AAAI 2021 论文列表

A polymatrix game is a multi-player game over n players, where each player chooses a pure strategy from a list of its own pure strategies. The utility of each player is a sum of payoffs it gains from the two player's game from all its neighbors, under its chosen strategy and that of its neighbor. As a natural extension to two-player games (a.k.a. bimatrix games), polymatrix games are widely used for multi-agent games in real world scenarios. In this paper we show that the problem of approximating a Nash equilibrium in a polymatrix game within the polynomial precision is PPAD-hard, even in sparse and win-lose ones. This result further challenges the predictability of Nash equilibria as a solution concept in the multi-agent setting. We also propose a simple and efficient algorithm, when the game is further restricted. Together, we establish a new dichotomy theorem for this class of games. It is also of independent interest for exploring the computational and structural properties in Nash equilibria.

Adaptive optimization algorithms such as Adam (Kingma and Ba, 2014) are widely used in deep learning. The stability of such algorithms is often improved with a warmup schedule for the learning rate. Motivated by the difficulty of choosing and tuning warmup schedules, recent work proposes automatic variance rectification of Adam's adaptive learning rate, claiming that this rectified approach ("RAdam") surpasses the vanilla Adam algorithm and reduces the need for expensive tuning of Adam with warmup. In this work, we refute this analysis and provide an alternative explanation for the necessity of warmup based on the magnitude of the update term, which is of greater relevance to training stability. We then provide some "rule-of-thumb" warmup schedules, and we demonstrate that simple untuned warmup of Adam performs more-or-less identically to RAdam in typical practical settings. We conclude by suggesting that practitioners stick to linear warmup with Adam, with a sensible default being linear warmup over 2 / (1 - β₂) training iterations.

Attention with positional encoding has been demonstrated as a powerful component in modern neural network models, such as transformers. However, why positional encoding works well in attention models remains largely unanswered. In this paper, we study the scalar relative positional encoding (SRPE) proposed in the T5 transformer. Such an encoding method has two features. First, it uses a scalar to embed relative positions. Second, the relative positions are bucketized using a fixed heuristic algorithm, and positions in the same bucket share the same embedding. In this work, we show that SRPE in attention has an elegant probabilistic interpretation. More specifically, the positional encoding serves to produce a prior distribution for the attended positions. The resulting attentive distribution can be viewed as a posterior distribution of the attended position given the observed input sequence. Furthermore, we propose a new SRPE (AT5) that adopts a learnable bucketization protocol and automatically adapts to the dependency range specific to the learning task. Empirical studies show that the AT5 achieves superior performance than the T5's SRPE.

The regret bound of dynamic online learning algorithms is often expressed in terms of the variation in the function sequence (V_T) and/or the path-length of the minimizer sequence after T rounds. For strongly convex and smooth functions, Zhang et al. (2017) establish the squared path-length of the minimizer sequence (C*_{2,T}) as a lower bound on regret. They also show that online gradient descent (OGD) achieves this lower bound using multiple gradient queries per round. In this paper, we focus on unconstrained online optimization. We first show that a preconditioned variant of OGD achieves O(min{C*_T,C*_{2,T}}) with one gradient query per round (C*_T refers to the normal path-length). We then propose online optimistic Newton (OON) method for the case when the first and second order information of the function sequence is predictable. The regret bound of OON is captured via the quartic path-length of the minimizer sequence (C*_{4,T}), which can be much smaller than C*_{2,T}. We finally show that by using multiple gradients for OGD, we can achieve an upper bound of O(min{C*_{2,T},V_T}) on regret.

This paper tackles the problem of Lipschitz regularization of Convolutional Neural Networks. Lipschitz regularity is now established as a key property of modern deep learning with implications in training stability, generalization, robustness against adversarial examples, etc. However, computing the exact value of the Lipschitz constant of a neural network is known to be NP-hard. Recent attempts from the literature introduce upper bounds to approximate this constant that are either efficient but loose or accurate but computationally expensive. In this work, by leveraging the theory of Toeplitz matrices, we introduce a new upper bound for convolutional layers that is both tight and easy to compute. Based on this result we devise an algorithm to train Lipschitz regularized Convolutional Neural Networks.

There is a growing concern that the recent progress made in AI, especially regarding the predictive competence of deep learning models, will be undermined by a failure to properly explain their operation and outputs. In response to this disquiet, counterfactual explanations have become very popular in eXplainable AI (XAI) due to their asserted computational, psychological, and legal benefits. In contrast however, semi-factuals (which appear to be equally useful) have surprisingly received no attention. Most counterfactual methods address tabular rather than image data, partly because the non-discrete nature of images makes good counterfactuals difficult to define; indeed, generating plausible counterfactual images which lie on the data manifold is also problematic. This paper advances a novel method for generating plausible counterfactuals and semi-factuals for black-box CNN classifiers doing computer vision. The present method, called PlausIble Exceptionality-based Contrastive Explanations (PIECE), modifies all “exceptional” features in a test image to be “normal” from the perspective of the counterfactual class, to generate plausible counterfactual images. Two controlled experiments compare this method to others in the literature, showing that PIECE generates highly plausible counterfactuals (and the best semi-factuals) on several benchmark measures.

We study the problem of fair and efficient allocation of a set of indivisible goods to agents with additive valuations using the popular fairness notions of envy-freeness up to one good (EF1) and equitability up to one good (EQ1) in conjunction with Pareto-optimality (PO). There exists a pseudo-polynomial time algorithm to compute an EF1+PO allocation, and a non-constructive proof of existence of allocations that are both EF1 and fractionally Pareto-optimal (fPO). We present a pseudo-polynomial time algorithm to compute an EF1+fPO allocation, thereby improving the earlier results. Our techniques also enable us to show that an EQ1+fPO allocation always exists when the values are positive, and that it can be computed in pseudo-polynomial time. We also consider the class of k-ary instances where k is a constant, i.e., each agent has at most k different values for the goods. We show that for such instances an EF1+fPO allocation can be computed in polynomial time. When all values are positive, we show that an EQ1+fPO allocation for such instances can be computed in polynomial time. Next, we consider instances where the number of agents is constant, and show that an EF1+PO (also EQ1+PO) allocation can be computed in polynomial time. These results significantly extend the polynomial-time computability beyond the known cases of binary or identical valuations. Further, we show that the problem of computing an EF1+PO allocation polynomial-time reduces to a problem in the complexity class PLS. We also design a polynomial-time algorithm that computes Nash welfare maximizing allocations when there are constantly many agents with constant many different values for the goods.

We study fair allocation of indivisible goods among additive agents with feasibility constraints. In these settings, every agent is restricted to get a bundle among a specified set of feasible bundles. Such scenarios have been of great interest to the AI community due to their applicability to real-world problems. Following some impossibility results, we restrict attention to matroid feasibility constraints that capture natural scenarios, such as the allocation of shifts to medical doctors, and the allocation of conference papers to referees. We focus on the common fairness notion of envy-freeness up to one good (EF1). Previous algorithms for finding EF1 allocations are either restricted to agents with identical feasibility constraints, or allow free disposal of items. An open problem is the existence of EF1 complete allocations among heterogeneous agents, where the heterogeneity is both in the agents' feasibility constraints and in their valuations. In this work, we make progress on this problem by providing positive and negative results for different matroid and valuation types. Among other results, we devise poly-time algorithms for finding EF1 allocations in the following settings: (i) n agents with heterogeneous partition matroids and heterogeneous binary valuations, (ii) 2 agents with heterogeneous partition matroids and heterogeneous valuations, and (iii) at most 3 agents with heterogeneous binary valuations and identical base-orderable matroids.

In human-aware planning, a planning agent may need to provide an explanation to a human user on why its plan is optimal. A popular approach to do this is called model reconciliation, where the agent tries to reconcile the differences in its model and the human's model such that the plan is also optimal in the human's model. In this paper, we present a logic-based framework for model reconciliation that extends beyond the realm of planning. More specifically, given a knowledge base KB1 entailing a formula phi and a second knowledge base KB2 not entailing it, model reconciliation seeks an explanation, in the form of a cardinality-minimal subset of KB1, whose integration into KB2 makes the entailment possible. Our approach, based on ideas originating in the context of analysis of inconsistencies, exploits the existing hitting set duality between minimal correction sets (MCSes) and minimal unsatisfiable sets (MUSes) in order to identify an appropriate explanation. However, differently from those works targeting inconsistent formulas, which assume a single knowledge base, MCSes and MUSes are computed over two distinct knowledge bases. We conclude our paper with an empirical evaluation of the newly introduced approach on planning instances, where we show how it outperforms an existing state-of-the-art solver, and generic non-planning instances from recent SAT competitions, for which no other solver exists.

Since the recent studies (KDD'20) done by Krichene and Rendle on the sampling based top-k evaluation metric for recommendation, there have been a lot of debate on the validity of using sampling for evaluating recommendation algorithms. Though their work and the recent work done by Li et. al. (KDD'20) have proposed some basic approach for mapping the sampling based metrics to their counter-part in the global evaluation which uses the entire dataset, there is still lack of understanding how sampling should be used for recommendation evaluation, and the proposed approaches either are rather ad-hoc or can only work on simple metrics, like Recall/Hit-Ratio. In this paper, we introduce some principled approach to derive the estimators of top-k metric based on sampling. Our approaches utilize the weighted MLE and maximal entropy approach to recover the global rank distribution and then utilize that for estimation. The experimental results shows significant advantages of using our approaches for evaluating recommendation algorithms based on top-k metrics.

We study the convergence of Expected Sarsa(λ) with function approximation. We show that with off-line es- timate (multi-step bootstrapping) to ExpectedSarsa(λ) is unstable for off-policy learning. Furthermore, based on convex-concave saddle-point framework, we propose a con- vergent Gradient Expected Sarsa(λ) (GES(λ)) algorithm. The theoretical analysis shows that the proposed GES(λ) converges to the optimal solution at a linear convergence rate under true gradient setting. Furthermore, we develop a Lyapunov function technique to investigate how the step- size influences finite-time performance of GES(λ). Addition- ally, such a technique of Lyapunov function can be poten- tially generalized to other gradient temporal difference algo- rithms. Finally, our experiments verify the effectiveness of our GES(λ). For the details of proof, please refer to https: //arxiv.org/pdf/2012.07199.pdf.

We explore the potential of continuous local search (CLS) in SAT solving by proposing a novel approach for finding a solution of a hybrid system of Boolean constraints. The algorithm is based on CLS combined with belief propagation on binary decision diagrams (BDDs). Our framework accepts all Boolean constraints that admit compact BDDs, including symmetric Boolean constraints and small-coefficient pseudo-Boolean constraints as interesting families. We propose a novel algorithm for efficiently computing the gradient needed by CLS. We study the capabilities and limitations of our versatile CLS solver, GradSAT, by applying it on many benchmark instances. The experimental results indicate that GradSAT can be a useful addition to the portfolio of existing SAT and MaxSAT solvers for solving Boolean satisfiability and optimization problems.

We initiate a comprehensive experimental study of objective-based hierarchical clustering methods on massive datasets consisting of deep embedding vectors from computer vision and NLP applications. This includes a large variety of image embedding (ImageNet, ImageNetV2, NaBirds), word embedding (Twitter, Wikipedia), and sentence embedding (SST-2) vectors from several popular recent models (e.g. ResNet, ResNext, Inception V3, SBERT). Our study includes datasets with up to 4.5 million data points with embedding dimensions up to 2048. In order to address the challenge of scaling up hierarchical clustering to such large datasets we propose a new practical hierarchical clustering algorithm B++&C. It gives a 5%/20% improvement on average for the popular Moseley-Wang (MW)/ Cohen-Addad et al. (CKMM) objectives (normalized) compared to a wide range of classic methods and recent heuristics. We also introduce a theoretical algorithm B2SAT&C which achieves a 0.74-approximation for the CKMM objective in polynomial time. This is the first substantial improvement over the trivial 2/3-approximation achieved by a random binary tree. Prior to this work, the best poly-time approximation of ≈2/3 + 0.0004 was due to Charikar et al. (SODA’19)

We begin with the hypothesis that a model must be able to understand individual objects and relationships between objects in order to generate complex scenes with multiple objects well. Our layout-to-image-generation method, which we call Object-Centric Generative Adversarial Network (or OC-GAN), relies on a novel Scene-Graph Similarity Module (SGSM). The SGSM learns representations of the spatial relationships between objects in the scene, which lead to our model's improved layout-fidelity. We also propose changes to the conditioning mechanism of the generator that enhance its object instance-awareness. Apart from improving image quality, our contributions mitigate two failure modes in previous approaches: (1) spurious objects being generated without corresponding bounding boxes in the layout, and (2) overlapping bounding boxes in the layout leading to merged objects in images. Extensive quantitative evaluation and ablation studies demonstrate the impact of our contributions, with our model outperforming previous state-of-the-art approaches on both the COCO-Stuff and Visual Genome datasets. Finally, we address an important limitation of evaluation metrics used in previous works by introducing SceneFID -- an object-centric adaptation of the popular Fréchet Inception Distance metric, that is better suited for multi-object images.

Image paragraph captioning aims to automatically generate a paragraph from a given image. It is an extension of image captioning in terms of generating multiple sentences instead of a single one, and it is more challenging because paragraphs are longer, more informative, and more linguistically complicated. Because a paragraph consists of several sentences, an effective image paragraph captioning method should generate consistent sentences rather than contradictory ones. It is still an open question how to achieve this goal, and for it we propose a method to incorporate objects' spatial coherence into a language-generating model. For every two overlapping objects, the proposed method concatenates their raw visual features to create two directional pair features and learns weights optimizing those pair features as relation-aware object features for a language-generating model. Experimental results show that the proposed network extracts effective object features for image paragraph captioning and achieves promising performance against existing methods.

A normalizing flow is an invertible mapping between an arbitrary probability distribution and a standard normal distribution; it can be used for density estimation and statistical inference. Computing the flow follows the change of variables formula and thus requires invertibility of the mapping and an efficient way to compute the determinant of its Jacobian. To satisfy these requirements, normalizing flows typically consist of carefully chosen components. Continuous normalizing flows (CNFs) are mappings obtained by solving a neural ordinary differential equation (ODE). The neural ODE's dynamics can be chosen almost arbitrarily while ensuring invertibility. Moreover, the log-determinant of the flow's Jacobian can be obtained by integrating the trace of the dynamics' Jacobian along the flow. Our proposed OT-Flow approach tackles two critical computational challenges that limit a more widespread use of CNFs. First, OT-Flow leverages optimal transport (OT) theory to regularize the CNF and enforce straight trajectories that are easier to integrate. Second, OT-Flow features exact trace computation with time complexity equal to trace estimators used in existing CNFs. On five high-dimensional density estimation and generative modeling tasks, OT-Flow performs competitively to state-of-the-art CNFs while on average requiring one-fourth of the number of weights with an 8x speedup in training time and 24x speedup in inference.

As Deep Neural Networks (DNNs) usually are overparameterized and have millions of weight parameters, it is challenging to deploy these large DNN models on resource-constrained hardware platforms, e.g., smartphones. Numerous network compression methods such as pruning and quantization are proposed to reduce the model size significantly, of which the key is to find suitable compression allocation (e.g., pruning sparsity and quantization codebook) of each layer. Existing solutions obtain the compression allocation in an iterative/manual fashion while finetuning the compressed model, thus suffering from the efficiency issue. Different from the prior art, we propose a novel One-shot Pruning-Quantization (OPQ) in this paper, which analytically solves the compression allocation with pre-trained weight parameters only. During finetuning, the compression module is fixed and only weight parameters are updated. To our knowledge, OPQ is the first work that reveals pre-trained model is sufficient for solving pruning and quantization simultaneously, without any complex iterative/manual optimization at the finetuning stage. Furthermore, we propose a unified channel-wise quantization method that enforces all channels of each layer to share a common codebook, which leads to low bit-rate allocation without introducing extra overhead brought by traditional channel-wise quantization. Comprehensive experiments on ImageNet with AlexNet/MobileNet-V1/ResNet-50 show that our method improves accuracy and training efficiency while obtains significantly higher compression rates compared to the state-of-the-art.

Transformers have emerged as a powerful tool for a broad range of natural language processing tasks. A key component that drives the impressive performance of Transformers is the self-attention mechanism that encodes the influence or dependence of other tokens on each specific token. While beneficial, the quadratic complexity of self-attention on the input sequence length has limited its application to longer sequences - a topic being actively studied in the community. To address this limitation, we propose Nyströmformer - a model that exhibits favorable scalability as a function of sequence length. Our idea is based on adapting the Nyström method to approximate standard self-attention with O(n) complexity. The scalability of Nyströmformer enables application to longer sequences with thousands of tokens. We perform evaluations on multiple downstream tasks on the GLUE benchmark and IMDB reviews with standard sequence length, and find that our Nyströmformer performs comparably, or in a few cases, even slightly better, than standard self-attention. On longer sequence tasks in the Long Range Arena (LRA) benchmark, Nyströmformer performs favorably relative to other efficient self-attention methods. Our code is available at https://github.com/mlpen/Nystromformer.

We introduce Nutri-bullets, a multi-document summarization task for health and nutrition. First, we present two datasets of food and health summaries from multiple scientific studies. Furthermore, we propose a novel extract-compose model to solve the problem in the regime of limited parallel data. We explicitly select key spans from several abstracts using a policy network, followed by composing the selected spans to present a summary via a task specific language model. Compared to state-of-the-art methods, our approach leads to more faithful, relevant and diverse summarization -- properties imperative to this application. For instance, on the BreastCancer dataset our approach gets a more than 50% improvement on relevance and faithfulness.