NeurIPS 2021 Papers

To tackle interpretability in deep learning, we present a novel framework to jointly learn a predictive model and its associated interpretation model. The interpreter provides both local and global interpretability about the predictive model in terms of human-understandable high level attribute functions, with minimal loss of accuracy. This is achieved by a dedicated architecture and well chosen regularization penalties. We seek for a small-size dictionary of high level attribute functions that take as inputs the outputs of selected hidden layers and whose outputs feed a linear classifier. We impose strong conciseness on the activation of attributes with an entropy-based criterion while enforcing fidelity to both inputs and outputs of the predictive model. A detailed pipeline to visualize the learnt features is also developed. Moreover, besides generating interpretable models by design, our approach can be specialized to provide post-hoc interpretations for a pre-trained neural network. We validate our approach against several state-of-the-art methods on multiple datasets and show its efficacy on both kinds of tasks.

During online decision making in Multi-Armed Bandits (MAB), one needs to conduct inference on the true mean reward of each arm based on data collected so far at each step. However, since the arms are adaptively selected--thereby yielding non-iid data--conducting inference accurately is not straightforward. In particular, sample averaging, which is used in the family of UCB and Thompson sampling (TS) algorithms, does not provide a good choice as it suffers from bias and a lack of good statistical properties (e.g. asymptotic normality). Our thesis in this paper is that more sophisticated inference schemes that take into account the adaptive nature of the sequentially collected data can unlock further performance gains, even though both UCB and TS type algorithms are optimal in the worst case. In particular, we propose a variant of TS-style algorithms--which we call doubly adaptive TS--that leverages recent advances in causal inference and adaptively reweights the terms of a doubly robust estimator on the true mean reward of each arm. Through 20 synthetic domain experiments and a semi-synthetic experiment based on data from an A/B test of a web service, we demonstrate that using an adaptive inferential scheme (while still retaining the exploration efficacy of TS) provides clear benefits in online decision making: the proposed DATS algorithm has superior empirical performance to existing baselines (UCB and TS) in terms of regret and sample complexity in identifying the best arm. In addition, we also provide a finite-time regret bound of doubly adaptive TS that matches (up to log factors) those of UCB and TS algorithms, thereby establishing that its improved practical benefits do not come at the expense of worst-case suboptimality.

As large-scale language model pretraining pushes the state-of-the-art in text generation, recent work has turned to controlling attributes of the text such models generate. While modifying the pretrained models via fine-tuning remains the popular approach, it incurs a significant computational cost and can be infeasible due to a lack of appropriate data. As an alternative, we propose \textsc{MuCoCO}---a flexible and modular algorithm for controllable inference from pretrained models. We formulate the decoding process as an optimization problem that allows for multiple attributes we aim to control to be easily incorporated as differentiable constraints. By relaxing this discrete optimization to a continuous one, we make use of Lagrangian multipliers and gradient-descent-based techniques to generate the desired text. We evaluate our approach on controllable machine translation and style transfer with multiple sentence-level attributes and observe significant improvements over baselines.

Discovering ideal Graph Neural Networks (GNNs) architectures for different tasks is labor intensive and time consuming. To save human efforts, Neural Architecture Search (NAS) recently has been used to automatically discover adequate GNN architectures for certain tasks in order to achieve competitive or even better performance compared with manually designed architectures. However, existing works utilizing NAS to search GNN structures fail to answer the question: how NAS is able to select the desired GNN architectures? In this paper, we investigate this question to solve the problem, for the first time. We conduct a measurement study with experiments to discover that gradient based NAS methods tend to select proper architectures based on the usefulness of different types of information with respect to the target task. Our explorations further show that gradient based NAS also suffers from noises hidden in the graph, resulting in searching suboptimal GNN architectures. Based on our findings, we propose a Graph differentiable Architecture Search model with Structure Optimization (GASSO), which allows differentiable search of the architecture with gradient descent and is able to discover graph neural architectures with better performance through employing graph structure learning as a denoising process in the search procedure. The proposed GASSO model is capable of simultaneously searching the optimal architecture and adaptively adjusting graph structure by jointly optimizing graph architecture search and graph structure denoising. Extensive experiments on real-world graph datasets demonstrate that our proposed GASSO model is able to achieve state-of-the-art performance compared with existing baselines.

Defocus blur is one kind of blur effects often seen in images, which is challenging to remove due to its spatially variant amount. This paper presents an end-to-end deep learning approach for removing defocus blur from a single image, so as to have an all-in-focus image for consequent vision tasks. First, a pixel-wise Gaussian kernel mixture (GKM) model is proposed for representing spatially variant defocus blur kernels in an efficient linear parametric form, with higher accuracy than existing models. Then, a deep neural network called GKMNet is developed by unrolling a fixed-point iteration of the GKM-based deblurring. The GKMNet is built on a lightweight scale-recurrent architecture, with a scale-recurrent attention module for estimating the mixing coefficients in GKM for defocus deblurring. Extensive experiments show that the GKMNet not only noticeably outperforms existing defocus deblurring methods, but also has its advantages in terms of model complexity and computational efficiency.

Attention is sparse in vision transformers. We observe the final prediction in vision transformers is only based on a subset of most informative tokens, which is sufficient for accurate image recognition. Based on this observation, we propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically based on the input. Specifically, we devise a lightweight prediction module to estimate the importance score of each token given the current features. The module is added to different layers to prune redundant tokens hierarchically. To optimize the prediction module in an end-to-end manner, we propose an attention masking strategy to differentiably prune a token by blocking its interactions with other tokens. Benefiting from the nature of self-attention, the unstructured sparse tokens are still hardware friendly, which makes our framework easy to achieve actual speed-up. By hierarchically pruning 66% of the input tokens, our method greatly reduces 31% $\sim$ 37% FLOPs and improves the throughput by over 40% while the drop of accuracy is within 0.5% for various vision transformers. Equipped with the dynamic token sparsification framework, DynamicViT models can achieve very competitive complexity/accuracy trade-offs compared to state-of-the-art CNNs and vision transformers on ImageNet. Code is available at https://github.com/raoyongming/DynamicViT

Shift neural networks reduce computation complexity by removing expensive multiplication operations and quantizing continuous weights into low-bit discrete values, which are fast and energy-efficient compared to conventional neural networks. However, existing shift networks are sensitive to the weight initialization and yield a degraded performance caused by vanishing gradient and weight sign freezing problem. To address these issues, we propose S$^3$ re-parameterization, a novel technique for training low-bit shift networks. Our method decomposes a discrete parameter in a sign-sparse-shift 3-fold manner. This way, it efficiently learns a low-bit network with weight dynamics similar to full-precision networks and insensitive to weight initialization. Our proposed training method pushes the boundaries of shift neural networks and shows 3-bit shift networks compete with their full-precision counterparts in terms of top-1 accuracy on ImageNet.

Several problems in neuroimaging and beyond require inference on the parameters of multi-task sparse hierarchical regression models. Examples include M/EEG inverse problems, neural encoding models for task-based fMRI analyses, and climate science. In these domains, both the model parameters to be inferred and the measurement noise may exhibit a complex spatio-temporal structure. Existing work either neglects the temporal structure or leads to computationally demanding inference schemes. Overcoming these limitations, we devise a novel flexible hierarchical Bayesian framework within which the spatio-temporal dynamics of model parameters and noise are modeled to have Kronecker product covariance structure. Inference in our framework is based on majorization-minimization optimization and has guaranteed convergence properties. Our highly efficient algorithms exploit the intrinsic Riemannian geometry of temporal autocovariance matrices. For stationary dynamics described by Toeplitz matrices, the theory of circulant embeddings is employed. We prove convex bounding properties and derive update rules of the resulting algorithms. On both synthetic and real neural data from M/EEG, we demonstrate that our methods lead to improved performance.

Deep learning (DL) systems have been gaining popularity in critical tasks such as credit evaluation and crime prediction. Such systems demand fairness. Recent work shows that DL software implementations introduce variance: identical DL training runs (i.e., identical network, data, configuration, software, and hardware) with a fixed seed produce different models. Such variance could make DL models and networks violate fairness compliance laws, resulting in negative social impact. In this paper, we conduct the first empirical study to quantify the impact of software implementation on the fairness and its variance of DL systems. Our study of 22 mitigation techniques and five baselines reveals up to 12.6% fairness variance across identical training runs with identical seeds. In addition, most debiasing algorithms have a negative impact on the model such as reducing model accuracy, increasing fairness variance, or increasing accuracy variance. Our literature survey shows that while fairness is gaining popularity in artificial intelligence (AI) related conferences, only 34.4% of the papers use multiple identical training runs to evaluate their approach, raising concerns about their results’ validity. We call for better fairness evaluation and testing protocols to improve fairness and fairness variance of DL systems as well as DL research validity and reproducibility at large.

In recent years there has been a considerable amount of research on local post hoc explanations for neural networks. However, work on building interpretable neural architectures has been relatively sparse. In this paper, we present a novel neural architecture, CoFrNet, inspired by the form of continued fractions which are known to have many attractive properties in number theory, such as fast convergence of approximations to real numbers. We show that CoFrNets can be efficiently trained as well as interpreted leveraging their particular functional form. Moreover, we prove that such architectures are universal approximators based on a proof strategy that is different than the typical strategy used to prove universal approximation results for neural networks based on infinite width (or depth), which is likely to be of independent interest. We experiment on nonlinear synthetic functions and are able to accurately model as well as estimate feature attributions and even higher order terms in some cases, which is a testament to the representational power as well as interpretability of such architectures. To further showcase the power of CoFrNets, we experiment on seven real datasets spanning tabular, text and image modalities, and show that they are either comparable or significantly better than other interpretable models and multilayer perceptrons, sometimes approaching the accuracies of state-of-the-art models.

We develop nested variational inference (NVI), a family of methods that learn proposals for nested importance samplers by minimizing an forward or reverse KL divergence at each level of nesting. NVI is applicable to many commonly-used importance sampling strategies and provides a mechanism for learning intermediate densities, which can serve as heuristics to guide the sampler. Our experiments apply NVI to (a) sample from a multimodal distribution using a learned annealing path (b) learn heuristics that approximate the likelihood of future observations in a hidden Markov model and (c) to perform amortized inference in hierarchical deep generative models. We observe that optimizing nested objectives leads to improved sample quality in terms of log average weight and effective sample size.

Imitation learning often needs a large demonstration set in order to handle the full range of situations that an agent might find itself in during deployment. However, collecting expert demonstrations can be expensive. Recent work in vision, reinforcement learning, and NLP has shown that auxiliary representation learning objectives can reduce the need for large amounts of expensive, task-specific data. Our Empirical Investigation of Representation Learning for Imitation (EIRLI) investigates whether similar benefits apply to imitation learning. We propose a modular framework for constructing representation learning algorithms, then use our framework to evaluate the utility of representation learning for imitation across several environment suites. In the settings we evaluate, we find that existing algorithms for image-based representation learning provide limited value relative to a well-tuned baseline with image augmentations. To explain this result, we investigate differences between imitation learning and other settings where representation learning *has* provided significant benefit, such as image classification. Finally, we release a well-documented codebase which both replicates our findings and provides a modular framework for creating new representation learning algorithms out of reusable components.

A recent line of research investigates how algorithms can be augmented with machine-learned predictions to overcome worst case lower bounds. This area has revealed interesting algorithmic insights into problems, with particular success in the design of competitive online algorithms. However, the question of improving algorithm running times with predictions has largely been unexplored. We take a first step in this direction by combining the idea of machine-learned predictions with the idea of ``warm-starting" primal-dual algorithms. We consider one of the most important primitives in combinatorial optimization: weighted bipartite matching and its generalization to $b$-matching. We identify three key challenges when using learned dual variables in a primal-dual algorithm. First, predicted duals may be infeasible, so we give an algorithm that efficiently maps predicted infeasible duals to nearby feasible solutions. Second, once the duals are feasible, they may not be optimal, so we show that they can be used to quickly find an optimal solution. Finally, such predictions are useful only if they can be learned, so we show that the problem of learning duals for matching has low sample complexity. We validate our theoretical findings through experiments on both real and synthetic data. As a result we give a rigorous, practical, and empirically effective method to compute bipartite matchings.

Controllable text generation concerns two fundamental tasks of wide applications, namely generating text of given attributes (i.e., attribute-conditional generation), and minimally editing existing text to possess desired attributes (i.e., text attribute transfer). Extensive prior work has largely studied the two problems separately, and developed different conditional models which, however, are prone to producing biased text (e.g., various gender stereotypes). This paper proposes to formulate controllable text generation from a principled causal perspective which models the two tasks with a unified framework. A direct advantage of the causal formulation is the use of rich causality tools to mitigate generation biases and improve control. We treat the two tasks as interventional and counterfactual causal inference based on a structural causal model, respectively. We then apply the framework to the challenging practical setting where confounding factors (that induce spurious correlations) are observable only on a small fraction of data. Experiments show significant superiority of the causal approach over previous conditional models for improved control accuracy and reduced bias.

Graph Similarity Computation (GSC) is essential to wide-ranging graph applications such as retrieval, plagiarism/anomaly detection, etc. The exact computation of graph similarity, e.g., Graph Edit Distance (GED), is an NP-hard problem that cannot be exactly solved within an adequate time given large graphs. Thanks to the strong representation power of graph neural network (GNN), a variety of GNN-based inexact methods emerged. To capture the subtle difference across graphs, the key success is designing the dense interaction with features fusion at the early stage, which, however, is a trade-off between speed and accuracy. For slow learning of graph similarity, this paper proposes a novel early-fusion approach by designing a co-attention-based feature fusion network on multilevel GNN features. To further improve the speed without much accuracy drop, we introduce an efficient GSC solution by distilling the knowledge from the slow early-fusion model to the student one for fast inference. Such a student model also enables the offline collection of individual graph embeddings, speeding up the inference time in orders. To address the instability through knowledge transfer, we decompose the dynamic joint embedding into the static pseudo individual ones for precise teacher-student alignment. The experimental analysis on the real-world datasets demonstrates the superiority of our approach over the state-of-the-art methods on both accuracy and efficiency. Particularly, we speed up the prior art by more than 10x on the benchmark AIDS data.

(Non-)robustness of neural networks to small, adversarial pixel-wise perturbations, and as more recently shown, to even random spatial transformations (e.g., translations, rotations) entreats both theoretical and empirical understanding. Spatial robustness to random translations and rotations is commonly attained via equivariant models (e.g., StdCNNs, GCNNs) and training augmentation, whereas adversarial robustness is typically achieved by adversarial training. In this paper, we prove a quantitative trade-off between spatial and adversarial robustness in a simple statistical setting. We complement this empirically by showing that: (a) as the spatial robustness of equivariant models improves by training augmentation with progressively larger transformations, their adversarial robustness worsens progressively, and (b) as the state-of-the-art robust models are adversarially trained with progressively larger pixel-wise perturbations, their spatial robustness drops progressively. Towards achieving Pareto-optimality in this trade-off, we propose a method based on curriculum learning that trains gradually on more difficult perturbations (both spatial and adversarial) to improve spatial and adversarial robustness simultaneously.

Automated hyperparameter optimization (HPO) can support practitioners to obtain peak performance in machine learning models. However, there is often a lack of valuable insights into the effects of different hyperparameters on the final model performance. This lack of explainability makes it difficult to trust and understand the automated HPO process and its results. We suggest using interpretable machine learning (IML) to gain insights from the experimental data obtained during HPO with Bayesian optimization (BO). BO tends to focus on promising regions with potential high-performance configurations and thus induces a sampling bias. Hence, many IML techniques, such as the partial dependence plot (PDP), carry the risk of generating biased interpretations. By leveraging the posterior uncertainty of the BO surrogate model, we introduce a variant of the PDP with estimated confidence bands. We propose to partition the hyperparameter space to obtain more confident and reliable PDPs in relevant sub-regions. In an experimental study, we provide quantitative evidence for the increased quality of the PDPs within sub-regions.

Knowledge distillation has shown great success in classification, however, it is still challenging for detection. In a typical image for detection, representations from different locations may have different contributions to detection targets, making the distillation hard to balance. In this paper, we propose a conditional distillation framework to distill the desired knowledge, namely knowledge that is beneficial in terms of both classification and localization for every instance. The framework introduces a learnable conditional decoding module, which retrieves information given each target instance as query. Specifically, we encode the condition information as query and use the teacher's representations as key. The attention between query and key is used to measure the contribution of different features, guided by a localization-recognition-sensitive auxiliary task. Extensive experiments demonstrate the efficacy of our method: we observe impressive improvements under various settings. Notably, we boost RetinaNet with ResNet-50 backbone from $37.4$ to $40.7$ mAP ($+3.3$) under $1\times$ schedule, that even surpasses the teacher ($40.4$ mAP) with ResNet-101 backbone under $3\times$ schedule. Code has been released on https://github.com/megvii-research/ICD.

Compressive phase retrieval is a popular variant of the standard compressive sensing problem in which the measurements only contain magnitude information. In this paper, motivated by recent advances in deep generative models, we provide recovery guarantees with near-optimal sample complexity for phase retrieval with generative priors. We first show that when using i.i.d. Gaussian measurements and an $L$-Lipschitz continuous generative model with bounded $k$-dimensional inputs, roughly $O(k \log L)$ samples suffice to guarantee that any signal minimizing an amplitude-based empirical loss function is close to the true signal. Attaining this sample complexity with a practical algorithm remains a difficult challenge, and finding a good initialization for gradient-based methods has been observed to pose a major bottleneck. To partially address this, we further show that roughly $O(k \log L)$ samples ensure sufficient closeness between the underlying signal and any {\em globally optimal} solution to an optimization problem designed for spectral initialization (though finding such a solution may still be challenging). We also adapt this result to sparse phase retrieval, and show that $O(s \log n)$ samples are sufficient for a similar guarantee when the underlying signal is $s$-sparse and $n$-dimensional, matching an information-theoretic lower bound. While these guarantees do not directly correspond to a practical algorithm, we propose a practical spectral initialization method motivated by our findings, and experimentally observe performance gains over various existing spectral initialization methods for sparse phase retrieval.