NeurIPS 2019 论文列表

Numerous neurophysiological studies have revealed that a large number of the primary visual cortex neurons operate in a regime called surround modulation. Surround modulation has a substantial effect on various perceptual tasks, and it also plays a crucial role in the efficient neural coding of the visual cortex. Inspired by the notion of surround modulation, we designed new excitatory-inhibitory connections between a unit and its surrounding units in the convolutional neural network (CNN) to achieve a more biologically plausible network. Our experiments show that this simple mechanism can considerably improve both the performance and training speed of traditional CNNs in visual tasks. We further explore additional outcomes of the proposed structure. We first evaluate the model under several visual challenges, such as the presence of clutter or change in lighting conditions and show its superior generalization capability in handling these challenging situations. We then study possible changes in the statistics of neural activities such as sparsity and decorrelation and provide further insight into the underlying efficiencies of surround modulation. Experimental results show that importing surround modulation into the convolutional layers ensues various effects analogous to those derived by surround modulation in the visual cortex.

We investigate batch policy optimization for cost-sensitive classification and contextual bandits---two related tasks that obviate exploration but require generalizing from observed rewards to action selections in unseen contexts. When rewards are fully observed, we show that the expected reward objective exhibits suboptimal plateaus and exponentially many local optima in the worst case. To overcome the poor landscape, we develop a convex surrogate that is calibrated with respect to entropy regularized expected reward. We then consider the partially observed case, where rewards are recorded for only a subset of actions. Here we generalize the surrogate to partially observed data, and uncover novel objectives for batch contextual bandit training. We find that surrogate objectives remain provably sound in this setting and empirically demonstrate state-of-the-art performance.

One-bit compressed sensing (1bCS) is a method of signal acquisition under extreme measurement quantization that gives important insights on the limits of signal compression and analog-to-digital conversion. The setting is also equivalent to the problem of learning a sparse hyperplane-classifier. In this paper, we propose a generic approach for signal recovery in nonadaptive 1bCS that leads to improved sample complexity for approximate recovery for a variety of signal models, including nonnegative signals and binary signals. We construct 1bCS matrices that are universal - i.e. work for all signals under a model - and at the same time recover very general random sparse signals with high probability. In our approach, we divide the set of samples (measurements) into two parts, and use the first part to recover the superset of the support of a sparse vector. The second set of measurements is then used to approximate the signal within the superset. While support recovery in 1bCS is well-studied, recovery of superset of the support requires fewer samples, which then leads to an overall reduction in sample complexity for approximate recovery.

We present a method for storing multiple models within a single set of parameters. Models can coexist in superposition and still be retrieved individually. In experiments with neural networks, we show that a surprisingly large number of models can be effectively stored within a single parameter instance. Furthermore, each of these models can undergo thousands of training steps without significantly interfering with other models within the superposition. This approach may be viewed as the online complement of compression: rather than reducing the size of a network after training, we make use of the unrealized capacity of a network during training.

Posterior sampling for reinforcement learning (PSRL) is an effective method for balancing exploration and exploitation in reinforcement learning. Randomised value functions (RVF) can be viewed as a promising approach to scaling PSRL. However, we show that most contemporary algorithms combining RVF with neural network function approximation do not possess the properties which make PSRL effective, and provably fail in sparse reward problems. Moreover, we find that propagation of uncertainty, a property of PSRL previously thought important for exploration, does not preclude this failure. We use these insights to design Successor Uncertainties (SU), a cheap and easy to implement RVF algorithm that retains key properties of PSRL. SU is highly effective on hard tabular exploration benchmarks. Furthermore, on the Atari 2600 domain, it surpasses human performance on 38 of 49 games tested (achieving a median human normalised score of 2.09), and outperforms its closest RVF competitor, Bootstrapped DQN, on 36 of those.

Computing optimal transport (OT) between measures in high dimensions is doomed by the curse of dimensionality. A popular approach to avoid this curse is to project input measures on lower-dimensional subspaces (1D lines in the case of sliced Wasserstein distances), solve the OT problem between these reduced measures, and settle for the Wasserstein distance between these reductions, rather than that between the original measures. This approach is however difficult to extend to the case in which one wants to compute an OT map (a Monge map) between the original measures. Since computations are carried out on lower-dimensional projections, classical map estimation techniques can only produce maps operating in these reduced dimensions. We propose in this work two methods to extrapolate, from an transport map that is optimal on a subspace, one that is nearly optimal in the entire space. We prove that the best optimal transport plan that takes such "subspace detours" is a generalization of the Knothe-Rosenblatt transport. We show that these plans can be explicitly formulated when comparing Gaussian measures (between which the Wasserstein distance is commonly referred to as the Bures or Fréchet distance). We provide an algorithm to select optimal subspaces given pairs of Gaussian measures, and study scenarios in which that mediating subspace can be selected using prior information. We consider applications to semantic mediation between elliptic word embeddings and domain adaptation with Gaussian mixture models.

Unlike the white-box counterparts that are widely studied and readily accessible, adversarial examples in black-box settings are generally more Herculean on account of the difficulty of estimating gradients. Many methods achieve the task by issuing numerous queries to target classification systems, which makes the whole procedure costly and suspicious to the systems. In this paper, we aim at reducing the query complexity of black-box attacks in this category. We propose to exploit gradients of a few reference models which arguably span some promising search subspaces. Experimental results show that, in comparison with the state-of-the-arts, our method can gain up to 2x and 4x reductions in the requisite mean and medium numbers of queries with much lower failure rates even if the reference models are trained on a small and inadequate dataset disjoint to the one for training the victim model. Code and models for reproducing our results will be made publicly available.

We consider the widely-used average-linkage, single-linkage, and Ward's methods for computing hierarchical clusterings of high-dimensional Euclidean inputs. It is easy to show that there is no efficient implementation of these algorithms in high dimensional Euclidean space since it implicitly requires to solve the closest pair problem, a notoriously difficult problem. However, how fast can these algorithms be implemented if we allow approximation? More precisely: these algorithms successively merge the clusters that are at closest average (for average-linkage), minimum distance (for single-linkage), or inducing the least sum-of-square error (for Ward's). We ask whether one could obtain a significant running-time improvement if the algorithm can merge $\gamma$-approximate closest clusters (namely, clusters that are at distance (average, minimum, or sum-of-square error) at most $\gamma$ times the distance of the closest clusters). We show that one can indeed take advantage of the relaxation and compute the approximate hierarchical clustering tree using $\widetilde{O}(n)$ $\gamma$-approximate nearest neighbor queries. This leads to an algorithm running in time $\widetilde{O}(nd) + n^{1+O(1/\gamma)}$ for $d$-dimensional Euclidean space. We then provide experiments showing that these algorithms perform as well as the non-approximate version for classic classification tasks while achieving a significant speed-up.

This paper considers submodular function minimization with \textit{noisy evaluation oracles} that return the function value of a submodular objective with zero-mean additive noise. For this problem, we provide an algorithm that returns an $O(n^{3/2}/\sqrt{T})$-additive approximate solution in expectation, where $n$ and $T$ stand for the size of the problem and the number of oracle calls, respectively. There is no room for reducing this error bound by a factor smaller than $O(1/\sqrt{n})$. Indeed, we show that any algorithm will suffer additive errors of $\Omega(n/\sqrt{T})$ in the worst case. Further, we consider an extended problem setting with \textit{multiple-point feedback} in which we can get the feedback of $k$ function values with each oracle call. Under the additional assumption that each noisy oracle is submodular and that $2 \leq k = O(1)$, we provide an algorithm with an $O(n/\sqrt{T})$-additive error bound as well as a worst-case analysis including a lower bound of $\Omega(n/\sqrt{T})$, which together imply that the algorithm achieves an optimal error bound up to a constant.

Synaptic plasticity is widely accepted to be the mechanism behind learning in the brain’s neural networks. A central question is how synapses, with access to only local information about the network, can still organize collectively and perform circuit-wide learning in an efficient manner. In single-layered and all-to-all connected neural networks, local plasticity has been shown to implement gradient-based learning on a class of cost functions that contain a term that aligns the similarity of outputs to the similarity of inputs. Whether such cost functions exist for networks with other architectures is not known. In this paper, we introduce structured and deep similarity matching cost functions, and show how they can be optimized in a gradient-based manner by neural networks with local learning rules. These networks extend F\"oldiak’s Hebbian/Anti-Hebbian network to deep architectures and structured feedforward, lateral and feedback connections. Credit assignment problem is solved elegantly by a factorization of the dual learning objective to synapse specific local objectives. Simulations show that our networks learn meaningful features.

We propose a scalable framework for inference in a continuous sigmoidal Cox process that assumes the corresponding intensity function is given by a Gaussian process (GP) prior transformed with a scaled logistic sigmoid function. We present a tractable representation of the likelihood through augmentation with a superposition of Poisson processes. This view enables a structured variational approximation capturing dependencies across variables in the model. Our framework avoids discretization of the domain, does not require accurate numerical integration over the input space and is not limited to GPs with squared exponential kernels. We evaluate our approach on synthetic and real-world data showing that its benefits are particularly pronounced on multivariate input settings where it overcomes the limitations of mean-field methods and sampling schemes. We provide the state of-the-art in terms of speed, accuracy and uncertainty quantification trade-offs.

We propose in this paper a general framework for deriving loss functions for structured prediction. In our framework, the user chooses a convex set including the output space and provides an oracle for projecting onto that set. Given that oracle, our framework automatically generates a corresponding convex and smooth loss function. As we show, adding a projection as output layer provably makes the loss smaller. We identify the marginal polytope, the output space's convex hull, as the best convex set on which to project. However, because the projection onto the marginal polytope can sometimes be expensive to compute, we allow to use any convex superset instead, with potentially cheaper-to-compute projection. Since efficient projection algorithms are available for numerous convex sets, this allows us to construct loss functions for a variety of tasks. On the theoretical side, when combined with calibrated decoding, we prove that our loss functions can be used as a consistent surrogate for a (potentially non-convex) target loss function of interest. We demonstrate our losses on label ranking, ordinal regression and multilabel classification, confirming the improved accuracy enabled by projections.

Learning a graph with a specific structure is essential for interpretability and identification of the relationships among data. But structured graph learning from observed samples is an NP-hard combinatorial problem. In this paper, we first show, for a set of important graph families it is possible to convert the combinatorial constraints of structure into eigenvalue constraints of the graph Laplacian matrix. Then we introduce a unified graph learning framework lying at the integration of the spectral properties of the Laplacian matrix with Gaussian graphical modeling, which is capable of learning structures of a large class of graph families. The proposed algorithms are provably convergent and practically amenable for big-data specific tasks. Extensive numerical experiments with both synthetic and real datasets demonstrate the effectiveness of the proposed methods. An R package containing codes for all the experimental results is submitted as a supplementary file.

Graphical models encode the stochastic dependencies among random variables (RVs). The vertices represent the RVs, and the edges signify the conditional dependencies among the RVs. Structure learning is the process of inferring the edges by observing realizations of the RVs, and it has applications in a wide range of technological, social, and biological networks. Learning the structure of graphs when the vertices are treated in isolation from inferential information known about them is well-investigated. In a wide range of domains, however, often there exist additional inferred knowledge about the structure, which can serve as valuable side information. For instance, the gene networks that represent different subtypes of the same cancer share similar edges across all subtypes and also have exclusive edges corresponding to each subtype, rendering partially similar graphical models for gene expression in different cancer subtypes. Hence, an inferential decision regarding a gene network can serve as side information for inferring other related gene networks. When such side information is leveraged judiciously, it can translate to significant improvement in structure learning. Leveraging such side information can be abstracted as inferring structures of distinct graphical models that are {\sl partially} similar. This paper focuses on Ising graphical models, and considers the problem of simultaneously learning the structures of two {\sl partially} similar graphs, where any inference about the structure of one graph offers side information for the other graph. The bounded edge subclass of Ising models is considered, and necessary conditions (information-theoretic ), as well as sufficient conditions (algorithmic) for the sample complexity for achieving a bounded probability of error, are established. Furthermore, specific regimes are identified in which the necessary and sufficient conditions coincide, rendering the optimal sample complexity.

A key challenge in crowdsourcing is inferring the ground truth from noisy and unreliable data. To do so, existing approaches rely on collecting redundant information from the crowd, and aggregating it with some probabilistic method. However, oftentimes such methods are computationally inefficient, are restricted to some specific settings, or lack theoretical guarantees. In this paper, we revisit the problem of binary classification from crowdsourced data. Specifically we propose Streaming Bayesian Inference for Crowdsourcing (SBIC), a new algorithm that does not suffer from any of these limitations. First, SBIC has low complexity and can be used in a real-time online setting. Second, SBIC has the same accuracy as the best state-of-the-art algorithms in all settings. Third, SBIC has provable asymptotic guarantees both in the online and offline settings.

We consider a generic empirical composition optimization problem, where there are empirical averages present both outside and inside nonlinear loss functions. Such a problem is of interest in various machine learning applications, and cannot be directly solved by standard methods such as stochastic gradient descent (SGD). We take a novel approach to solving this problem by reformulating the original minimization objective into an equivalent min-max objective, which brings out all the empirical averages that are originally inside the nonlinear loss functions. We exploit the rich structures of the reformulated problem and develop a stochastic primal-dual algorithms, SVRPDA-I, to solve the problem efficiently. We carry out extensive theoretical analysis of the proposed algorithm, obtaining the convergence rate, the total computation complexity and the storage complexity. In particular, the algorithm is shown to converge at a linear rate when the problem is strongly convex. Moreover, we also develop an approximate version of the algorithm, named SVRPDA-II, which further reduces the memory requirement. Finally, we evaluate the performance of our algorithms on several real-world benchmarks and experimental results show that they significantly outperform existing techniques.

In deep neural nets, lower level embedding layers account for a large portion of the total number of parameters. Tikhonov regularization, graph-based regularization, and hard parameter sharing are approaches that introduce explicit biases into training in a hope to reduce statistical complexity. Alternatively, we propose stochastically shared embeddings (SSE), a data-driven approach to regularizing embedding layers, which stochastically transitions between embeddings during stochastic gradient descent (SGD). Because SSE integrates seamlessly with existing SGD algorithms, it can be used with only minor modifications when training large scale neural networks. We develop two versions of SSE: SSE-Graph using knowledge graphs of embeddings; SSE-SE using no prior information. We provide theoretical guarantees for our method and show its empirical effectiveness on 6 distinct tasks, from simple neural networks with one hidden layer in recommender systems, to the transformer and BERT in natural languages. We find that when used along with widely-used regularization methods such as weight decay and dropout, our proposed SSE can further reduce overfitting, which often leads to more favorable generalization results.

Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) algorithms have received increasing attention in both theory and practice. In this paper, we propose a Stochastic Recursive Variance-Reduced gradient HMC (SRVR-HMC) algorithm. It makes use of a semi-stochastic gradient estimator that recursively accumulates the gradient information to reduce the variance of the stochastic gradient. We provide a convergence analysis of SRVR-HMC for sampling from a class of non-log-concave distributions and show that SRVR-HMC converges faster than all existing HMC-type algorithms based on underdamped Langevin dynamics. Thorough experiments on synthetic and real-world datasets validate our theory and demonstrate the superiority of SRVR-HMC.

A broad class of convex optimization problems can be formulated as a semidefinite program (SDP), minimization of a convex function over the positive-semidefinite cone subject to some affine constraints. The majority of classical SDP solvers are designed for the deterministic setting where problem data is readily available. In this setting, generalized conditional gradient methods (aka Frank-Wolfe-type methods) provide scalable solutions by leveraging the so-called linear minimization oracle instead of the projection onto the semidefinite cone. Most problems in machine learning and modern engineering applications, however, contain some degree of stochasticity. In this work, we propose the first conditional-gradient-type method for solving stochastic optimization problems under affine constraints. Our method guarantees O(k^{-1/3}) convergence rate in expectation on the objective residual and O(k^{-5/12}) on the feasibility gap.