ICML 2018 Papers

There is growing interest in automating neural network architecture design. Existing architecture search methods can be computationally expensive, requiring thousands of different architectures to be trained from scratch. Recent work has explored

This work aims to provide understandings on the remarkable success of deep convolutional neural networks (CNNs) by theoretically analyzing their generalization performance and establishing optimization guarantees for gradient descent based training algorithms. Specifically, for a CNN model consisting of $l$ convolutional layers and one fully connected layer, we prove that its generalization error is bounded by $\mathcal{O}(\sqrt{\theta\widetilde{\varrho}/n})$ where $\theta$ denotes freedom degree of the network parameters and $\widetilde{\varrho}=\mathcal{O}(\log(\prod_{i=1}^{l}b_{i} (k_{i}-s_{i}+1)/p)+\log(b_{l+1}))$ encapsulates architecture parameters including the kernel size $k_{i}$, stride $s_{i}$, pooling size $p$ and parameter magnitude $b_{i}$. To our best knowledge, this is the first generalization bound that only depends on $\mathcal{O}(\log(\prod_{i=1}^{l+1}b_{i}))$, tighter than existing ones that all involve an exponential term like $\mathcal{O}(\prod_{i=1}^{l+1}b_{i})$. Besides, we prove that for an arbitrary gradient descent algorithm, the computed approximate stationary point by minimizing empirical risk is also an approximate stationary point to the population risk. This well explains why gradient descent training algorithms usually perform sufficiently well in practice. Furthermore, we prove the one-to-one correspondence and convergence guarantees for the non-degenerate stationary points between the empirical and population risks. It implies that the computed local minimum for the empirical risk is also close to a local minimum for the population risk, thus ensuring that the optimized CNN model well generalizes to new data.

For data-driven decision-making, one promising approach, called predictive optimization, is to solve maximization problems i n which the objective function to be maximized is estimated from data. Predictive optimization, however, suffers from the problem of a calculated optimal solution’s being evaluated too optimistically, i.e., the value of the objective function is overestimated. This paper investigates such optimistic bias and presents two methods for correcting it. The first, which is analogous to cross-validation, successfully corrects the optimistic bias but results in underestimation of the true value. Our second method employs resampling techniques to avoid both overestimation and underestimation. We show that the second method, referred to as the parameter perturbation method, achieves asymptotically unbiased estimation. Empirical results for both artificial and real-world datasets demonstrate that our proposed approach successfully corrects the optimistic bias.

We establish, for the first time, explicit connections between feedforward neural networks with ReLU activation and tropical geometry — we show that the family of such neural networks is equivalent to the family of tropical rational maps. Among other things, we deduce that feedforward ReLU neural networks with one hidden layer can be characterized by zonotopes, which serve as building blocks for deeper networks; we relate decision boundaries of such neural networks to tropical hypersurfaces, a major object of study in tropical geometry; and we prove that linear regions of such neural networks correspond to vertices of polytopes associated with tropical rational functions. An insight from our tropical formulation is that a deeper network is exponentially more expressive than a shallow network.

Metric learning has the aim to improve classification accuracy by learning a distance measure which brings data points from the same class closer together and pushes data points from different classes further apart. Recent research has demonstrated that metric learning approaches can also be applied to trees, such as molecular structures, abstract syntax trees of computer programs, or syntax trees of natural language, by learning the cost function of an edit distance, i.e. the costs of replacing, deleting, or inserting nodes in a tree. However, learning such costs directly may yield an edit distance which violates metric axioms, is challenging to interpret, and may not generalize well. In this contribution, we propose a novel metric learning approach for trees which we call embedding edit distance learning (BEDL) and which learns an edit distance indirectly by embedding the tree nodes as vectors, such that the Euclidean distance between those vectors supports class discrimination. We learn such embeddings by reducing the distance to prototypical trees from the same class and increasing the distance to prototypical trees from different classes. In our experiments, we show that BEDL improves upon the state-of-the-art in metric learning for trees on six benchmark data sets, ranging from computer science over biomedical data to a natural-language processing data set containing over 300,000 nodes.

The fundamental task of general density estimation $p(x)$ has been of keen interest to machine learning. In this work, we attempt to systematically characterize methods for density estimation. Broadly speaking, most of the existing methods can be categorized into either using:

The ability to transfer skills across tasks has the potential to scale up reinforcement learning (RL) agents to environments currently out of reach. Recently, a framework based on two ideas, successor features (SFs) and generalised policy improvement (GPI), has been introduced as a principled way of transferring skills. In this paper we extend the SF&GPI framework in two ways. One of the basic assumptions underlying the original formulation of SF&GPI is that rewards for all tasks of interest can be computed as linear combinations of a fixed set of features. We relax this constraint and show that the theoretical guarantees supporting the framework can be extended to any set of tasks that only differ in the reward function. Our second contribution is to show that one can use the reward functions themselves as features for future tasks, without any loss of expressiveness, thus removing the need to specify a set of features beforehand. This makes it possible to combine SF&GPI with deep learning in a more stable way. We empirically verify this claim on a complex 3D environment where observations are images from a first-person perspective. We show that the transfer promoted by SF&GPI leads to very good policies on unseen tasks almost instantaneously. We also describe how to learn policies specialised to the new tasks in a way that allows them to be added to the agent’s set of skills, and thus be reused in the future.

In transfer learning, what and how to transfer are two primary issues to be addressed, as different transfer learning algorithms applied between a source and a target domain result in different knowledge transferred and thereby the performance improvement in the target domain. Determining the optimal one that maximizes the performance improvement requires either exhaustive exploration or considerable expertise. Meanwhile, it is widely accepted in educational psychology that human beings improve transfer learning skills of deciding what to transfer through meta-cognitive reflection on inductive transfer learning practices. Motivated by this, we propose a novel transfer learning framework known as Learning to Transfer (L2T) to automatically determine what and how to transfer are the best by leveraging previous transfer learning experiences. We establish the L2T framework in two stages: 1) we learn a reflection function encrypting transfer learning skills from experiences; and 2) we infer what and how to transfer are the best for a future pair of domains by optimizing the reflection function. We also theoretically analyse the algorithmic stability and generalization bound of L2T, and empirically demonstrate its superiority over several state-of-the-art transfer learning algorithms.

We investigate the effectiveness of trace-based supervision methods for training existing neural abstract machines. To define the class of neural machines amenable to trace-based supervision, we introduce the concept of a differential neural computational machine (dNCM) and show that several existing architectures (NTMs, NRAMs) can be described as dNCMs. We performed a detailed experimental evaluation with NTM and NRAM machines, showing that additional supervision on the interpretable portions of these architectures leads to better convergence and generalization capabilities of the learning phase than standard training, in both noise-free and noisy scenarios.

Modern neural networks have recently been found to be poorly calibrated, primarily in the direction of over-confidence. Methods like entropy penalty and temperature smoothing improve calibration by clamping confidence, but in doing so compromise the many legitimately confident predictions. We propose a more principled fix that minimizes an explicit calibration error during training. We present MMCE, a RKHS kernel based measure of calibration that is efficiently trainable alongside the negative likelihood loss without careful hyper-parameter tuning. Theoretically too, MMCE is a sound measure of calibration that is minimized at perfect calibration, and whose finite sample estimates are consistent and enjoy fast convergence rates. Extensive experiments on several network architectures demonstrate that MMCE is a fast, stable, and accurate method to minimize calibration error while maximally preserving the number of high confidence predictions.

Recently, the decentralized optimization problem is attracting growing attention. Most existing methods are deterministic with high per-iteration cost and have a convergence rate quadratically depending on the problem condition number. Besides, the dense communication is necessary to ensure the convergence even if the dataset is sparse. In this paper, we generalize the decentralized optimization problem to a monotone operator root finding problem, and propose a stochastic algorithm named DSBA that (1) converges geometrically with a rate linearly depending on the problem condition number, and (2) can be implemented using sparse communication only. Additionally, DSBA handles important learning problems like AUC-maximization which can not be tackled efficiently in the previous problem setting. Experiments on convex minimization and AUC-maximization validate the efficiency of our method.

Verifying the robustness property of a general Rectified Linear Unit (ReLU) network is an NP-complete problem. Although finding the exact minimum adversarial distortion is hard, giving a certified lower bound of the minimum distortion is possible. Current available methods of computing such a bound are either time-consuming or deliver low quality bounds that are too loose to be useful. In this paper, we exploit the special structure of ReLU networks and provide two computationally efficient algorithms (Fast-Lin, Fast-Lip) that are able to certify non-trivial lower bounds of minimum adversarial distortions. Experiments show that (1) our methods deliver bounds close to (the gap is 2-3X) exact minimum distortions found by Reluplex in small networks while our algorithms are more than 10,000 times faster; (2) our methods deliver similar quality of bounds (the gap is within 35% and usually around 10%; sometimes our bounds are even better) for larger networks compared to the methods based on solving linear programming problems but our algorithms are 33-14,000 times faster; (3) our method is capable of solving large MNIST and CIFAR networks up to 7 layers with more than 10,000 neurons within tens of seconds on a single CPU core. In addition, we show that there is no polynomial time algorithm that can approximately find the minimum $\ell_1$ adversarial distortion of a ReLU network with a $0.99\ln n$ approximation ratio unless NP=P, where $n$ is the number of neurons in the network.

We present an extension to the Tacotron speech synthesis architecture that learns a latent embedding space of prosody, derived from a reference acoustic representation containing the desired prosody. We show that conditioning Tacotron on this learned embedding space results in synthesized audio that matches the prosody of the reference signal with fine time detail even when the reference and synthesis speakers are different. Additionally, we show that a reference prosody embedding can be used to synthesize text that is different from that of the reference utterance. We define several quantitative and subjective metrics for evaluating prosody transfer, and report results with accompanying audio samples from single-speaker and 44-speaker Tacotron models on a prosody transfer task.

In this paper, we make an important step towards the black-box machine teaching by considering the cross-space machine teaching, where the teacher and the learner use different feature representations and the teacher can not fully observe the learner’s model. In such scenario, we study how the teacher is still able to teach the learner to achieve faster convergence rate than the traditional passive learning. We propose an active teacher model that can actively query the learner (i.e., make the learner take exams) for estimating the learner’s status and provably guide the learner to achieve faster convergence. The sample complexities for both teaching and query are provided. In the experiments, we compare the proposed active teacher with the omniscient teacher and verify the effectiveness of the active teacher model.

Long Short-Term Memory (LSTM) is one of the most widely used recurrent structures in sequence modeling. It aims to use gates to control information flow (e.g., whether to skip some information or not) in the recurrent computations, although its practical implementation based on soft gates only partially achieves this goal. In this paper, we propose a new way for LSTM training, which pushes the output values of the gates towards 0 or 1. By doing so, we can better control the information flow: the gates are mostly open or closed, instead of in a middle state, which makes the results more interpretable. Empirical studies show that (1) Although it seems that we restrict the model capacity, there is no performance drop: we achieve better or comparable performances due to its better generalization ability; (2) The outputs of gates are not sensitive to their inputs: we can easily compress the LSTM unit in multiple ways, e.g., low-rank approximation and low-precision approximation. The compressed models are even better than the baseline models without compression.

We introduce topological mixture estimation, a completely nonparametric and computationally efficient solution to the problem of estimating a one-dimensional mixture with generic unimodal components. We repeatedly perturb the unimodal decomposition of Baryshnikov and Ghrist to produce a topologically and information-theoretically optimal unimodal mixture. We also detail a smoothing process that optimally exploits topological persistence of the unimodal category in a natural way when working directly with sample data. Finally, we illustrate these techniques through examples.

Generalization performance of classifiers in deep learning has recently become a subject of intense study. Deep models, which are typically heavily over-parametrized, tend to fit the training data exactly. Despite this “overfitting", they perform well on test data, a phenomenon not yet fully understood. The first point of our paper is that strong performance of overfitted classifiers is not a unique feature of deep learning. Using six real-world and two synthetic datasets, we establish experimentally that kernel machines trained to have zero classification error or near zero regression error (interpolation) perform very well on test data. We proceed to give a lower bound on the norm of zero loss solutions for smooth kernels, showing that they increase nearly exponentially with data size. None of the existing bounds produce non-trivial results for interpolating solutions. We also show experimentally that (non-smooth) Laplacian kernels easily fit random labels, a finding that parallels results recently reported for ReLU neural networks. In contrast, fitting noisy data requires many more epochs for smooth Gaussian kernels. Similar performance of overfitted Laplacian and Gaussian classifiers on test, suggests that generalization is tied to the properties of the kernel function rather than the optimization process. Some key phenomena of deep learning are manifested similarly in kernel methods in the modern “overfitted" regime. The combination of the experimental and theoretical results presented in this paper indicates a need for new theoretical ideas for understanding properties of classical kernel methods. We argue that progress on understanding deep learning will be difficult until more tractable “shallow” kernel methods are better understood.

In reinforcement learning, it is common to let an agent interact for a fixed amount of time with its environment before resetting it and repeating the process in a series of episodes. The task that the agent has to learn can either be to maximize its performance over (i) that fixed period, or (ii) an indefinite period where time limits are only used during training to diversify experience. In this paper, we provide a formal account for how time limits could effectively be handled in each of the two cases and explain why not doing so can cause state-aliasing and invalidation of experience replay, leading to suboptimal policies and training instability. In case (i), we argue that the terminations due to time limits are in fact part of the environment, and thus a notion of the remaining time should be included as part of the agent’s input to avoid violation of the Markov property. In case (ii), the time limits are not part of the environment and are only used to facilitate learning. We argue that this insight should be incorporated by bootstrapping from the value of the state at the end of each partial episode. For both cases, we illustrate empirically the significance of our considerations in improving the performance and stability of existing reinforcement learning algorithms, showing state-of-the-art results on several control tasks.

We provide theoretical and empirical evidence that using tighter evidence lower bounds (ELBOs) can be detrimental to the process of learning an inference network by reducing the signal-to-noise ratio of the gradient estimator. Our results call into question common implicit assumptions that tighter ELBOs are better variational objectives for simultaneous model learning and inference amortization schemes. Based on our insights, we introduce three new algorithms: the partially importance weighted auto-encoder (PIWAE), the multiply importance weighted auto-encoder (MIWAE), and the combination importance weighted autoencoder (CIWAE), each of which includes the standard importance weighted auto-encoder (IWAE) as a special case. We show that each can deliver improvements over IWAE, even when performance is measured by the IWAE target itself. Furthermore, our results suggest that PIWAE may be able to deliver simultaneous improvements in the training of both the inference and generative networks.