ICML 2020 论文列表

Classical linear metric learning methods have recently been extended along two distinct lines: deep metric learning methods for learning embeddings of the data using neural networks, and Bregman divergence learning approaches for extending learning Euclidean distances to more general divergence measures such as divergences over distributions. In this paper, we introduce deep Bregman divergences, which are based on learning and parameterizing functional Bregman divergences using neural networks, and which unify and extend these existing lines of work. We show in particular how deep metric learning formulations, kernel metric learning, Mahalanobis metric learning, and moment-matching functions for comparing distributions arise as special cases of these divergences in the symmetric setting. We then describe a deep learning framework for learning general functional Bregman divergences, and show in experiments that this method yields superior performance on benchmark datasets as compared to existing deep metric learning approaches. We also discuss novel applications, including a semi-supervised distributional clustering problem, and a new loss function for unsupervised data generation.

This paper introduces the deep coordination graph (DCG) for collaborative multi-agent reinforcement learning. DCG strikes a flexible trade-off between representational capacity and generalization by factoring the joint value function of all agents according to a coordination graph into payoffs between pairs of agents. The value can be maximized by local message passing along the graph, which allows training of the value function end-to-end with Q-learning. Payoff functions are approximated with deep neural networks that employ parameter sharing and low-rank approximations to significantly improve sample efficiency. We show that DCG can solve predator-prey tasks that highlight the relative overgeneralization pathology, as well as challenging StarCraft II micromanagement tasks.

A commonly cited inefficiency of neural network training by back-propagation is the update locking problem: each layer must wait for the signal to propagate through the network before updating. In recent years multiple authors have considered alternatives that can alleviate this issue. In this context, we consider a simpler, but more effective, substitute that uses minimal feedback, which we call Decoupled Greedy Learning (DGL). It is based on a greedy relaxation of the joint training objective, recently shown to be effective in the context of Convolutional Neural Networks (CNNs) on large-scale image classification. We consider an optimization of this objective that permits us to decouple the layer training, allowing for layers or modules in networks to be trained with a potentially linear parallelization in layers. We show theoretically and empirically that this approach converges. Then, we empirically find that it can lead to better generalization than sequential greedy optimization and sometimes end-to-end back-propagation. We show an extension of this approach to asynchronous settings, where modules can operate with large communication delays, is possible with the use of a replay buffer. We demonstrate the effectiveness of DGL on the CIFAR-10 dataset against alternatives and on the large-scale ImageNet dataset.

We consider the use of decision trees for decision-making problems under the predict-then-optimize framework. That is, we would like to first use a decision tree to predict unknown input parameters of an optimization problem, and then make decisions by solving the optimization problem using the predicted parameters. A natural loss function in this framework is to measure the suboptimality of the decisions induced by the predicted input parameters, as opposed to measuring loss using input parameter prediction error. This natural loss function is known in the literature as the Smart Predict-then-Optimize (SPO) loss, and we propose a tractable methodology called SPO Trees (SPOTs) for training decision trees under this loss. SPOTs benefit from the interpretability of decision trees, providing an interpretable segmentation of contextual features into groups with distinct optimal solutions to the optimization problem of interest. We conduct several numerical experiments on synthetic and real data including the prediction of travel times for shortest path problems and predicting click probabilities for news article recommendation. We demonstrate on these datasets that SPOTs simultaneously provide higher quality decisions and significantly lower model complexity than other machine learning approaches (e.g., CART) trained to minimize prediction error.

This paper seeks to establish a framework for directing a society of simple, specialized, self-interested agents to solve what traditionally are posed as monolithic single-agent sequential decision problems. What makes it challenging to use a decentralized approach to collectively optimize a central objective is the difficulty in characterizing the equilibrium strategy profile of non-cooperative games. To overcome this challenge, we design a mechanism for defining the learning environment of each agent for which we know that the optimal solution for the global objective coincides with a Nash equilibrium strategy profile of the agents optimizing their own local objectives. The society functions as an economy of agents that learn the credit assignment process itself by buying and selling to each other the right to operate on the environment state. We derive a class of decentralized reinforcement learning algorithms that are broadly applicable not only to standard reinforcement learning but also for selecting options in semi-MDPs and dynamically composing computation graphs. Lastly, we demonstrate the potential advantages of a society’s inherent modular structure for more efficient transfer learning.

We investigate the generalisation performance of Distributed Gradient Descent with implicit regularisation and random features in the homogenous setting where a network of agents are given data sampled independently from the same unknown distribution. Along with reducing the memory footprint, random features are particularly convenient in this setting as they provide a common parameterisation across agents that allows to overcome previous difficulties in implementing decentralised kernel regression. Under standard source and capacity assumptions, we establish high probability bounds on the predictive performance for each agent as a function of the step size, number of iterations, inverse spectral gap of the communication matrix and number of random features. By tuning these parameters, we obtain statistical rates that are minimax optimal with respect to the total number of samples in the network. The algorithm provides a linear improvement over single-machine gradient descent in memory cost and, when agents hold enough data with respect to the network size and inverse spectral gap, a linear speed up in computational run-time for any network topology. We present simulations that show how the number of random features, iterations and samples impact predictive performance.

Entropy regularization in optimal transport (OT) has been the driver of many recent interests for Wasserstein metrics and barycenters in machine learning. It allows to keep the appealing geometrical properties of the unregularized Wasserstein distance while having a significantly lower complexity thanks to Sinkhorn’s algorithm. However, entropy brings some inherent smoothing bias, resulting for example in blurred barycenters. This side effect has prompted an increasing temptation in the community to settle for a slower algorithm such as log-domain stabilized Sinkhorn which breaks the parallel structure that can be leveraged on GPUs, or even go back to unregularized OT. Here we show how this bias is tightly linked to the reference measure that defines the entropy regularizer and propose debiased Sinkhorn barycenters that preserve the best of worlds: fast Sinkhorn-like iterations without entropy smoothing. Theoretically, we prove that this debiasing is perfect for Gaussian distributions with equal variance. Empirically, we illustrate the reduced blurring and the computational advantage.

As machine learning algorithms are increasingly deployed for high-impact automated decision making, ethical and increasingly also legal standards demand that they treat all individuals fairly, without discrimination based on their age, gender, race or other sensitive traits. In recent years much progress has been made on ensuring fairness and reducing bias in standard machine learning settings. Yet, for network embedding, with applications in vulnerable domains ranging from social network analysis to recommender systems, current options remain limited both in number and performance. We thus propose DeBayes: a conceptually elegant Bayesian method that is capable of learning debiased embeddings by using a biased prior. Our experiments show that these representations can then be used to perform link prediction that is significantly more fair in terms of popular metrics such as demographic parity and equalized opportunity.

Human observers can learn to recognize new categories of images from a handful of examples, yet doing so with artificial ones remains an open challenge. We hypothesize that data-efficient recognition is enabled by representations which make the variability in natural signals more predictable. We therefore revisit and improve Contrastive Predictive Coding, an unsupervised objective for learning such representations. This new implementation produces features which support state-of-the-art linear classification accuracy on the ImageNet dataset. When used as input for non-linear classification with deep neural networks, this representation allows us to use 2-5x less labels than classifiers trained directly on image pixels. Finally, this unsupervised representation substantially improves transfer learning to object detection on the PASCAL VOC dataset, surpassing fully supervised pre-trained ImageNet classifiers.

Directed graphical models (DGMs) are a class of probabilistic models that are widely used for predictive analysis in sensitive domains such as medical diagnostics. In this paper, we present an algorithm for differentially-private learning of the parameters of a DGM. Our solution optimizes for the utility of inference queries over the DGM and \emph{adds noise that is customized to the properties of the private input dataset and the graph structure of the DGM}. To the best of our knowledge, this is the first explicit data-dependent privacy budget allocation algorithm in the context of DGMs. We compare our algorithm with a standard data-independent approach over a diverse suite of benchmarks and demonstrate that our solution requires a privacy budget that is roughly $3\times$ smaller to obtain the same or higher utility.

Data containing human or social attributes may over- or under-represent groups with respect to salient social attributes such as gender or race, which can lead to biases in downstream applications. This paper presents an algorithmic framework that can be used as a data preprocessing method towards mitigating such bias. Unlike prior work, it can efficiently learn distributions over large domains, controllably adjust the representation rates of protected groups and achieve target fairness metrics such as statistical parity, yet remains close to the empirical distribution induced by the given dataset. Our approach leverages the principle of maximum entropy {–} amongst all distributions satisfying a given set of constraints, we should choose the one closest in KL-divergence to a given prior. While maximum entropy distributions can succinctly encode distributions over large domains, they can be difficult to compute. Our main contribution is an instantiation of this framework for our set of constraints and priors, which encode our bias mitigation goals, and that runs in time polynomial in the dimension of the data. Empirically, we observe that samples from the learned distribution have desired representation rates and statistical rates, and when used for training a classifier incurs only a slight loss in accuracy while maintaining fairness properties.

Quantifying the value of data is a fundamental problem in machine learning and has multiple important use cases: (1) building insights about the dataset and task, (2) domain adaptation, (3) corrupted sample discovery, and (4) robust learning. We propose Data Valuation using Reinforcement Learning (DVRL), to adaptively learn data values jointly with the predictor model. DVRL uses a data value estimator (DVE) to learn how likely each datum is used in training of the predictor model. DVE is trained using a reinforcement signal that reflects performance on the target task. We demonstrate that DVRL yields superior data value estimates compared to alternative methods across numerous datasets and application scenarios. The corrupted sample discovery performance of DVRL is close to optimal in many regimes (i.e. as if the noisy samples were known apriori), and for domain adaptation and robust learning DVRL significantly outperforms state-of-the-art by 14.6% and 10.8%, respectively.

The best-known and most commonly used technique for distribution-property estimation uses a plug-in estimator, with empirical frequency replacing the underlying distribution. We present novel linear-time-computable estimators that significantly “amplify” the effective amount of data available. For a large variety of distribution properties including four of the most popular ones and for every underlying distribution, they achieve the accuracy that the empirical-frequency plug-in estimators would attain using a logarithmic-factor more samples. Specifically, for Shannon entropy and a broad class of Lipschitz properties including the $L_1$ distance to a fixed distribution, the new estimators use $n$ samples to achieve the accuracy attained by the empirical estimators with $n\log n$ samples. For support-size and coverage, the new estimators use $n$ samples to achieve the performance of empirical frequency with sample size $n$ times the logarithm of the property value. Significantly strengthening the traditional min-max formulation, these results hold not only for the worst distributions, but for each and every underlying distribution. Furthermore, the logarithmic amplification factors are optimal. Experiments on a wide variety of distributions show that the new estimators outperform the previous state-of-the-art estimators designed for each specific property.

Differentiable renderers have been used successfully for unsupervised 3D structure learning from 2D images because they can bridge the gap between 3D and 2D. To optimize 3D shape parameters, current renderers rely on pixel-wise losses between rendered images of 3D reconstructions and ground truth images from corresponding viewpoints. Hence they require interpolation of the recovered 3D structure at each pixel, visibility handling, and optionally evaluating a shading model. In contrast, here we propose a Differentiable Renderer Without Rendering (DRWR) that omits these steps. DRWR only relies on a simple but effective loss that evaluates how well the projections of reconstructed 3D point clouds cover the ground truth object silhouette. Specifically, DRWR employs a smooth silhouette loss to pull the projection of each individual 3D point inside the object silhouette, and a structure-aware repulsion loss to push each pair of projections that fall inside the silhouette far away from each other. Although we omit surface interpolation, visibility handling, and shading, our results demonstrate that DRWR achieves state-of-the-art accuracies under widely used benchmarks, outperforming previous methods both qualitatively and quantitatively. In addition, our training times are significantly lower due to the simplicity of DRWR.

Classical approaches for one-class problems such as one-class SVM and isolation forest require careful feature engineering when applied to structured domains like images. State-of-the-art methods aim to leverage deep learning to learn appropriate features via two main approaches. The first approach based on predicting transformations (Golan & El-Yaniv, 2018; Hendrycks et al., 2019a) while successful in some domains, crucially depends on an appropriate domain-specific set of transformations that are hard to obtain in general. The second approach of minimizing a classical one-class loss on the learned final layer representations, e.g., DeepSVDD (Ruff et al., 2018) suffers from the fundamental drawback of representation collapse. In this work, we propose Deep Robust One Class Classification (DROCC) that is both applicable to most standard domains without requiring any side-information and robust to representation collapse. DROCC is based on the assumption that the points from the class of interest lie on a well-sampled, locally linear low dimensional manifold. Empirical evaluation demonstrates that DROCC is highly effective in two different one-class problem settings and on a range of real-world datasets across different domains: tabular data, images (CIFAR and ImageNet), audio, and time-series, offering up to 20% increase in accuracy over the state-of-the-art in anomaly detection. Code is available at https://github.com/microsoft/EdgeML

We present a novel communication-efficient Newton-type algorithm for finite-sum optimization over a distributed computing environment. Our method, named DINO, overcomes both theoretical and practical shortcomings of similar existing methods. Under minimal assumptions, we guarantee global sub-linear convergence of DINO to a first-order stationary point for general non-convex functions and arbitrary data distribution over the network. Furthermore, for functions satisfying Polyak-Lojasiewicz (PL) inequality, we show that DINO enjoys a linear convergence rate. Our proposed algorithm is practically parameter free, in that it will converge regardless of the selected hyper-parameters, which are easy to tune. Additionally, its sub-problems are simple linear least-squares, for which efficient solvers exist, and numerical simulations demonstrate the efficiency of DINO as compared with similar alternatives.

A popular line of recent research incorporates ML advice in the design of online algorithms to improve their performance in typical instances. These papers treat the ML algorithm as a black-box, and redesign online algorithms to take advantage of ML predictions. In this paper, we ask the complementary question: can we redesign ML algorithms to provide better predictions for online algorithms? We explore this question in the context of the classic rent-or-buy problem, and show that incorporating optimization benchmarks in ML loss functions leads to significantly better performance, while maintaining a worst-case adversarial result when the advice is completely wrong. We support this finding both through theoretical bounds and numerical simulations.

Deep neural networks exhibit complex learning dynamics due to their non-convex loss landscapes. Second-order optimization methods facilitate learning dynamics by compensating for ill-conditioned curvature. In this work, we investigate how curvature correction modifies the learning dynamics in deep linear neural networks and provide analytical solutions. We derive a generalized conservation law that preserves the path of parameter dynamics from curvature correction, which shows that curvature correction only modifies the temporal profiles of dynamics along the path. We show that while curvature correction accelerates the convergence dynamics of the input-output map, it can also negatively affect the generalization performance. Our analysis also reveals an undesirable effect of curvature correction that compromises stability of parameters dynamics during learning, especially with block-diagonal approximation of natural gradient descent. We introduce fractional curvature correction that resolves this problem while retaining most of the acceleration benefits of full curvature correction.

Randomized smoothing, using just a simple isotropic Gaussian distribution, has been shown to produce good robustness guarantees against $\ell_2$-norm bounded adversaries. In this work, we show that extending the smoothing technique to defend against other attack models can be challenging, especially in the high-dimensional regime. In particular, for a vast class of i.i.d. smoothing distributions, we prove that the largest $\ell_p$-radius that can be certified decreases as $O(1/d^{\frac{1}{2} - \frac{1}{p}})$ with dimension $d$ for $p > 2$. Notably, for $p \geq 2$, this dependence on $d$ is no better than that of the $\ell_p$-radius that can be certified using isotropic Gaussian smoothing, essentially putting a matching lower bound on the robustness radius. When restricted to \emph{generalized} Gaussian smoothing, these two bounds can be shown to be within a constant factor of each other in an asymptotic sense, establishing that Gaussian smoothing provides the best possible results, up to a constant factor, when $p \geq 2$. We present experimental results on CIFAR to validate our theory. For other smoothing distributions, such as, a uniform distribution within an $\ell_1$ or an $\ell_\infty$-norm ball, we show upper bounds of the form $O(1 / d)$ and $O(1 / d^{1 - \frac{1}{p}})$ respectively, which have an even worse dependence on $d$.