arXiv subjects

Philip M. Long

Publications and source records attributed to Philip M. Long.

27 records · Page 2Linked to original sources

arXiv2019

On the effect of the activation function on the distribution of hidden nodes in a deep network

We analyze the joint probability distribution on the lengths of the vectors of hidden variables in different layers of a fully connected deep network, when the weights and biases are chosen randomly according to Gaussian distributions, and the input is in $\{ -1, 1\}^N$. We show that, if the activation function $ϕ$ satisfies a minimal set of assumptions, satisfied by all activation functions that we know that are used in practice, then, as the width of the network gets large, the `length process' converges in probability to a length map that is determined as a simple function of the variances of the random weights and biases, and the activation function $ϕ$. We also show that this convergence may fail for $ϕ$ that violate our assumptions.

cs.LG↗

arXiv2018

Density estimation for shift-invariant multidimensional distributions

We study density estimation for classes of shift-invariant distributions over $\mathbb{R}^d$. A multidimensional distribution is "shift-invariant" if, roughly speaking, it is close in total variation distance to a small shift of it in any direction. Shift-invariance relaxes smoothness assumptions commonly used in non-parametric density estimation to allow jump discontinuities. The different classes of distributions that we consider correspond to different rates of tail decay. For each such class we give an efficient algorithm that learns any distribution in the class from independent samples with respect to total variation distance. As a special case of our general result, we show that $d$-dimensional shift-invariant distributions which satisfy an exponential tail bound can be learned to total variation distance error $ε$ using $\tilde{O}_d(1/ ε^{d+2})$ examples and $\tilde{O}_d(1/ ε^{2d+2})$ time. This implies that, for constant $d$, multivariate log-concave distributions can be learned in $\tilde{O}_d(1/ε^{2d+2})$ time using $\tilde{O}_d(1/ε^{d+2})$ samples, answering a question of [Diakonikolas, Kane and Stewart, 2016] All of our results extend to a model of noise-tolerant density estimation using Huber's contamination model, in which the target distribution to be learned is a $(1-ε,ε)$ mixture of some unknown distribution in the class with some other arbitrary and unknown distribution, and the learning algorithm must output a hypothesis distribution with total variation distance error $O(ε)$ from the target distribution. We show that our general results are close to best possible by proving a simple $Ω\left(1/ε^d\right)$ information-theoretic lower bound on sample complexity even for learning bounded distributions that are shift-invariant.

cs.LG↗

arXiv2018

Gradient descent with identity initialization efficiently learns positive definite linear transformations by deep residual networks

We analyze algorithms for approximating a function $f(x) = Φx$ mapping $\Re^d$ to $\Re^d$ using deep linear neural networks, i.e. that learn a function $h$ parameterized by matrices $Θ_1,...,Θ_L$ and defined by $h(x) = Θ_L Θ_{L-1} ... Θ_1 x$. We focus on algorithms that learn through gradient descent on the population quadratic loss in the case that the distribution over the inputs is isotropic. We provide polynomial bounds on the number of iterations for gradient descent to approximate the least squares matrix $Φ$, in the case where the initial hypothesis $Θ_1 = ... = Θ_L = I$ has excess loss bounded by a small enough constant. On the other hand, we show that gradient descent fails to converge for $Φ$ whose distance from the identity is a larger constant, and we show that some forms of regularization toward the identity in each layer do not help. If $Φ$ is symmetric positive definite, we show that an algorithm that initializes $Θ_i = I$ learns an $ε$-approximation of $f$ using a number of updates polynomial in $L$, the condition number of $Φ$, and $\log(d/ε)$. In contrast, we show that if the least squares matrix $Φ$ is symmetric and has a negative eigenvalue, then all members of a class of algorithms that perform gradient descent with identity initialization, and optionally regularize toward the identity in each layer, fail to converge. We analyze an algorithm for the case that $Φ$ satisfies $u^{\top} Φu > 0$ for all $u$, but may not be symmetric. This algorithm uses two regularizers: one that maintains the invariant $u^{\top} Θ_L Θ_{L-1} ... Θ_1 u > 0$ for all $u$, and another that "balances" $Θ_1, ..., Θ_L$ so that they have the same singular values.

cs.LG↗

arXiv2018

The Power of Localization for Efficiently Learning Linear Separators with Noise

We introduce a new approach for designing computationally efficient learning algorithms that are tolerant to noise, and demonstrate its effectiveness by designing algorithms with improved noise tolerance guarantees for learning linear separators. We consider both the malicious noise model and the adversarial label noise model. For malicious noise, where the adversary can corrupt both the label and the features, we provide a polynomial-time algorithm for learning linear separators in $\Re^d$ under isotropic log-concave distributions that can tolerate a nearly information-theoretically optimal noise rate of $η= Ω(ε)$. For the adversarial label noise model, where the distribution over the feature vectors is unchanged, and the overall probability of a noisy label is constrained to be at most $η$, we also give a polynomial-time algorithm for learning linear separators in $\Re^d$ under isotropic log-concave distributions that can handle a noise rate of $η= Ω\left(ε\right)$. We show that, in the active learning model, our algorithms achieve a label complexity whose dependence on the error parameter $ε$ is polylogarithmic. This provides the first polynomial-time active learning algorithm for learning linear separators in the presence of malicious noise or adversarial label noise.

cs.LG↗

arXiv2018

Representing smooth functions as compositions of near-identity functions with implications for deep network optimization

We show that any smooth bi-Lipschitz $h$ can be represented exactly as a composition $h_m \circ ... \circ h_1$ of functions $h_1,...,h_m$ that are close to the identity in the sense that each $\left(h_i-\mathrm{Id}\right)$ is Lipschitz, and the Lipschitz constant decreases inversely with the number $m$ of functions composed. This implies that $h$ can be represented to any accuracy by a deep residual network whose nonlinear layers compute functions with a small Lipschitz constant. Next, we consider nonlinear regression with a composition of near-identity nonlinear maps. We show that, regarding Fréchet derivatives with respect to the $h_1,...,h_m$, any critical point of a quadratic criterion in this near-identity region must be a global minimizer. In contrast, if we consider derivatives with respect to parameters of a fixed-size residual network with sigmoid activation functions, we show that there are near-identity critical points that are suboptimal, even in the realizable case. Informally, this means that functional gradient methods for residual networks cannot get stuck at suboptimal critical points corresponding to near-identity layers, whereas parametric gradient methods for sigmoidal residual networks suffer from suboptimal critical points in the near-identity region.

cs.LG↗

arXiv2017

Surprising properties of dropout in deep networks

We analyze dropout in deep networks with rectified linear units and the quadratic loss. Our results expose surprising differences between the behavior of dropout and more traditional regularizers like weight decay. For example, on some simple data sets dropout training produces negative weights even though the output is the sum of the inputs. This provides a counterpoint to the suggestion that dropout discourages co-adaptation of weights. We also show that the dropout penalty can grow exponentially in the depth of the network while the weight-decay penalty remains essentially linear, and that dropout is insensitive to various re-scalings of the input features, outputs, and network weights. This last insensitivity implies that there are no isolated local minima of the dropout training criterion. Our work uncovers new properties of dropout, extends our understanding of why dropout succeeds, and lays the foundation for further progress.

cs.LG↗

arXiv2015

On the Inductive Bias of Dropout

Dropout is a simple but effective technique for learning in neural networks and other settings. A sound theoretical understanding of dropout is needed to determine when dropout should be applied and how to use it most effectively. In this paper we continue the exploration of dropout as a regularizer pioneered by Wager, et.al. We focus on linear classification where a convex proxy to the misclassification loss (i.e. the logistic loss used in logistic regression) is minimized. We show: (a) when the dropout-regularized criterion has a unique minimizer, (b) when the dropout-regularization penalty goes to infinity with the weights, and when it remains bounded, (c) that the dropout regularization can be non-monotonic as individual weights increase from 0, and (d) that the dropout regularization penalty may not be convex. This last point is particularly surprising because the combination of dropout regularization with any convex loss proxy is always a convex function. In order to contrast dropout regularization with $L_2$ regularization, we formalize the notion of when different sources are more compatible with different regularizers. We then exhibit distributions that are provably more compatible with dropout regularization than $L_2$ regularization, and vice versa. These sources provide additional insight into how the inductive biases of dropout and $L_2$ regularization differ. We provide some similar results for $L_1$ regularization.

cs.LG↗

arXiv2013

Active and passive learning of linear separators under log-concave distributions

We provide new results concerning label efficient, polynomial time, passive and active learning of linear separators. We prove that active learning provides an exponential improvement over PAC (passive) learning of homogeneous linear separators under nearly log-concave distributions. Building on this, we provide a computationally efficient PAC algorithm with optimal (up to a constant factor) sample complexity for such problems. This resolves an open question concerning the sample complexity of efficient PAC algorithms under the uniform distribution in the unit ball. Moreover, it provides the first bound for a polynomial-time PAC algorithm that is tight for an interesting infinite class of hypothesis functions under a general and natural class of data-distributions, providing significant progress towards a longstanding open question. We also provide new bounds for active and passive learning in the case that the data might not be linearly separable, both in the agnostic case and and under the Tsybakov low-noise condition. To derive our results, we provide new structural results for (nearly) log-concave distributions, which might be of independent interest as well.

cs.LG↗

arXiv2012

On the Necessity of Irrelevant Variables

This work explores the effects of relevant and irrelevant boolean variables on the accuracy of classifiers. The analysis uses the assumption that the variables are conditionally independent given the class, and focuses on a natural family of learning algorithms for such sources when the relevant variables have a small advantage over random guessing. The main result is that algorithms relying predominately on irrelevant variables have error probabilities that quickly go to 0 in situations where algorithms that limit the use of irrelevant variables have errors bounded below by a positive constant. We also show that accurate learning is possible even when there are so few examples that one cannot determine with high confidence whether or not any individual variable is relevant.

cs.LG↗