Deep networks often perform well on the data distribution on which they are trained, yet give incorrect (and often very confident) answers when evaluated on points from off of the training distribution. This is exemplified by the adversarial examples phenomenon but can also be seen in terms of model generalization and domain shift. Ideally, a model would assign lower confidence to points unlike those from the training distribution. We propose a regularizer which addresses this issue by training with interpolated hidden states and encouraging the classifier to be less confident at these points. Because the hidden states are learned, this has an important effect of encouraging the hidden states for a class to be concentrated in such a way so that interpolations within the same class or between two different classes do not intersect with the real data points from other classes. This has a major advantage in that it avoids the underfitting which can result from interpolating in the input space. We prove that the exact condition for this problem of underfitting to be avoided by Manifold Mixup is that the dimensionality of the hidden states exceeds the number of classes, which is often the case in practice. Additionally, this concentration can be seen as making the features in earlier layers more discriminative. We show that despite requiring no significant additional computation, Manifold Mixup achieves large improvements over strong baselines in supervised learning, robustness to single-step adversarial attacks, semi-supervised learning, and Negative Log-Likelihood on held out samples.
Games generalize the optimization paradigm by introducing different objective functions for different optimizing agents, known as players. Generative Adversarial Networks (GANs) are arguably the most popular game formulation in recent machine learning literature. GANs achieve great results on generating realistic natural images, however they are known for being difficult to train. Training them involves finding a Nash equilibrium, typically performed using gradient descent on the two players' objectives. Game dynamics can induce rotations that slow down convergence to a Nash equilibrium, or prevent it altogether. We provide a theoretical analysis of the game dynamics. Our analysis, supported by experiments, shows that gradient descent with a negative momentum term can improve the convergence properties of some GANs.
Three-dimensional geometric data offer an excellent domain for studying representation learning and generative modeling. In this paper, we look at geometric data represented as point clouds. We introduce a deep AutoEncoder (AE) network with state-of-the-art reconstruction quality and generalization ability. The learned representations outperform existing methods on 3D recognition tasks and enable shape editing via simple algebraic manipulations, such as semantic part editing, shape analogies and shape interpolation, as well as shape completion. We perform a thorough study of different generative models including GANs operating on the raw point clouds, significantly improved GANs trained in the fixed latent space of our AEs, and Gaussian Mixture Models (GMMs). To quantitatively evaluate generative models we introduce measures of sample fidelity and diversity based on matchings between sets of point clouds. Interestingly, our evaluation of generalization, fidelity and diversity reveals that GMMs trained in the latent space of our AEs yield the best results overall.
Deep networks have achieved impressive results across a variety of important tasks. However a known weakness is a failure to perform well when evaluated on data which differ from the training distribution, even if these differences are very small, as is the case with adversarial examples. We propose Fortified Networks, a simple transformation of existing networks, which fortifies the hidden layers in a deep network by identifying when the hidden states are off of the data manifold, and maps these hidden states back to parts of the data manifold where the network performs well. Our principal contribution is to show that fortifying these hidden states improves the robustness of deep networks and our experiments (i) demonstrate improved robustness to standard adversarial attacks in both black-box and white-box threat models; (ii) suggest that our improvements are not primarily due to the gradient masking problem and (iii) show the advantage of doing this fortification in the hidden layers instead of the input space.
Hyperparameter tuning is one of the most time-consuming workloads in deep learning. State-of-the-art optimizers, such as AdaGrad, RMSProp and Adam, reduce this labor by adaptively tuning an individual learning rate for each variable. Recently researchers have shown renewed interest in simpler methods like momentum SGD as they may yield better test metrics. Motivated by this trend, we ask: can simple adaptive methods based on SGD perform as well or better? We revisit the momentum SGD algorithm and show that hand-tuning a single learning rate and momentum makes it competitive with Adam. We then analyze its robustness to learning rate misspecification and objective curvature variation. Based on these insights, we design YellowFin, an automatic tuner for momentum and learning rate in SGD. YellowFin optionally uses a negative-feedback loop to compensate for the momentum dynamics in asynchronous settings on the fly. We empirically show that YellowFin can converge in fewer iterations than Adam on ResNets and LSTMs for image recognition, language modeling and constituency parsing, with a speedup of up to 3.28x in synchronous and up to 2.69x in asynchronous settings.
This paper presents the first, 15-PetaFLOP Deep Learning system for solving scientific pattern classification problems on contemporary HPC architectures. We develop supervised convolutional architectures for discriminating signals in high-energy physics data as well as semi-supervised architectures for localizing and classifying extreme weather in climate data. Our Intelcaffe-based implementation obtains $\sim$2TFLOP/s on a single Cori Phase-II Xeon-Phi node. We use a hybrid strategy employing synchronous node-groups, while using asynchronous communication across groups. We use this strategy to scale training of a single model to $\sim$9600 Xeon-Phi nodes; obtaining peak performance of 11.73-15.07 PFLOP/s and sustained performance of 11.41-13.27 PFLOP/s. At scale, our HEP architecture produces state-of-the-art classification accuracy on a dataset with 10M images, exceeding that achieved by selections on high-level physics-motivated features. Our semi-supervised architecture successfully extracts weather patterns in a 15TB climate dataset. Our results demonstrate that Deep Learning can be optimized and scaled effectively on many-core, HPC systems.
The pairwise influence matrix of Dobrushin has long been used as an analytical tool to bound the rate of convergence of Gibbs sampling. In this work, we use Dobrushin influence as the basis of a practical tool to certify and efficiently improve the quality of a discrete Gibbs sampler. Our Dobrushin-optimized Gibbs samplers (DoGS) offer customized variable selection orders for a given sampling budget and variable subset of interest, explicit bounds on total variation distance to stationarity, and certifiable improvements over the standard systematic and uniform random scan Gibbs samplers. In our experiments with joint image segmentation and object recognition, Markov chain Monte Carlo maximum likelihood estimation, and Ising model inference, DoGS consistently deliver higher-quality inferences with significantly smaller sampling budgets than standard Gibbs samplers.
Principal component analysis (PCA) is one of the most powerful tools in machine learning. The simplest method for PCA, the power iteration, requires $\mathcal O(1/\Delta)$ full-data passes to recover the principal component of a matrix with eigen-gap $\Delta$. Lanczos, a significantly more complex method, achieves an accelerated rate of $\mathcal O(1/\sqrt{\Delta})$ passes. Modern applications, however, motivate methods that only ingest a subset of available data, known as the stochastic setting. In the online stochastic setting, simple algorithms like Oja's iteration achieve the optimal sample complexity $\mathcal O(\sigma^2/\Delta^2)$. Unfortunately, they are fully sequential, and also require $\mathcal O(\sigma^2/\Delta^2)$ iterations, far from the $\mathcal O(1/\sqrt{\Delta})$ rate of Lanczos. We propose a simple variant of the power iteration with an added momentum term, that achieves both the optimal sample and iteration complexity. In the full-pass setting, standard analysis shows that momentum achieves the accelerated rate, $\mathcal O(1/\sqrt{\Delta})$. We demonstrate empirically that naively applying momentum to a stochastic method, does not result in acceleration. We perform a novel, tight variance analysis that reveals the "breaking-point variance" beyond which this acceleration does not occur. By combining this insight with modern variance reduction techniques, we construct stochastic PCA algorithms, for the online and offline setting, that achieve an accelerated iteration complexity $\mathcal O(1/\sqrt{\Delta})$. Due to the embarassingly parallel nature of our methods, this acceleration translates directly to wall-clock time if deployed in a parallel environment. Our approach is very general, and applies to many non-convex optimization problems that can now be accelerated using the same technique.
Asynchronous methods are widely used in deep learning, but have limited theoretical justification when applied to non-convex problems. We show that running stochastic gradient descent (SGD) in an asynchronous manner can be viewed as adding a momentum-like term to the SGD iteration. Our result does not assume convexity of the objective function, so it is applicable to deep learning systems. We observe that a standard queuing model of asynchrony results in a form of momentum that is commonly used by deep learning practitioners. This forges a link between queuing theory and asynchrony in deep learning systems, which could be useful for systems builders. For convolutional neural networks, we experimentally validate that the degree of asynchrony directly correlates with the momentum, confirming our main result. An important implication is that tuning the momentum parameter is important when considering different levels of asynchrony. We assert that properly tuned momentum reduces the number of steps required for convergence. Finally, our theory suggests new ways of counteracting the adverse effects of asynchrony: a simple mechanism like using negative algorithmic momentum can improve performance under high asynchrony. Since asynchronous methods have better hardware efficiency, this result may shed light on when asynchronous execution is more efficient for deep learning systems.
We study the factors affecting training time in multi-device deep learning systems. Given a specification of a convolutional neural network, our goal is to minimize the time to train this model on a cluster of commodity CPUs and GPUs. We first focus on the single-node setting and show that by using standard batching and data-parallel techniques, throughput can be improved by at least 5.5x over state-of-the-art systems on CPUs. This ensures an end-to-end training speed directly proportional to the throughput of a device regardless of its underlying hardware, allowing each node in the cluster to be treated as a black box. Our second contribution is a theoretical and empirical study of the tradeoffs affecting end-to-end training time in a multiple-device setting. We identify the degree of asynchronous parallelization as a key factor affecting both hardware and statistical efficiency. We see that asynchrony can be viewed as introducing a momentum term. Our results imply that tuning momentum is critical in asynchronous parallel configurations, and suggest that published results that have not been fully tuned might report suboptimal performance for some configurations. For our third contribution, we use our novel understanding of the interaction between system and optimization dynamics to provide an efficient hyperparameter optimizer. Our optimizer involves a predictive model for the total time to convergence and selects an allocation of resources to minimize that time. We demonstrate that the most popular distributed deep learning systems fall within our tradeoff space, but do not optimize within the space. By doing this optimization, our prototype runs 1.9x to 12x faster than the fastest state-of-the-art systems.