NeurIPS 2019 Papers

A new approach to explainable AI, denoted {\it deliberative explanations,\/} is proposed. Deliberative explanations are a visualization technique that aims to go beyond the simple visualization of the image regions (or, more generally, input variables) responsible for a network prediction. Instead, they aim to expose the deliberations carried by the network to arrive at that prediction, by uncovering the insecurities of the network about the latter. The explanation consists of a list of insecurities, each composed of 1) an image region (more generally, a set of input variables), and 2) an ambiguity formed by the pair of classes responsible for the network uncertainty about the region. Since insecurity detection requires quantifying the difficulty of network predictions, deliberative explanations combine ideas from the literatures on visual explanations and assessment of classification difficulty. More specifically, the proposed implementation combines attributions with respect to both class predictions and a difficulty score. An evaluation protocol that leverages object recognition (CUB200) and scene classification (ADE20K) datasets that combine part and attribute annotations is also introduced to evaluate the accuracy of deliberative explanations. Finally, an experimental evaluation shows that the most accurate explanations are achieved by combining non self-referential difficulty scores and second-order attributions. The resulting insecurities are shown to correlate with regions of attributes that are shared by different classes. Since these regions are also ambiguous for humans, deliberative explanations are intuitive, suggesting that the deliberative process of modern networks correlates with human reasoning.

We introduce a feature scattering-based adversarial training approach for improving model robustness against adversarial attacks. Conventional adversarial training approaches leverage a supervised scheme (either targeted or non-targeted) in generating attacks for training, which typically suffer from issues such as label leaking as noted in recent works. Differently, the proposed approach generates adversarial images for training through feature scattering in the latent space, which is unsupervised in nature and avoids label leaking. More importantly, this new approach generates perturbed images in a collaborative fashion, taking the inter-sample relationships into consideration. We conduct analysis on model robustness and demonstrate the effectiveness of the proposed approach through extensively experiments on different datasets compared with state-of-the-art approaches.

Neural backdoor attack is emerging as a severe security threat to deep learning, while the capability of existing defense methods is limited, especially for complex backdoor triggers. In the work, we explore the space formed by the pixel values of all possible backdoor triggers. An original trigger used by an attacker to build the backdoored model represents only a point in the space. It then will be generalized into a distribution of valid triggers, all of which can influence the backdoored model. Thus, previous methods that model only one point of the trigger distribution is not sufficient. Getting the entire trigger distribution, e.g., via generative modeling, is a key of effective defense. However, existing generative modeling techniques for image generation are not applicable to the backdoor scenario as the trigger distribution is completely unknown. In this work, we propose max-entropy staircase approximator (MESA) for high-dimensional sampling-free generative modeling and use it to recover the trigger distribution. We also develop a defense technique to remove the triggers from the backdoored model. Our experiments on Cifar10/100 dataset demonstrate the effectiveness of MESA in modeling the trigger distribution and the robustness of the proposed defense method.

Recent progress in natural language generation has raised dual-use concerns. While applications like summarization and translation are positive, the underlying technology also might enable adversaries to generate neural fake news: targeted propaganda that closely mimics the style of real news.

We propose a recurrent neural-network for real-time reconstruction of acoustic camera spherical maps. The network, dubbed DeepWave, is both physically and algorithmically motivated: its recurrent architecture mimics iterative solvers from convex optimisation, and its parsimonious parametrisation is based on the natural structure of acoustic imaging problems. Each network layer applies successive filtering, biasing and activation steps to its input, which can be interpreted as generalised deblurring and sparsification steps. To comply with the irregular geometry of spherical maps, filtering operations are implemented efficiently by means of graph signal processing techniques. Unlike commonly-used imaging network architectures, DeepWave is moreover capable of directly processing the complex-valued raw microphone correlations, learning how to optimally back-project these into a spherical map. We propose moreover a smart physically-inspired initialisation scheme that attains much faster training and higher performance than random initialisation. Our real-data experiments show DeepWave has similar computational speed to the state-of-the-art delay-and-sum imager with vastly superior resolution. While developed primarily for acoustic cameras, DeepWave could easily be adapted to neighbouring signal processing fields, such as radio astronomy, radar and sonar.

Deep neural network (DNN) based salient object detection in images based on high-quality labels is expensive. Alternative unsupervised approaches rely on careful selection of multiple handcrafted saliency methods to generate noisy pseudo-ground-truth labels. In this work, we propose a two-stage mechanism for robust unsupervised object saliency prediction, where the first stage involves refinement of the noisy pseudo labels generated from different handcrafted methods. Each handcrafted method is substituted by a deep network that learns to generate the pseudo labels. These labels are refined incrementally in multiple iterations via our proposed self-supervision technique. In the second stage, the refined labels produced from multiple networks representing multiple saliency methods are used to train the actual saliency detection network. We show that this self-learning procedure outperforms all the existing unsupervised methods over different datasets. Results are even comparable to those of fully-supervised state-of-the-art approaches.

Many measurement modalities arise from well-understood physical processes and result in information-rich but difficult-to-interpret data. Much of this data still requires laborious human interpretation. This is the case in nuclear magnetic resonance (NMR) spectroscopy, where the observed spectrum of a molecule provides a distinguishing fingerprint of its bond structure. Here we solve the resulting inverse problem: given a molecular formula and a spectrum, can we infer the chemical structure? We show for a wide variety of molecules we can quickly compute the correct molecular structure, and can detect with reasonable certainty when our method cannot. We treat this as a problem of graph-structured prediction, where armed with per-vertex information on a subset of the vertices, we infer the edges and edge types. We frame the problem as a Markov decision process (MDP) and incrementally construct molecules one bond at a time, training a deep neural network via imitation learning, where we learn to imitate a subisomorphic oracle which knows which remaining bonds are correct. Our method is fast, accurate, and is the first among recent chemical-graph generation approaches to exploit per-vertex information and generate graphs with vertex constraints. Our method points the way towards automation of molecular structure identification and potentially active learning for spectroscopy.

We address the problem of finding representative points of datasets by learning from multiple datasets and their ground-truth summaries. We develop a supervised subset selection framework, based on the facility location utility function, which learns to map datasets to their ground-truth representatives. To do so, we propose to learn representations of data so that the input of transformed data to the facility location recovers their ground-truth representatives. Given the NP-hardness of the utility function, we consider its convex relaxation based on sparse representation and investigate conditions under which the solution of the convex optimization recovers ground-truth representatives of each dataset. We design a loss function whose minimization over the parameters of the data representation network leads to satisfying the theoretical conditions, hence guaranteeing recovering ground-truth summaries. Given the non-convexity of the loss function, we develop an efficient learning scheme that alternates between representation learning by minimizing our proposed loss given the current assignments of points to ground-truth representatives and updating assignments given the current data representation. By experiments on the problem of learning key-steps (subactivities) of instructional videos, we show that our proposed framework improves the state-of-the-art supervised subset selection algorithms.

Existing deep learning based facial landmark detection methods have achieved excellent performance. These methods, however, do not explicitly embed the structural dependencies among landmark points. They hence cannot preserve the geometric relationships between landmark points or generalize well to challenging conditions or unseen data. This paper proposes a method for deep structured facial landmark detection based on combining a deep Convolutional Network with a Conditional Random Field. We demonstrate its superior performance to existing state-of-the-art techniques in facial landmark detection, especially a better generalization ability on challenging datasets that include large pose and occlusion.

The signature is an infinite graded sequence of statistics known to characterise a stream of data up to a negligible equivalence class. It is a transform which has previously been treated as a fixed feature transformation, on top of which a model may be built. We propose a novel approach which combines the advantages of the signature transform with modern deep learning frameworks. By learning an augmentation of the stream prior to the signature transform, the terms of the signature may be selected in a data-dependent way. More generally, we describe how the signature transform may be used as a layer anywhere within a neural network. In this context it may be interpreted as a pooling operation. We present the results of empirical experiments to back up the theoretical justification. Code available at \texttt{github.com/patrick-kidger/Deep-Signature-Transforms}.

Current approaches for predicting sets from feature vectors ignore the unordered nature of sets and suffer from discontinuity issues as a result. We propose a general model for predicting sets that properly respects the structure of sets and avoids this problem. With a single feature vector as input, we show that our model is able to auto-encode point sets, predict the set of bounding boxes of objects in an image, and predict the set of attributes of these objects.

We introduce deep scale-spaces, a generalization of convolutional neural networks, exploiting the scale symmetry structure of conventional image recognition tasks. Put plainly, the class of an image is invariant to the scale at which it is viewed. We construct scale equivariant cross-correlations based on a principled extension of convolutions, grounded in the theory of scale-spaces and semigroups. As a very basic operation, these cross-correlations can be used in almost any modern deep learning architecture in a plug-and-play manner. We demonstrate our networks on the Patch Camelyon and Cityscapes datasets, to prove their utility and perform introspective studies to further understand their properties.

The success of deep networks has been attributed in part to their expressivity: per parameter, deep networks can approximate a richer class of functions than shallow networks. In ReLU networks, the number of activation patterns is one measure of expressivity; and the maximum number of patterns grows exponentially with the depth. However, recent work has showed that the practical expressivity of deep networks - the functions they can learn rather than express - is often far from the theoretical maximum. In this paper, we show that the average number of activation patterns for ReLU networks at initialization is bounded by the total number of neurons raised to the input dimension. We show empirically that this bound, which is independent of the depth, is tight both at initialization and during training, even on memorization tasks that should maximize the number of activation patterns. Our work suggests that realizing the full expressivity of deep networks may not be possible in practice, at least with current methods.

Gaussian processes are the leading class of distributions on random functions, but they suffer from well known issues including difficulty scaling and inflexibility with respect to certain shape constraints (such as nonnegativity). Here we propose Deep Random Splines, a flexible class of random functions obtained by transforming Gaussian noise through a deep neural network whose output are the parameters of a spline. Unlike Gaussian processes, Deep Random Splines allow us to readily enforce shape constraints while inheriting the richness and tractability of deep generative models. We also present an observational model for point process data which uses Deep Random Splines to model the intensity function of each point process and apply it to neural population data to obtain a low-dimensional representation of spiking activity. Inference is performed via a variational autoencoder that uses a novel recurrent encoder architecture that can handle multiple point processes as input. We use a newly collected dataset where a primate completes a pedaling task, and observe better dimensionality reduction with our model than with competing alternatives.

In this paper, we propose our Correlation For Completion Network (CFCNet), an end-to-end deep learning model that uses the correlation between two data sources to perform sparse depth completion. CFCNet learns to capture, to the largest extent, the semantically correlated features between RGB and depth information. Through pairs of image pixels and the visible measurements in a sparse depth map, CFCNet facilitates feature-level mutual transformation of different data sources. Such a transformation enables CFCNet to predict features and reconstruct data of missing depth measurements according to their corresponding, transformed RGB features. We extend canonical correlation analysis to a 2D domain and formulate it as one of our training objectives (i.e. 2d deep canonical correlation, or “2D^2CCA loss"). Extensive experiments validate the ability and flexibility of our CFCNet compared to the state-of-the-art methods on both indoor and outdoor scenes with different real-life sparse patterns. Codes are available at: https://github.com/choyingw/CFCNet.

Tensor-based multimodal fusion techniques have exhibited great predictive performance. However, one limitation is that existing approaches only consider bilinear or trilinear pooling, which fails to unleash the complete expressive power of multilinear fusion with restricted orders of interactions. More importantly, simply fusing features all at once ignores the complex local intercorrelations, leading to the deterioration of prediction. In this work, we first propose a polynomial tensor pooling (PTP) block for integrating multimodal features by considering high-order moments, followed by a tensorized fully connected layer. Treating PTP as a building block, we further establish a hierarchical polynomial fusion network (HPFN) to recursively transmit local correlations into global ones. By stacking multiple PTPs, the expressivity capacity of HPFN enjoys an exponential growth w.r.t. the number of layers, which is shown by the equivalence to a very deep convolutional arithmetic circuits. Various experiments demonstrate that it can achieve the state-of-the-art performance.

We present a new deep multi-state Dynamic Recurrent Neural Network (DRNN) architecture for Brain Machine Interface (BMI) applications. Our DRNN is used to predict Cartesian representation of a computer cursor movement kinematics from open-loop neural data recorded from the posterior parietal cortex (PPC) of a human subject in a BMI system. We design the algorithm to achieve a reasonable trade-off between performance and robustness, and we constrain memory usage in favor of future hardware implementation. We feed the predictions of the network back to the input to improve prediction performance and robustness. We apply a scheduled sampling approach to the model in order to solve a statistical distribution mismatch between the ground truth and predictions. Additionally, we configure a small DRNN to operate with a short history of input, reducing the required buffering of input data and number of memory accesses. This configuration lowers the expected power consumption in a neural network accelerator. Operating on wavelet-based neural features, we show that the average performance of DRNN surpasses other state-of-the-art methods in the literature on both single- and multi-day data recorded over 43 days. Results show that multi-state DRNN has the potential to model the nonlinear relationships between the neural data and kinematics for robust BMIs.

Exploring the transferability between heterogeneous tasks sheds light on their intrinsic interconnections, and consequently enables knowledge transfer from one task to another so as to reduce the training effort of the latter. In this paper, we propose an embarrassingly simple yet very efficacious approach to estimating the transferability of deep networks, especially those handling vision tasks. Unlike the seminal work of \emph{taskonomy} that relies on a large number of annotations as supervision and is thus computationally cumbersome, the proposed approach requires no human annotations and imposes no constraints on the architectures of the networks. This is achieved, specifically, via projecting deep networks into a \emph{model space}, wherein each network is treated as a point and the distances between two points are measured by deviations of their produced attribution maps. The proposed approach is several-magnitude times faster than taskonomy, and meanwhile preserves a task-wise topological structure highly similar to the one obtained by taskonomy. Code is available at \url{https://github.com/zju-vipa/TransferbilityFromAttributionMaps}.

Current algorithms for deep learning probably cannot run in the brain because they rely on weight transport, where forward-path neurons transmit their synaptic weights to a feedback path, in a way that is likely impossible biologically. An algorithm called feedback alignment achieves deep learning without weight transport by using random feedback weights, but it performs poorly on hard visual-recognition tasks. Here we describe two mechanisms — a neural circuit called a weight mirror and a modification of an algorithm proposed by Kolen and Pollack in 1994 — both of which let the feedback path learn appropriate synaptic weights quickly and accurately even in large networks, without weight transport or complex wiring. Tested on the ImageNet visual-recognition task, these mechanisms outperform both feedback alignment and the newer sign-symmetry method, and nearly match backprop, the standard algorithm of deep learning, which uses weight transport.