NeurIPS 2019 论文列表

Transfer learning from natural image datasets, particularly ImageNet, using standard large models and corresponding pretrained weights has become a de-facto method for deep learning applications to medical imaging. However, there are fundamental differences in data sizes, features and task specifications between natural image classification and the target medical tasks, and there is little understanding of the effects of transfer. In this paper, we explore properties of transfer learning for medical imaging. A performance evaluation on two large scale medical imaging tasks shows that surprisingly, transfer offers little benefit to performance, and simple, lightweight models can perform comparably to ImageNet architectures. Investigating the learned representations and features, we find that some of the differences from transfer learning are due to the over-parametrization of standard models rather than sophisticated feature reuse. We isolate where useful feature reuse occurs, and outline the implications for more efficient model exploration. We also explore feature independent benefits of transfer arising from weight scalings.

Deep neural networks (DNNs) excel at learning representations when trained on large-scale datasets. Pre-trained DNNs also show strong transferability when fine-tuned to other labeled datasets. However, such transferability becomes weak when the target dataset is fully unlabeled as in Unsupervised Domain Adaptation (UDA). We envision that the loss of transferability may stem from the intrinsic limitation of the architecture design of DNNs. In this paper, we delve into the components of DNN architectures and propose Transferable Normalization (TransNorm) in place of existing normalization techniques. TransNorm is an end-to-end trainable layer to make DNNs more transferable across domains. As a general method, TransNorm can be easily applied to various deep neural networks and domain adaption methods, without introducing any extra hyper-parameters or learnable parameters. Empirical results justify that TransNorm not only improves classification accuracies but also accelerates convergence for mainstream DNN-based domain adaptation methods.

We propose a new principle for transfer learning, based on a straightforward intuition: if two domains are similar to each other, the model trained on one domain should also perform well on the other domain, and vice versa. To formalize this intuition, we define the performance gap as a measure of the discrepancy between the source and target domains. We derive generalization bounds for the instance weighting approach to transfer learning, showing that the performance gap can be viewed as an algorithm-dependent regularizer, which controls the model complexity. Our theoretical analysis provides new insight into transfer learning and motivates a set of general, principled rules for designing new instance weighting schemes for transfer learning. These rules lead to gapBoost, a novel and principled boosting approach for transfer learning. Our experimental evaluation on benchmark data sets shows that gapBoost significantly outperforms previous boosting-based transfer learning algorithms.

We propose a method to improve the anomaly detection performance on target domains by transferring knowledge on related domains. Although anomaly labels are valuable to learn anomaly detectors, they are difficult to obtain due to their rarity. To alleviate this problem, existing methods use anomalous and normal instances in the related domains as well as target normal instances. These methods require training on each target domain. However, this requirement can be problematic in some situations due to the high computational cost of training. The proposed method can infer the anomaly detectors for target domains without re-training by introducing the concept of latent domain vectors, which are latent representations of the domains and are used for inferring the anomaly detectors. The latent domain vector for each domain is inferred from the set of normal instances in the domain. The anomaly score function for each domain is modeled on the basis of autoencoders, and its domain-specific property is controlled by the latent domain vector. The anomaly score function for each domain is trained so that the scores of normal instances become low and the scores of anomalies become higher than those of the normal instances, while considering the uncertainty of the latent domain vectors. When target normal instances can be used during training, the proposed method can also use them for training in a unified framework. The effectiveness of the proposed method is demonstrated through experiments using one synthetic and four real-world datasets. Especially, the proposed method without re-training outperforms existing methods with target specific training.

To recognize objects of the unseen classes, most existing Zero-Shot Learning (ZSL) methods first learn a compatible projection function between the common semantic space and the visual space based on the data of source seen classes, then directly apply it to the target unseen classes. However, in real scenarios, the data distribution between the source and target domain might not match well, thus causing the well-known domain shift problem. Based on the observation that visual features of test instances can be separated into different clusters, we propose a new visual structure constraint on class centers for transductive ZSL, to improve the generality of the projection function (\ie alleviate the above domain shift problem). Specifically, three different strategies (symmetric Chamfer-distance,Bipartite matching distance, and Wasserstein distance) are adopted to align the projected unseen semantic centers and visual cluster centers of test instances. We also propose a new training strategy to handle the real cases where many unrelated images exist in the test dataset, which is not considered in previous methods. Experiments on many widely used datasets demonstrate that the proposed visual structure constraint can bring substantial performance gain consistently and achieve state-of-the-art results.

The alternating direction method of multipliers (ADMM) is one of the most widely used first-order optimisation methods in the literature owing to its simplicity, flexibility and efficiency. Over the years, numerous efforts are made to improve the performance of the method, such as the inertial technique. By studying the geometric properties of ADMM, we discuss the limitations of current inertial accelerated ADMM and then present and analyze an adaptive acceleration scheme for the method. Numerical experiments on problems arising from image processing, statistics and machine learning demonstrate the advantages of the proposed acceleration approach.

Generative Adversarial Networks (GANs) enjoy great success at image generation, but have proven difficult to train in the domain of natural language. Challenges with gradient estimation, optimization instability, and mode collapse have lead practitioners to resort to maximum likelihood pre-training, followed by small amounts of adversarial fine-tuning. The benefits of GAN fine-tuning for language generation are unclear, as the resulting models produce comparable or worse samples than traditional language models. We show it is in fact possible to train a language GAN from scratch --- without maximum likelihood pre-training. We combine existing techniques such as large batch sizes, dense rewards and discriminator regularization to stabilize and improve language GANs. The resulting model, ScratchGAN, performs comparably to maximum likelihood training on EMNLP2017 News and WikiText-103 corpora according to quality and diversity metrics.

Deep neural networks have been very successful in compressive-sensing and image restoration applications, as a means to estimate images from partial, blurry, or otherwise degraded measurements. These networks are trained on a large number of corresponding pairs of measurements and ground-truth images, and thus implicitly learn to exploit domain-specific image statistics. But unlike measurement data, it is often expensive or impractical to collect a large training set of ground-truth images in many application settings. In this paper, we introduce an unsupervised framework for training image estimation networks, from a training set that contains only measurements---with two varied measurements per image---but no ground-truth for the full images desired as output. We demonstrate that our framework can be applied for both regular and blind image estimation tasks, where in the latter case parameters of the measurement model (e.g., the blur kernel) are unknown: during inference, and potentially, also during training. We evaluate our framework for training networks for compressive-sensing and blind deconvolution, considering both non-blind and blind training for the latter. Our framework yields models that are nearly as accurate as those from fully supervised training, despite not having access to any ground-truth images.

Neural architecture search methods are able to find high performance deep learning architectures with minimal effort from an expert. However, current systems focus on specific use-cases (e.g. convolutional image classifiers and recurrent language models), making them unsuitable for general use-cases that an expert might wish to write. Hyperparameter optimization systems are general-purpose but lack the constructs needed for easy application to architecture search. In this work, we propose a formal language for encoding search spaces over general computational graphs. The language constructs allow us to write modular, composable, and reusable search space encodings and to reason about search space design. We use our language to encode search spaces from the architecture search literature. The language allows us to decouple the implementations of the search space and the search algorithm, allowing us to expose search spaces to search algorithms through a consistent interface. Our experiments show the ease with which we can experiment with different combinations of search spaces and search algorithms without having to implement each combination from scratch. We release an implementation of our language with this paper.

Amongst the very first variance reduced stochastic methods for solving the empirical risk minimization problem was the SVRG method. SVRG is an inner-outer loop based method, where in the outer loop a reference full gradient is evaluated, after which $m \in \N$ steps of an inner loop are executed where the reference gradient is used to build a variance reduced estimate of the current gradient. The simplicity of the SVRG method and its analysis have lead to multiple extensions and variants for even non-convex optimization. Yet there is a significant gap between the parameter settings that the analysis suggests and what is known to work well in practice. Our first contribution is that we take several steps towards closing this gap. In particular, the current analysis shows that $m$ should be of the order of the condition number so that the resulting method has a favorable complexity. Yet in practice $m=n$ works well regardless of the condition number, where $n$ is the number of data points. Furthermore, the current analysis shows that the inner iterates have to be reset using averaging after every outer loop. Yet in practice SVRG works best when the inner iterates are updated continuously and not reset. We provide an analysis of these aforementioned practical settings and show that they achieve the same favorable complexity as the original analysis (with slightly better constants). Our second contribution is to provide a more general analysis than had been previously done by using arbitrary sampling, which allows us to analyze virtually all forms of mini-batching through a single theorem. Since our setup and analysis reflect what is done in practice, we are able to set the parameters such as the mini-batch size and step size using our theory in such a way that produces a more efficient algorithm in practice, as we show in extensive numerical experiments.

There are a number of approximation algorithms for NP-hard versions of low rank approximation, such as finding a rank-$k$ matrix $B$ minimizing the sum of absolute values of differences to a given $n$-by-$n$ matrix $A$, $\min_{\textrm{rank-}k~B}\|A-B\|_1$, or more generally finding a rank-$k$ matrix $B$ which minimizes the sum of $p$-th powers of absolute values of differences, $\min_{\textrm{rank-}k~B}\|A-B\|_p^p$. Many of these algorithms are linear time columns subset selection algorithms, returning a subset of $\poly(k \log n)$ columns whose cost is no more than a $\poly(k)$ factor larger than the cost of the best rank-$k$ matrix. The above error measures are special cases of the following general entrywise low rank approximation problem: given an arbitrary function $g:\mathbb{R} \rightarrow \mathbb{R}_{\geq 0}$, find a rank-$k$ matrix $B$ which minimizes $\|A-B\|_g = \sum_{i,j}g(A_{i,j}-B_{i,j})$. A natural question is which functions $g$ admit efficient approximation algorithms? Indeed, this is a central question of recent work studying generalized low rank models. In this work we give approximation algorithms for {\it every} function $g$ which is approximately monotone and satisfies an approximate triangle inequality, and we show both of these conditions are necessary. Further, our algorithm is efficient if the function $g$ admits an efficient approximate regression algorithm. Our approximation algorithms handle functions which are not even scale-invariant, such as the Huber loss function, which we show have very different structural properties than $\ell_p$-norms, e.g., one can show the lack of scale-invariance causes any column subset selection algorithm to provably require a $\sqrt{\log n}$ factor larger number of columns than $\ell_p$-norms; nevertheless we design the first efficient column subset selection algorithms for such error measures.

Residual Network (ResNet) is undoubtedly a milestone in deep learning. ResNet is equipped with shortcut connections between layers, and exhibits efficient training using simple first order algorithms. Despite of the great empirical success, the reason behind is far from being well understood. In this paper, we study a two-layer non-overlapping convolutional ResNet. Training such a network requires solving a non-convex optimization problem with a spurious local optimum. We show, however, that gradient descent combined with proper normalization, avoids being trapped by the spurious local optimum, and converges to a global optimum in polynomial time, when the weight of the first layer is initialized at 0, and that of the second layer is initialized arbitrarily in a ball. Numerical experiments are provided to support our theory.

Alternating least-squares (ALS) is a simple yet effective solver for canonical correlation analysis (CCA). In terms of ease of use, ALS is arguably practitioners' first choice. Despite recent provably guaranteed variants, the empirical performance often remains unsatisfactory. To promote the practical use of ALS for CCA, we propose truly alternating least-squares. Instead of approximately solving two independent linear systems, in each iteration, it simply solves two coupled linear systems of half the size. It turns out that this coupling procedure is able to bring significant performance improvements in practice. Inspired by accelerated power method, we further propose faster alternating least-squares, where momentum terms are introduced into the update equations. Both algorithms enjoy linear convergence. To make faster ALS even more practical, we put forward adaptive alternating least-squares to avoid tuning the momentum parameter, which is as easy to use as the plain ALS while retaining advantages of the fast version. Experiments on several datasets empirically demonstrate the superiority of the proposed algorithms to recent variants.

Motivated by the many real-world applications of reinforcement learning (RL) that require safe-policy iterations, we consider the problem of off-policy evaluation (OPE) --- the problem of evaluating a new policy using the historical data obtained by different behavior policies --- under the model of nonstationary episodic Markov Decision Processes (MDP) with a long horizon and a large action space. Existing importance sampling (IS) methods often suffer from large variance that depends exponentially on the RL horizon $H$. To solve this problem, we consider a marginalized importance sampling (MIS) estimator that recursively estimates the state marginal distribution for the target policy at every step. MIS achieves a mean-squared error of $$ \frac{1}{n} \sum_{t=1}^H\mathbb{E}_{\mu}\left[\frac{d_t^\pi(s_t)^2}{d_t^\mu(s_t)^2} \Var_{\mu}\left[\frac{\pi_t(a_t|s_t)}{\mu_t(a_t|s_t)}\big( V_{t+1}^\pi(s_{t+1}) + r_t\big) \middle| s_t\right]\right] + \tilde{O}(n^{-1.5}) $$ where $\mu$ and $\pi$ are the logging and target policies, $d_t^{\mu}(s_t)$ and $d_t^{\pi}(s_t)$ are the marginal distribution of the state at $t$th step, $H$ is the horizon, $n$ is the sample size and $V_{t+1}^\pi$ is the value function of the MDP under $\pi$. The result matches the Cramer-Rao lower bound in [Jiang and Li, 2016] up to a multiplicative factor of $H$. To the best of our knowledge, this is the first OPE estimation error bound with a polynomial dependence on $H$. Besides theory, we show empirical superiority of our method in time-varying, partially observable, and long-horizon RL environments.

Inspired by recent work in attention models for image captioning and question answering, we present a soft attention model for the reinforcement learning domain. This model bottlenecks the view of an agent by a soft, top-down attention mechanism, forcing the agent to focus on task-relevant information by sequentially querying its view of the environment. The output of the attention mechanism allows direct observation of the information used by the agent to select its actions, enabling easier interpretation of this model than of traditional models. We analyze the different strategies the agents learn and show that a handful of strategies arise repeatedly across different games. We also show that the model learns to query separately about space and content (

Smart portable applications increasingly rely on edge computing due to privacy and latency concerns. But guaranteeing always-on functionality comes with two major challenges: heavily resource-constrained hardware; and dynamic application conditions. Probabilistic models present an ideal solution to these challenges: they are robust to missing data, allow for joint predictions and have small data needs. In addition, ongoing efforts in field of tractable learning have resulted in probabilistic models with strict inference efficiency guarantees. However, the current notions of tractability are often limited to model complexity, disregarding the hardware's specifications and constraints. We propose a novel resource-aware cost metric that takes into consideration the hardware's properties in determining whether the inference task can be efficiently deployed. We use this metric to evaluate the performance versus resource trade-off relevant to the application of interest, and we propose a strategy that selects the device-settings that can optimally meet users' requirements. We showcase our framework on a mobile activity recognition scenario, and on a variety of benchmark datasets representative of the field of tractable learning and of the applications of interest.

Stochastic gradient descent with a large initial learning rate is widely used for training modern neural net architectures. Although a small initial learning rate allows for faster training and better test performance initially, the large learning rate achieves better generalization soon after the learning rate is annealed. Towards explaining this phenomenon, we devise a setting in which we can prove that a two layer network trained with large initial learning rate and annealing provably generalizes better than the same network trained with a small learning rate from the start. The key insight in our analysis is that the order of learning different types of patterns is crucial: because the small learning rate model first memorizes low-noise, hard-to-fit patterns, it generalizes worse on hard-to-generalize, easier-to-fit patterns than its large learning rate counterpart. This concept translates to a larger-scale setting: we demonstrate that one can add a small patch to CIFAR-10 images that is immediately memorizable by a model with small initial learning rate, but ignored by the model with large learning rate until after annealing. Our experiments show that this causes the small learning rate model's accuracy on unmodified images to suffer, as it relies too much on the patch early on.

Interpretability has become an important topic of research as more machine learning (ML) models are deployed and widely used to make important decisions. Most of the current explanation methods provide explanations through feature importance scores, which identify features that are important for each individual input. However, how to systematically summarize and interpret such per sample feature importance scores itself is challenging. In this work, we propose principles and desiderata for \emph{concept} based explanation, which goes beyond per-sample features to identify higher level human-understandable concepts that apply across the entire dataset. We develop a new algorithm, ACE, to automatically extract visual concepts. Our systematic experiments demonstrate that \alg discovers concepts that are human-meaningful, coherent and important for the neural network's predictions.

In this work we study loss functions for learning and evaluating probability distributions over large discrete domains. Unlike classification or regression where a wide variety of loss functions are used, in the distribution learning and density estimation literature, very few losses outside the dominant \emph{log loss} are applied. We aim to understand this fact, taking an axiomatic approach to the design of loss functions for distributions. We start by proposing a set of desirable criteria that any good loss function should satisfy. Intuitively, these criteria require that the loss function faithfully evaluates a candidate distribution, both in expectation and when estimated on a few samples. Interestingly, we observe that \emph{no loss function} possesses all of these criteria. However, one can circumvent this issue by introducing a natural restriction on the set of candidate distributions. Specifically, we require that candidates are \emph{calibrated} with respect to the target distribution, i.e., they may contain less information than the target but otherwise do not significantly distort the truth. We show that, after restricting to this set of distributions, the log loss and a large variety of other losses satisfy the desired criteria. These results pave the way for future investigations of distribution learning that look beyond the log loss, choosing a loss function based on application or domain need.