AAAI 2023 Papers

Source free domain adaptation (SFDA) transfers a single-source model to the unlabeled target domain without accessing the source data. With the intelligence development of various fields, a zoo of source models is more commonly available, arising in a new setting called multi-source-free domain adaptation (MSFDA). We find that the critical inborn challenge of MSFDA is how to estimate the importance (contribution) of each source model. In this paper, we shed new Bayesian light on the fact that the posterior probability of source importance connects to discriminability and transferability. We propose Discriminability And Transferability Estimation (DATE), a universal solution for source importance estimation. Specifically, a proxy discriminability perception module equips with habitat uncertainty and density to evaluate each sample's surrounding environment. A source-similarity transferability perception module quantifies the data distribution similarity and encourages the transferability to be reasonably distributed with a domain diversity loss. Extensive experiments show that DATE can precisely and objectively estimate the source importance and outperform prior arts by non-trivial margins. Moreover, experiments demonstrate that DATE can take the most popular SFDA networks as backbones and make them become advanced MSFDA solutions.

Recent model-extraction attacks on Machine Learning as a Service (MLaaS) systems have moved towards data-free approaches, showing the feasibility of stealing models trained with difficult-to-access data. However, these attacks are ineffective or limited due to the low accuracy of extracted models and the high number of queries to the models under attack. The high query cost makes such techniques infeasible for online MLaaS systems that charge per query. We create a novel approach to get higher accuracy and query efficiency than prior data-free model extraction techniques. Specifically, we introduce a novel generator training scheme that maximizes the disagreement loss between two clone models that attempt to copy the model under attack. This loss, combined with diversity loss and experience replay, enables the generator to produce better instances to train the clone models. Our evaluation on popular datasets CIFAR-10 and CIFAR-100 shows that our approach improves the final model accuracy by up to 3.42% and 18.48% respectively. The average number of queries required to achieve the accuracy of the prior state of the art is reduced by up to 64.95%. We hope this will promote future work on feasible data-free model extraction and defenses against such attacks.

Estimating the structure of directed acyclic graphs (DAGs) of features (variables) plays a vital role in revealing the latent data generation process and providing causal insights in various applications. Although there have been many studies on structure learning with various types of data, the structure learning on the dynamic graph has not been explored yet, and thus we study the learning problem of node feature generation mechanism on such ubiquitous dynamic graph data. In a dynamic graph, we propose to simultaneously estimate contemporaneous relationships and time-lagged interaction relationships between the node features. These two kinds of relationships form a DAG, which could effectively characterize the feature generation process in a concise way. To learn such a DAG, we cast the learning problem as a continuous score-based optimization problem, which consists of a differentiable score function to measure the validity of the learned DAGs and a smooth acyclicity constraint to ensure the acyclicity of the learned DAGs. These two components are translated into an unconstraint augmented Lagrangian objective which could be minimized by mature continuous optimization techniques. The resulting algorithm, named GraphNOTEARS, outperforms baselines on simulated data across a wide range of settings that may encounter in real-world applications. We also apply the proposed approach on two dynamic graphs constructed from the real-world Yelp dataset, demonstrating our method could learn the connections between node features, which conforms with the domain knowledge.

Marketing is an important mechanism to increase user engagement and improve platform revenue, and heterogeneous causal learning can help develop more effective strategies. Most decision-making problems in marketing can be formulated as resource allocation problems and have been studied for decades. Existing works usually divide the solution procedure into two fully decoupled stages, i.e., machine learning (ML) and operation research (OR) --- the first stage predicts the model parameters and they are fed to the optimization in the second stage. However, the error of the predicted parameters in ML cannot be respected and a series of complex mathematical operations in OR lead to the increased accumulative errors. Essentially, the improved precision on the prediction parameters may not have a positive correlation on the final solution due to the side-effect from the decoupled design. In this paper, we propose a novel approach for solving resource allocation problems to mitigate the side-effects. Our key intuition is that we introduce the decision factor to establish a bridge between ML and OR such that the solution can be directly obtained in OR by only performing the sorting or comparison operations on the decision factor. Furthermore, we design a customized loss function that can conduct direct heterogeneous causal learning on the decision factor, an unbiased estimation of which can be guaranteed when the loss convergences. As a case study, we apply our approach to two crucial problems in marketing: the binary treatment assignment problem and the budget allocation problem with multiple treatments. Both large-scale simulations and online A/B Tests demonstrate that our approach achieves significant improvement compared with state-of-the-art.

Structural topology optimization, which aims to find the optimal physical structure that maximizes mechanical performance, is vital in engineering design applications in aerospace, mechanical, and civil engineering. Recently, generative adversarial networks (GANs) have emerged as a popular alternative to traditional iterative topology optimization methods. However, GANs can be challenging to train, have limited generalizability, and often neglect important performance objectives such as mechanical compliance and manufacturability. To address these issues, we propose a new architecture called TopoDiff that uses conditional diffusion models to perform performance-aware and manufacturability-aware topology optimization. Our method introduces a surrogate model-based guidance strategy that actively favors structures with low compliance and good manufacturability. Compared to a state-of-the-art conditional GAN, our approach reduces the average error on physical performance by a factor of eight and produces eleven times fewer infeasible samples. Our work demonstrates the potential of using diffusion models in topology optimization and suggests a general framework for solving engineering optimization problems using external performance with constraint-aware guidance. We provide access to our data, code, and trained models at the following link: https://decode.mit.edu/projects/topodiff/.

Learning a categorical distribution comes with its own set of challenges. A successful approach taken by state-of-the-art works is to cast the problem in a continuous domain to take advantage of the impressive performance of the generative models for continuous data. Amongst them are the recently emerging diffusion probabilistic models, which have the observed advantage of generating high-quality samples. Recent advances for categorical generative models have focused on log likelihood improvements. In this work, we propose a generative model for categorical data based on diffusion models with a focus on high-quality sample generation, and propose sampled-based evaluation methods. The efficacy of our method stems from performing diffusion in the continuous domain while having its parameterization informed by the structure of the categorical nature of the target distribution. Our method of evaluation highlights the capabilities and limitations of different generative models for generating categorical data, and includes experiments on synthetic and real-world protein datasets.

Efficient Transformers have been developed for long sequence modeling, due to their subquadratic memory and time complexity. Sparse Transformer is a popular approach to improving the efficiency of Transformers by restricting self-attention to locations specified by the predefined sparse patterns. However, leveraging sparsity may sacrifice expressiveness compared to full-attention, when important token correlations are multiple hops away. To combine advantages of both the efficiency of sparse transformer and the expressiveness of full-attention Transformer, we propose Diffuser, a new state-of-the-art efficient Transformer. Diffuser incorporates all token interactions within one attention layer while maintaining low computation and memory costs. The key idea is to expand the receptive field of sparse attention using Attention Diffusion, which computes multi-hop token correlations based on all paths between corresponding disconnected tokens, besides attention among neighboring tokens. Theoretically, we show the expressiveness of Diffuser as a universal sequence approximator for sequence-to-sequence modeling, and investigate its ability to approximate full-attention by analyzing the graph expander property from the spectral perspective. Experimentally, we investigate the effectiveness of Diffuser with extensive evaluations, including language modeling, image modeling, and Long Range Arena (LRA). Evaluation results show that Diffuser achieves improvements by an average of 0.94% on text classification tasks and 2.30% on LRA, with 1.67x memory savings compared to state-of-the-art benchmarks, which demonstrates superior performance of Diffuser in both expressiveness and efficiency aspects.

Recently, several methods such as private ANM, EM-PC and Priv-PC have been proposed to perform differentially private causal discovery in various scenarios including bivariate, multivariate Gaussian and categorical cases. However, there is little effort on how to conduct private nonlinear causal discovery from numerical data. This work tries to challenge this problem. To this end, we propose a method to infer nonlinear causal relations from observed numerical data by using regression-based conditional independence test (RCIT) that consists of kernel ridge regression (KRR) and Hilbert-Schmidt independence criterion (HSIC) with permutation approximation. Sensitivity analysis for RCIT is given and a private constraint-based causal discovery framework with differential privacy guarantee is developed. Extensive simulations and real-world experiments for both conditional independence test and causal discovery are conducted, which show that our method is effective in handling nonlinear numerical cases and easy to implement. The source code of our method and data are available at https://github.com/Causality-Inference/PCD.

Privacy in AI remains a topic that draws attention from researchers and the general public in recent years. As one way to implement privacy-preserving AI, differentially private learning is a framework that enables AI models to use differential privacy (DP). To achieve DP in the learning process, existing algorithms typically limit the magnitude of gradients with a constant clipping, which requires carefully tuned due to its significant impact on model performance. As a solution to this issue, latest works NSGD and Auto-S innovatively propose to use normalization instead of clipping to avoid hyperparameter tuning. However, normalization-based approaches like NSGD and Auto-S rely on a monotonic weight function, which imposes excessive weight on small gradient samples and introduces extra deviation to the update. In this paper, we propose a Differentially Private Per-Sample Adaptive Clipping (DP-PSAC) algorithm based on a non-monotonic adaptive weight function, which guarantees privacy without the typical hyperparameter tuning process of using a constant clipping while significantly reducing the deviation between the update and true batch-averaged gradient. We provide a rigorous theoretical convergence analysis and show that with convergence rate at the same order, the proposed algorithm achieves a lower non-vanishing bound, which is maintained over training iterations, compared with NSGD/Auto-S. In addition, through extensive experimental evaluation, we show that DP-PSAC outperforms or matches the state-of-the-art methods on multiple main-stream vision and language tasks.

We consider the task of producing heatmaps from users' aggregated data while protecting their privacy. We give a differentially private (DP) algorithm for this task and demonstrate its advantages over previous algorithms on real-world datasets. Our core algorithmic primitive is a DP procedure that takes in a set of distributions and produces an output that is close in Earth Mover's Distance (EMD) to the average of the inputs. We prove theoretical bounds on the error of our algorithm under a certain sparsity assumption and that these are essentially optimal.

Fairness and privacy are two important concerns in social decision-making processes such as resource allocation. We study privacy in the fair allocation of indivisible resources using the well-established framework of differential privacy. We present algorithms for approximate envy-freeness and proportionality when two instances are considered to be adjacent if they differ only on the utility of a single agent for a single item. On the other hand, we provide strong negative results for both fairness criteria when the adjacency notion allows the entire utility function of a single agent to change.

Designing private voting rules is an important and pressing problem for trustworthy democracy. In this paper, under the framework of differential privacy, we propose a novel famliy of randomized voting rules based on the well-known Condorcet method, and focus on three classes of voting rules in this family: Laplacian Condorcet method (CMLAP), exponential Condorcet method (CMEXP), and randomized response Condorcet method (CMRR), where λ represents the level of noise. We prove that all of our rules satisfy absolute monotonicity, lexi-participation, probabilistic Pareto efficiency, approximate probabilistic Condorcet criterion, and approximate SD-strategyproofness. In addition, CMRR satisfies (non-approximate) probabilistic Condorcet criterion, while CMLAP and CMEXP satisfy strong lexi-participation. Finally, we regard differential privacy as a voting axiom, and discuss its relations to other axioms.

Heterogeneous information networks (HINs) are widely employed for describing real-world data with intricate entities and relationships. To automatically utilize their semantic information, graph neural architecture search has recently been developed for various tasks of HINs. Existing works, on the other hand, show weaknesses in instability and inflexibility. To address these issues, we propose a novel method called Partial Message Meta Multigraph search (PMMM) to automatically optimize the neural architecture design on HINs. Specifically, to learn how graph neural networks (GNNs) propagate messages along various types of edges, PMMM adopts an efficient differentiable framework to search for a meaningful meta multigraph, which can capture more flexible and complex semantic relations than a meta graph. The differentiable search typically suffers from performance instability, so we further propose a stable algorithm called partial message search to ensure that the searched meta multigraph consistently surpasses the manually designed meta-structures, i.e., meta-paths. Extensive experiments on six benchmark datasets over two representative tasks, including node classification and recommendation, demonstrate the effectiveness of the proposed method. Our approach outperforms the state-of-the-art heterogeneous GNNs, finds out meaningful meta multigraphs, and is significantly more stable. Our code is available at https://github.com/JHL-HUST/PMMM.

Mutual Information (MI) and Conditional Mutual Information (CMI) are multi-purpose tools from information theory that are able to naturally measure the statistical dependencies between random variables, thus they are usually of central interest in several statistical and machine learning tasks, such as conditional independence testing and representation learning. However, estimating CMI, or even MI, is infamously challenging due the intractable formulation. In this study, we introduce DINE (Diffeomorphic Information Neural Estimator)–a novel approach for estimating CMI of continuous random variables, inspired by the invariance of CMI over diffeomorphic maps. We show that the variables of interest can be replaced with appropriate surrogates that follow simpler distributions, allowing the CMI to be efficiently evaluated via analytical solutions. Additionally, we demonstrate the quality of the proposed estimator in comparison with state-of-the-arts in three important tasks, including estimating MI, CMI, as well as its application in conditional independence testing. The empirical evaluations show that DINE consistently outperforms competitors in all tasks and is able to adapt very well to complex and high-dimensional relationships.

Molecular dynamics (MD) has long been the de facto choice for simulating complex atomistic systems from first principles. Recently deep learning models become a popular way to accelerate MD. Notwithstanding, existing models depend on intermediate variables such as the potential energy or force fields to update atomic positions, which requires additional computations to perform back-propagation. To waive this requirement, we propose a novel model called DiffMD by directly estimating the gradient of the log density of molecular conformations. DiffMD relies on a score-based denoising diffusion generative model that perturbs the molecular structure with a conditional noise depending on atomic accelerations and treats conformations at previous timeframes as the prior distribution for sampling. Another challenge of modeling such a conformation generation process is that a molecule is kinetic instead of static, which no prior works have strictly studied. To solve this challenge, we propose an equivariant geometric Transformer as the score function in the diffusion process to calculate corresponding gradients. It incorporates the directions and velocities of atomic motions via 3D spherical Fourier-Bessel representations. With multiple architectural improvements, we outperform state-of-the-art baselines on MD17 and isomers of C7O2H10 datasets. This work contributes to accelerating material and drug discovery.

In task-oriented dialogue systems, Dialogue State Tracking (DST) aims to extract users' intentions from the dialogue history. Currently, most existing approaches suffer from error propagation and are unable to dynamically select relevant information when utilizing previous dialogue states. Moreover, the relations between the updates of different slots provide vital clues for DST. However, the existing approaches rely only on predefined graphs to indirectly capture the relations. In this paper, we propose a Dialogue State Distillation Network (DSDN) to utilize relevant information of previous dialogue states and migrate the gap of utilization between training and testing. Thus, it can dynamically exploit previous dialogue states and avoid introducing error propagation simultaneously. Further, we propose an inter-slot contrastive learning loss to effectively capture the slot co-update relations from dialogue context. Experiments are conducted on the widely used MultiWOZ 2.0 and MultiWOZ 2.1 datasets. The experimental results show that our proposed model achieves the state-of-the-art performance for DST.

Dialogue rewriting aims to transform multi-turn, context-dependent dialogues into well-formed, context-independent text for most NLP systems. Previous dialogue rewriting benchmarks and systems assume a fluent and informative utterance to rewrite. Unfortunately, dialogue utterances from real-world systems are frequently noisy and with various kinds of errors that can make them almost uninformative. In this paper, we first present Real-world Dialogue Rewriting Corpus (RealDia), a new benchmark to evaluate how well current dialogue rewriting systems can deal with real-world noisy and uninformative dialogue utterances. RealDia contains annotated multi-turn dialogues from real scenes with ASR errors, spelling errors, redundancies and other noises that are ignored by previous dialogue rewriting benchmarks. We show that previous dialogue rewriting approaches are neither effective nor data-efficient to resolve RealDia. Then this paper presents Skeleton-Guided Rewriter (SGR), which can resolve the task of dialogue rewriting via a skeleton-guided generation paradigm. Experiments show that RealDia is a much more challenging benchmark for real-world dialogue rewriting, and SGR can effectively resolve the task and outperform previous approaches by a large margin.

Feature acquisition in predictive modeling is an important task in many practical applications. For example, in patient health prediction, we do not fully observe their personal features and need to dynamically select features to acquire. Our goal is to acquire a small subset of features that maximize prediction performance. Recently, some works reformulated feature acquisition as a Markov decision process and applied reinforcement learning (RL) algorithms, where the reward reflects both prediction performance and feature acquisition cost. However, RL algorithms only use zeroth-order information on the reward, which leads to slow empirical convergence, especially when there are many actions (number of features) to consider. For predictive modeling, it is possible to use first-order information on the reward, i.e., gradients, since we are often given an already collected dataset. Therefore, we propose differentiable feature acquisition (DiFA), which uses a differentiable representation of the feature selection policy to enable gradients to flow from the prediction loss to the policy parameters. We conduct extensive experiments on various real-world datasets and show that DiFA significantly outperforms existing feature acquisition methods when the number of features is large.

Agent-based modeling and simulation can provide a powerful test environment for crisis management scenarios. Human agent interaction has limitations in representing norms issued by an agent to a human agent that has emotions. In this study, we present an approach to the interaction between a virtual normative agent and a human agent in an evacuation scenario. Through simulation comparisons, it is shown that the method used in this study can more fully simulate the real-life out come of an emergency situation and also improves the au thenticity of the agent interaction.