it can generalize. Comparing with these results, our generalization bound only depends on training data and can be calculated without actually training the neural network.Another line of work assumed the existence of a true model, and showed that the (regularized) empirical risk minimizer has good generalization with sample complexity that depends on the true model (Du et al., 2018b; Ma et al., 2018; Imaizumi & Fukumizu, 2018). These papers ignored the difficulty of optimization, while we are able to prove generalization of the solution found by gradient descent. Furthermore, our generic generalization bound does not assume the existence of any true model.
Our paper is closely related to (AllenZhu et al., 2018a) which showed that twolayer overparametrized neural networks trained by randomly initialized stochastic gradient descent can learn a class of infiniteorder smooth functions. In contrast, our generalization bound depends on a datadependent complexity measure that can be computed for any dataset, without assuming any groundtruth model. As a consequence of our generic bound, we also show that twolayer neural networks can learn a class of infiniteorder smooth functions, with a less strict requirement for smoothness. Furthermore, our bound is completely independent of the number of hidden units, while there is a polylogarithmic dependence in AllenZhu et al. (2018a). AllenZhu et al. (2018a) also studied the generalization performance of threelayer neural networks.
3 Preliminaries and Overview of Results
Notation.
We use boldfaced letters for vectors and matrices. For a matrix
, let be its th entry. We use to denote the Euclidean norm of a vector or the spectral norm of a matrix, and use to denote the Frobenius norm of a matrix. Denote bythe minimum eigenvalue of a symmetric matrix
. Let be the vectorization of a matrix in columnfirst order. Letbe the identity matrix and
. Denote bythe Gaussian distribution with mean
and covariance . Denote by the ReLU function . Denote by the indicator function for an event .3.1 Setting: TwoLayer Neural Network Trained by Randomly Initialized Gradient Descent
We consider a twolayer ReLU activated neural network with neurons in the hidden layer:
where is the input, are weight vectors in the first layer, are weights in the second layer. For convenience we denote and .
We are given inputlabel samples drawn i.i.d. from an underlying data distribution over . We denote and . For simplicity, we assume that for sampled from , we have and .
We train the neural network by randomly initialized gradient descent (GD) on the quadratic loss over data . In particular, we first initialize the parameters randomly:
(1) 
where controls the magnitude of initialization, and all randomnesses are independent. We then fix the second layer and optimize the first layer through GD on the following objective function:
(2) 
The GD update rule can be written as:^{2}^{2}2Since ReLU is not differentiable at , we just define “gradient” using this formula, and this is indeed what is used in practice.
where is the learning rate.
3.2 The Gram Matrix from ReLU Kernel
Given , we define the following Gram matrix as follows:
(3)  
This matrix can be viewed as a Gram matrix from a kernel associated with the ReLU function, and has been studied in (Xie et al., 2017; Tsuchida et al., 2017; Du et al., 2018c).
In our setting of training a twolayer ReLU network, Du et al. (2018c) showed that if is positive definite, GD converges to training loss if is sufficiently large:
Theorem 3.1 ((Du et al., 2018c)^{3}^{3}3Du et al. (2018c) only considered the case , but it is straightforward to generalize their result to general at the price of an extra factor in .).
Assume . For , if and
, then with probability at least
over the random initialization (1), we have:
;

.
Our results on optimization and generalization also crucially depend on this matrix .
3.3 Overview of Our Results
Now we give an informal description of our main results. It assumes that the initialization magnitude is sufficiently small and the network width is sufficiently large (to be quantified later).
The following theorem gives a precise characterization of how the objective decreases to . It says that this process is essentially determined by a power method for matrix applied on the label vector .
Theorem 3.2 (Informal version of Theorem 4.1).
With high probability we have:
As a consequence, we are able to distinguish the convergence rates for different labels , which can be determined by the projections of on the eigenvectors of . This allows us to obtain an answer to Question 1. See Section 4 for details.
Our main result for generalization is the following:
Theorem 3.3 (Informal version of Theorem 5.1).
For any
Lipschitz loss function, the generalization error of the twolayer ReLU network found by GD is at most
(4) 
3.4 Additional Notation
We introduce some additional notation that will be used.
Define , i.e., the network’s prediction on the th input. We also use to denote all predictions. Then we have and the gradient of can be written as:
(5) 
where .
We define two matrices and which will play a key role in our analysis of the GD trajectory:
and . Note that
With this notation we have a more compact form of the gradient (5):
Then the GD update rule is:
(6) 
for . Throughout the paper, we use as the iteration number, and also use to index all variables that depend on . For example, we have , , etc.
4 Analysis of Convergence Rate
Although Theorem 3.1 already predicts linear convergence of GD to loss, it only provides an upper bound on the loss and does not distinguish different types of labels. In particular, it cannot answer Question 1. In this section we give a finegrained analysis of the convergence rate.
Recall the loss function . Thus, it is equivalent to study how fast the sequence converges to . Key to our analysis is the observation that when the size of initialization is small and the network width is large, the sequence stays close to another sequence which has a linear update rule:
(7)  
where is the Gram matrix defined in (3).
Write the eigendecomposition , where are orthonormal eigenvectors of and are corresponding eigenvalues. Our main theorem in this section is the following:
Theorem 4.1.
Suppose , , and . Then with probability at least over the random initialization, for all we have:
(8) 
In fact, the dominating term is exactly equal to , which we prove in Section 4.1.
In light of (8), it suffices to understand how fast converges to as grows. Define , and notice that each sequence is a geometric sequence which starts at and decreases at ratio . In other words, we can think of decomposing the label vector into its projections onto all eigenvectors of : , and the th portion shrinks exponentially at ratio . The larger is, the faster decreases to , so in order to have faster convergence we would like the projections of onto top eigenvectors to be larger. Therefore we obtain the following intuitive rule to compare the convergence rates on two sets of labels in a qualitative manner (for fixed ):

For a set of labels , if they align with the top eigenvectors, i.e., is large for large , then gradient descent converges quickly.

For a set of labels , if the projections on eigenvectors are uniform, or labels align with eigenvectors with respect to small eigenvalues, then gradient descent converges with a slow rate.
Answer to Question 1.
We now use this reasoning to answer Question 1. In Figure 1(b), we compute the eigenvalues of (blue curve) for the MNIST dataset. The plot shows the eigenvalues of admit a fast decay. We further compute the projections of true labels (red) and random labels (cyan). We observe that there is a significant difference between the projections of true labels and random labels: true labels align well with top eigenvectors whereas projections of random labels are close to being uniform. Furthermore, according to our theory, if a set of labels align with the eigenvector associated with the least eigenvalue, the convergence rate of gradient descent will be extremely slow. We construct such labels and in Figure 1(a) we indeed observe slow convergence. We repeat the same experiments on CIFAR and have similar observations (Figures 1(c) and 1(d)). These empirical findings support our theory on the convergence rate of gradient descent. See Appendix A for implementation details.
4.1 Proof Sketch of Theorem 4.1
Now we prove . The entire proof of Theorem 4.1 is given in Appendix C, which relies on the fact that the dynamics of is essentially a perturbed version of (7).
From (7) we have , which implies . Note that has eigendecomposition and that can be decomposed as . Then we have , which implies .
5 Analysis of Generalization
In this section, we study the generalization ability of the twolayer neural network trained by GD.
First, in order for optimization to succeed, i.e., zero training loss is achieved, we need a nondegeneracy assumption on the data distribution, defined below:
Definition 5.1.
A distribution over is nondegenerate, if for i.i.d. samples from , with probability at least we have .
Remark 5.1.
Note that as long as no two and are parallel to each other, we have . (See (Du et al., 2018c)). For most realworld distributions, any two training inputs are not parallel.
Our main theorem is the following:
Theorem 5.1.
Fix an error parameter and failure probability . Suppose our data are i.i.d. samples from a nondegenerate distribution , and . Consider any loss function that is Lipschitz in the first argument such that . Then with probability at least over the random initialization and the training samples, the twolayer neural network trained by GD for iterations has population loss bounded as:
(9) 
Note that in Theorem 5.1 there are three sources of possible failures: (i) failure of satisfying , (ii) failure of random initialization, and (iii) failure in the data sampling procedure (c.f. Theorem B.1). We ensure that all these failure probabilities are at most so that the final failure probability is at most .
As a corollary of Theorem 5.1, for binary classification problems (i.e., labels are ), Corollary 5.2 shows that (9) also bounds the population classification error of the learned classifier. See Appendix D for the proof.
Corollary 5.2.
Under the same assumptions as in Theorem 5.1 and additionally assuming that for , with probability at least , the population classification error is bounded as:
Now we discuss our generalization bound. The dominating term in (9) is:
(10) 
This can be viewed as a complexity measure of data that one can use to predict the test accuracy of the learned neural network. Our result has the following advantages: (i) our complexity measure (10) can be directly computed given data , without the need of training a neural network or assuming a groundtruth model; (ii) our bound is completely independent of the network width ; (iii) our theorem does not require early stopping of optimization as in AllenZhu et al. (2018a).
Evaluating our completixy measure (10).
To illustrate that the complexity measure in (10) effectively determines test error, in Figure 2 we compare this complexity measure versus the test error with true labels and random labels (and mixture of true and random labels). Random and true labels have significantly different complexity measures, and as the portion of random labels increases, our complexity measure also increases. See Appendix A for implementation details.
5.1 Proof Sketch of Theorem 5.1
The main ingredients in the proof of Theorem 5.1 are Lemmas 5.3 and 5.4. We defer the proofs of these lemmas as well as the full proof of Theorem 5.1 to Appendix D.
Our proof is based on a careful characterization of the trajectory of during GD. In particular, we bound its distance to initialization as follows:
Lemma 5.3.
Suppose and . Then with probability at least over the random initialization, we have for all :

, and

.
The bound on the movement of each was proved in Du et al. (2018c). Our main contribution is the bound on which corresponds to the total movement of all neurons. The main idea is to couple the trajectory of with another simpler trajectory defined as:
(11)  
We prove in Section 5.2.^{4}^{4}4Note that we have from standard concentration. See Lemma C.3. The actually proof of Lemma 5.3 is essentially a perturbed version of this.
Lemma 5.3 implies that the learned function from GD is in a restricted class of neural nets whose weights are close to initialization . The following lemma bounds the Rademacher complexity of this function class:
Lemma 5.4.
For given , with probability at least over the random initialization (, the following function class
has empirical Rademacher complexity bounded as:
5.2 Analysis of the Auxiliary Sequence
6 Provable Learning using TwoLayer ReLU Neural Networks
Theorem 5.1 determines that controls the generalization error. In this section, we study what functions can be provably learned in this setting. We assume the data satisfy for some underlying function . A simple observation is that if we can prove
for some quantity that is independent of the number of samples , then Theorem 5.1 implies we can provably learn the function on the underlying data distribution using samples. The following theorem shows that this is indeed the case for a broad class of functions.
Theorem 6.1.
Suppose we have
where or , and . Then we have
Notice that for two label vectors and , we have
This implies that the sum of learnable functions is also learnable. Therefore, the following is a direct corollary of Theorem 6.1:
Corollary 6.2.
Suppose we have
(12) 
where for each , , and . Then we have
(13) 
Corollary 6.2 shows that overparameterized twolayer ReLU network can learn any function of the form (12) for which (13) is bounded. One can view (12) as twolayer neural networks with polynomial activation , where are weights in the first layer and are the second layer. Below we give some specific examples.
Example 6.1 (Linear functions).
For , we have .
Example 6.2 (Quadratic functions).
For where is symmetric, we can write down the eigendecomposition . Then we have , so .^{5}^{5}5 is the tracenorm of . This is also the class of twolayer neural networks with quadratic activation.
Example 6.3 (Cosine activation).
Suppose for some . Using Taylor series we know . Thus we have .
Finally, we note that our “smoothness” requirement (13) is weaker than that in (AllenZhu et al., 2018a), as illustrated in the following example.
Example 6.4 (A notsosmooth function).
Suppose , where and . We have since . Thus , so our result implies that this function is learnable by layer ReLU nets.
However, for the above function, AllenZhu et al. (2018a)’s generalization theorem would require
to be bounded, where is a large constant and is the target generalization error. This is clearly not satisfied.
7 Conclusion
This paper shows how to give a finegrained analysis of the optimization trajectory and the generalization ability of overparameterized twolayer neural networks trained by gradient descent. We believe that our approach can also be useful in analyzing overparameterized deep neural networks and other machine learning models.
Acknowledgements
This work is supported by NSF, ONR, Simons Foundation, Schmidt Foundation, Mozilla Research, Amazon Research, DARPA and SRC. The authors would like to thank Yi Zhang for helpful discussions.
References
 AllenZhu et al. (2018a) AllenZhu, Z., Li, Y., and Liang, Y. Learning and generalization in overparameterized neural networks, going beyond two layers. arXiv preprint arXiv:1811.04918, 2018a.
 AllenZhu et al. (2018b) AllenZhu, Z., Li, Y., and Song, Z. A convergence theory for deep learning via overparameterization. arXiv preprint arXiv:1811.03962, 2018b.
 Arora et al. (2018) Arora, S., Ge, R., Neyshabur, B., and Zhang, Y. Stronger generalization bounds for deep nets via a compression approach. arXiv preprint arXiv:1802.05296, 2018.
 Bartlett & Mendelson (2002) Bartlett, P. L. and Mendelson, S. Rademacher and gaussian complexities: Risk bounds and structural results. Journal of Machine Learning Research, 3(Nov):463–482, 2002.
 Bartlett et al. (2017a) Bartlett, P. L., Foster, D. J., and Telgarsky, M. J. Spectrallynormalized margin bounds for neural networks. In Advances in Neural Information Processing Systems, pp. 6241–6250, 2017a.
 Bartlett et al. (2017b) Bartlett, P. L., Harvey, N., Liaw, C., and Mehrabian, A. Nearlytight VCdimension and pseudodimension bounds for piecewise linear neural networks. arXiv preprint arXiv:1703.02930, 2017b.
 Belkin et al. (2018) Belkin, M., Ma, S., and Mandal, S. To understand deep learning we need to understand kernel learning. arXiv preprint arXiv:1802.01396, 2018.
 Brutzkus & Globerson (2017) Brutzkus, A. and Globerson, A. Globally optimal gradient descent for a ConvNet with gaussian inputs. arXiv preprint arXiv:1702.07966, 2017.
 Chen et al. (2018) Chen, Y., Jin, C., and Yu, B. Stability and convergence tradeoff of iterative optimization algorithms. arXiv preprint arXiv:1804.01619, 2018.
 Chizat & Bach (2018a) Chizat, L. and Bach, F. On the global convergence of gradient descent for overparameterized models using optimal transport. arXiv preprint arXiv:1805.09545, 2018a.
 Chizat & Bach (2018b) Chizat, L. and Bach, F. A note on lazy training in supervised differentiable programming. arXiv preprint arXiv:1812.07956, 2018b.
 Daniely (2017) Daniely, A. SGD learns the conjugate kernel class of the network. arXiv preprint arXiv:1702.08503, 2017.
 Du & Lee (2018) Du, S. S. and Lee, J. D. On the power of overparametrization in neural networks with quadratic activation. arXiv preprint arXiv:1803.01206, 2018.
 Du et al. (2017a) Du, S. S., Lee, J. D., and Tian, Y. When is a convolutional filter easy to learn? arXiv preprint arXiv:1709.06129, 2017a.
 Du et al. (2017b) Du, S. S., Lee, J. D., Tian, Y., Poczos, B., and Singh, A. Gradient descent learns onehiddenlayer CNN: Don’t be afraid of spurious local minima. arXiv preprint arXiv:1712.00779, 2017b.
 Du et al. (2018a) Du, S. S., Lee, J. D., Li, H., Wang, L., and Zhai, X. Gradient descent finds global minima of deep neural networks. arXiv preprint arXiv:1811.03804, 2018a.

Du et al. (2018b)
Du, S. S., Wang, Y., Zhai, X., Balakrishnan, S., Salakhutdinov, R. R., and
Singh, A.
How many samples are needed to estimate a convolutional neural network?
In Advances in Neural Information Processing Systems, pp. 371–381, 2018b.  Du et al. (2018c) Du, S. S., Zhai, X., Poczos, B., and Singh, A. Gradient descent provably optimizes overparameterized neural networks. arXiv preprint arXiv:1810.02054, 2018c.
 Dziugaite & Roy (2017) Dziugaite, G. K. and Roy, D. M. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. arXiv preprint arXiv:1703.11008, 2017.
 Freeman & Bruna (2016) Freeman, C. D. and Bruna, J. Topology and geometry of halfrectified network optimization. arXiv preprint arXiv:1611.01540, 2016.

Ge et al. (2015)
Ge, R., Huang, F., Jin, C., and Yuan, Y.
Escaping from saddle points
online stochastic gradient for tensor decomposition.
In Proceedings of The 28th Conference on Learning Theory, pp. 797–842, 2015.  Golowich et al. (2017) Golowich, N., Rakhlin, A., and Shamir, O. Sizeindependent sample complexity of neural networks. arXiv preprint arXiv:1712.06541, 2017.
 Gunasekar et al. (2018a) Gunasekar, S., Lee, J., Soudry, D., and Srebro, N. Characterizing implicit bias in terms of optimization geometry. arXiv preprint arXiv:1802.08246, 2018a.
 Gunasekar et al. (2018b) Gunasekar, S., Lee, J., Soudry, D., and Srebro, N. Implicit bias of gradient descent on linear convolutional networks. arXiv preprint arXiv:1806.00468, 2018b.
 Haeffele & Vidal (2015) Haeffele, B. D. and Vidal, R. Global optimality in tensor factorization, deep learning, and beyond. arXiv preprint arXiv:1506.07540, 2015.
 Hardt & Ma (2016) Hardt, M. and Ma, T. Identity matters in deep learning. arXiv preprint arXiv:1611.04231, 2016.
 Hardt et al. (2015) Hardt, M., Recht, B., and Singer, Y. Train faster, generalize better: Stability of stochastic gradient descent. arXiv preprint arXiv:1509.01240, 2015.
 Imaizumi & Fukumizu (2018) Imaizumi, M. and Fukumizu, K. Deep neural networks learn nonsmooth functions effectively. arXiv preprint arXiv:1802.04474, 2018.
 Jacot et al. (2018) Jacot, A., Gabriel, F., and Hongler, C. Neural tangent kernel: Convergence and generalization in neural networks. arXiv preprint arXiv:1806.07572, 2018.
 Ji & Telgarsky (2018) Ji, Z. and Telgarsky, M. Gradient descent aligns the layers of deep linear networks. arXiv preprint arXiv:1810.02032, 2018.
 Jin et al. (2017) Jin, C., Ge, R., Netrapalli, P., Kakade, S. M., and Jordan, M. I. How to escape saddle points efficiently. In Proceedings of the 34th International Conference on Machine Learning, pp. 1724–1732, 2017.
 Kawaguchi (2016) Kawaguchi, K. Deep learning without poor local minima. In Advances In Neural Information Processing Systems, pp. 586–594, 2016.
 Konstantinos et al. (2017) Konstantinos, P., Davies, M., and Vandergheynst, P. PACBayesian margin bounds for convolutional neural networkstechnical report. arXiv preprint arXiv:1801.00171, 2017.
 Krizhevsky & Hinton (2009) Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009.
 LeCun et al. (1998) LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradientbased learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
 Lee et al. (2016) Lee, J. D., Simchowitz, M., Jordan, M. I., and Recht, B. Gradient descent only converges to minimizers. In Conference on Learning Theory, pp. 1246–1257, 2016.
 Li et al. (2018a) Li, X., Lu, J., Wang, Z., Haupt, J., and Zhao, T. On tighter generalization bound for deep neural networks: CNNs, ResNets, and beyond. arXiv preprint arXiv:1806.05159, 2018a.
 Li & Liang (2018) Li, Y. and Liang, Y. Learning overparameterized neural networks via stochastic gradient descent on structured data. arXiv preprint arXiv:1808.01204, 2018.
 Li & Yuan (2017) Li, Y. and Yuan, Y. Convergence analysis of twolayer neural networks with ReLU activation. arXiv preprint arXiv:1705.09886, 2017.
 Li et al. (2018b) Li, Y., Ma, T., and Zhang, H. Algorithmic regularization in overparameterized matrix sensing and neural networks with quadratic activations. In Conference On Learning Theory, pp. 2–47, 2018b.
 Ma et al. (2018) Ma, C., Wu, L., et al. A priori estimates of the generalization error for twolayer neural networks. arXiv preprint arXiv:1810.06397, 2018.
 Mei et al. (2018) Mei, S., Montanari, A., and Nguyen, P.M. A mean field view of the landscape of twolayers neural networks. arXiv preprint arXiv:1804.06561, 2018.
 Mohri et al. (2012) Mohri, M., Rostamizadeh, A., and Talwalkar, A. Foundations of machine learning. MIT Press, 2012.
 Mou et al. (2017) Mou, W., Wang, L., Zhai, X., and Zheng, K. Generalization bounds of SGLD for nonconvex learning: Two theoretical viewpoints. arXiv preprint arXiv:1707.05947, 2017.
 Neyshabur et al. (2015) Neyshabur, B., Tomioka, R., and Srebro, N. Normbased capacity control in neural networks. In Conference on Learning Theory, pp. 1376–1401, 2015.
 Neyshabur et al. (2017) Neyshabur, B., Bhojanapalli, S., McAllester, D., and Srebro, N. A PACBayesian approach to spectrallynormalized margin bounds for neural networks. arXiv preprint arXiv:1707.09564, 2017.
 Neyshabur et al. (2019) Neyshabur, B., Li, Z., Bhojanapalli, S., LeCun, Y., and Srebro, N. The role of overparametrization in generalization of neural networks. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=BygfghAcYX.
 Nguyen & Hein (2017) Nguyen, Q. and Hein, M. The loss surface of deep and wide neural networks. arXiv preprint arXiv:1704.08045, 2017.

Paszke et al. (2017)
Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z.,
Desmaison, A., Antiga, L., and Lerer, A.
Automatic differentiation in pytorch.
2017.  Rotskoff & VandenEijnden (2018) Rotskoff, G. M. and VandenEijnden, E. Neural networks as interacting particle systems: Asymptotic convexity of the loss landscape and universal scaling of the approximation error. arXiv preprint arXiv:1805.00915, 2018.
 Safran & Shamir (2017) Safran, I. and Shamir, O. Spurious local minima are common in twolayer relu neural networks. arXiv preprint arXiv:1712.08968, 2017.
 Sirignano & Spiliopoulos (2018) Sirignano, J. and Spiliopoulos, K. Mean field analysis of neural networks. arXiv preprint arXiv:1805.01053, 2018.
 Soltanolkotabi (2017) Soltanolkotabi, M. Learning ReLUs via gradient descent. arXiv preprint arXiv:1705.04591, 2017.
 Soltanolkotabi et al. (2018) Soltanolkotabi, M., Javanmard, A., and Lee, J. D. Theoretical insights into the optimization landscape of overparameterized shallow neural networks. IEEE Transactions on Information Theory, 2018.
 Soudry & Carmon (2016) Soudry, D. and Carmon, Y. No bad local minima: Data independent training error guarantees for multilayer neural networks. arXiv preprint arXiv:1605.08361, 2016.
 Soudry et al. (2018) Soudry, D., Hoffer, E., Nacson, M. S., Gunasekar, S., and Srebro, N. The implicit bias of gradient descent on separable data. Journal of Machine Learning Research, 19(70), 2018.
 Tian (2017) Tian, Y. An analytical formula of population gradient for twolayered ReLU network and its applications in convergence and critical point analysis. arXiv preprint arXiv:1703.00560, 2017.
 Tsuchida et al. (2017) Tsuchida, R., RoostaKhorasani, F., and Gallagher, M. Invariance of weight distributions in rectified mlps. arXiv preprint arXiv:1711.09090, 2017.
 Venturi et al. (2018) Venturi, L., Bandeira, A., and Bruna, J. Neural networks with finite intrinsic dimension have no spurious valleys. arXiv preprint arXiv:1802.06384, 2018.
 Wei et al. (2018) Wei, C., Lee, J. D., Liu, Q., and Ma, T. On the margin theory of feedforward neural networks. arXiv preprint arXiv:1810.05369, 2018.
 Xie et al. (2017) Xie, B., Liang, Y., and Song, L. Diverse neural network learns true target functions. In Artificial Intelligence and Statistics, pp. 1216–1224, 2017.
 Yun et al. (2018) Yun, C., Sra, S., and Jadbabaie, A. A critical view of global optimality in deep learning. arXiv preprint arXiv:1802.03487, 2018.
 Zhang et al. (2017) Zhang, C., Bengio, S., Hardt, M., Recht, B., and Vinyals, O. Understanding deep learning requires rethinking generalization. In Proceedings of the International Conference on Learning Representations (ICLR), 2017, 2017.
 Zhang et al. (2018) Zhang, X., Yu, Y., Wang, L., and Gu, Q. Learning onehiddenlayer relu networks via gradient descent. arXiv preprint arXiv:1806.07808, 2018.
 Zhong et al. (2017) Zhong, K., Song, Z., Jain, P., Bartlett, P. L., and Dhillon, I. S. Recovery guarantees for onehiddenlayer neural networks. arXiv preprint arXiv:1706.03175, 2017.
 Zhou et al. (2019) Zhou, W., Veitch, V., Austern, M., Adams, R. P., and Orbanz, P. Nonvacuous generalization bounds at the imagenet scale: a PACbayesian compression approach. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=BJgqqsAct7.
 Zhou & Liang (2017) Zhou, Y. and Liang, Y. Critical points of neural networks: Analytical forms and landscape properties. arXiv preprint arXiv:1710.11205, 2017.
 Zou et al. (2018) Zou, D., Cao, Y., Zhou, D., and Gu, Q. Stochastic gradient descent optimizes overparameterized deep ReLU networks. arXiv preprint arXiv:1811.08888, 2018.
Appendix
Appendix A Experimental Setup
The architecture of our neural networks is as described in Section 3.1. During the training process, we fix the second layer and only optimize the first layer, following the setting in Section 3.1. We fix the number of neurons to be in all experiments. We train the neural network using (fullbatch) gradient descent (GD), with a fixed learning rate . Our theory requires a small scaling factor during the initialization (cf. (1)). We fix in all experiments. We train the neural networks until the training loss converges.
We use two image datasets, the CIFAR dataset Krizhevsky & Hinton (2009) and the MNIST dataset LeCun et al. (1998), in our experiments. We only use the first two classes of images in the the CIFAR dataset and the MNIST dataset, with training images and validation images in total for each dataset. In both datasets, for each image , we set the corresponding label to be if the image belongs to the first class, and otherwise. For each image in the dataset, we normalize the image so that , following the setup in Section 3.1.
In the experiments reported in Figure 2, we choose a specific portion of (both training and test) data uniformly at random, and change their labels to .
Our neural networks are trained using the PyTorch package Paszke et al. (2017), using (possibly multiple) NVIDIA Tesla V100 GPUs.
Appendix B Background on Generalization and Rademacher Complexity
Consider a loss function . For a function , the population loss over data distribution as well as the empirical loss over samples from are defined as:
Generalization error refers to the gap for the learned function given sample .
Recall the standard definition of Rademacher complexity:
Definition B.1.
Given samples , the empirical Rademacher complexity of a function class (mapping from to ) is defined as:
where
contains i.i.d. random variables drawn from the Rademacher distribution
.Rademacher complexity directly gives an upper bound on generalization error (see e.g. (Mohri et al., 2012)):
Theorem B.1.
Suppose the loss function is bounded in and is Lipschitz in the first argument. Then with probability at least over sample of size :
Therefore, as long as we can bound the Rademacher complexity of a certain function class that contains our learned predictor, we can obtain a generalization bound.
Appendix C Proofs for Section 4
In this section we prove Theorem 4.1.
We first show some technical lemmas. Most of them are already proved in (Du et al., 2018c) and we give proofs for them for completeness.
First, we have the following lemma which gives an upper bound on how much each weight vector can move during optimization.
Lemma C.1.
Under the same setting as Theorem 3.1, i.e., , and , with probability at least over the random initialization we have