AAAI 2023 Papers

We study the algorithm configuration (AC) problem, in which one seeks to find an optimal parameter configuration of a given target algorithm in an automated way. Although this field of research has experienced much progress recently regarding approaches satisfying strong theoretical guarantees, there is still a gap between the practical performance of these approaches and the heuristic state-of-the-art approaches. Recently, there has been significant progress in designing AC approaches that satisfy strong theoretical guarantees. However, a significant gap still remains between the practical performance of these approaches and state-of-the-art heuristic methods. To this end, we introduce AC-Band, a general approach for the AC problem based on multi-armed bandits that provides theoretical guarantees while exhibiting strong practical performance. We show that AC-Band requires significantly less computation time than other AC approaches providing theoretical guarantees while still yielding high-quality configurations.

Deep neural networks (DNNs) can easily be manipulated (by an adversary) to output drastically different predictions and can be done so in a controlled and directed way. This process is known as adversarial attack and is considered one of the major hurdles in using DNNs in high-stakes and real-world applications. Although developing methods to secure DNNs against adversaries is now a primary research focus, it suffers from limitations such as lack of optimization generality and lack of optimization scalability. My research highlights will offer a holistic understanding of optimization foundations for robust AI, peer into their emerging challenges, and present recent solutions developed by my research group.

Recent Text-to-Speech (TTS) systems trained on reading or acted corpora have achieved near human-level naturalness. The diversity of human speech, however, often goes beyond the coverage of these corpora. We believe the ability to handle such diversity is crucial for AI systems to achieve human-level communication. Our work explores the use of more abundant real-world data for building speech synthesizers. We train TTS systems using real-world speech from YouTube and podcasts. We observe the mismatch between training and inference alignments in mel-spectrogram based autoregressive models, leading to unintelligible synthesis, and demonstrate that learned discrete codes within multiple code groups effectively resolves this issue. We introduce our MQTTS system whose architecture is designed for multiple code generation and monotonic alignment, along with the use of a clean silence prompt to improve synthesis quality. We conduct ablation analyses to identify the efficacy of our methods. We show that MQTTS outperforms existing TTS systems in several objective and subjective measures.

A common musical composition practice is to develop musical pieces using variations of musical themes. In this study, we present an interactive tool which can generate variations of musical themes in real-time using a variational autoencoder model. Our tool is controllable using semantically meaningful musical attributes via latent space regularisation technique to increase the explainability of the model. The tool is integrated into an industry standard digital audio workstation - Ableton Live - using the Max4Live device framework and can run locally on an average personal CPU rather than requiring a costly GPU cluster. In this way we demonstrate how cutting-edge AI research can be integrated into the exiting workflows of professional and practising musicians for use in the real-world beyond the research lab.

We study a normalizing flow in the latent space of a top-down generator model, in which the normalizing flow model plays the role of the informative prior model of the generator. We propose to jointly learn the latent space normalizing flow prior model and the top-down generator model by a Markov chain Monte Carlo (MCMC)-based maximum likelihood algorithm, where a short-run Langevin sampling from the intractable posterior distribution is performed to infer the latent variables for each observed example, so that the parameters of the normalizing flow prior and the generator can be updated with the inferred latent variables. We show that, under the scenario of non-convergent short-run MCMC, the finite step Langevin dynamics is a flow-like approximate inference model and the learning objective actually follows the perturbation of the maximum likelihood estimation (MLE). We further point out that the learning framework seeks to (i) match the latent space normalizing flow and the aggregated posterior produced by the short-run Langevin flow, and (ii) bias the model from MLE such that the short-run Langevin flow inference is close to the true posterior. Empirical results of extensive experiments validate the effectiveness of the proposed latent space normalizing flow model in the tasks of image generation, image reconstruction, anomaly detection, supervised image inpainting and unsupervised image recovery.

Despite achieving state-of-the-art performance on many NLP tasks, the high energy cost and long inference delay prevent Transformer-based pretrained language models (PLMs) from seeing broader adoption including for edge and mobile computing. Efficient NLP research aims to comprehensively consider computation, time and carbon emission for the entire life-cycle of NLP, including data preparation, model training and inference. In this survey, we focus on the inference stage and review the current state of model compression and acceleration for pretrained language models, including benchmarks, metrics and methodology.

There has been a consensus on integrating Computing into the teaching and learning of STEM (Science, Technology, Engineering and Math) subjects in K-12 (Kindergarten to 12th grade in the US education system). However, rigorous study on the impact of an integrated curriculum on students' learning in computing and/or the STEM subject(s) is still rare. In this paper, we report our research on how well an integrated curriculum helps middle school students learn Computing through the microgenetic analysis methods.

Synchronous dynamical systems are well-established models that have been used to capture a range of phenomena in networks, including opinion diffusion, spread of disease and product adoption. We study the three most notable problems in synchronous dynamical systems: whether the system will transition to a target configuration from a starting configuration, whether the system will reach convergence from a starting configuration, and whether the system is guaranteed to converge from every possible starting configuration. While all three problems were known to be intractable in the classical sense, we initiate the study of their exact boundaries of tractability from the perspective of structural parameters of the network by making use of the more fine-grained parameterized complexity paradigm. As our first result, we consider treewidth - as the most prominent and ubiquitous structural parameter - and show that all three problems remain intractable even on instances of constant treewidth. We complement this negative finding with fixed-parameter algorithms for the former two problems parameterized by treedepth, a well-studied restriction of treewidth. While it is possible to rule out a similar algorithm for convergence guarantee under treedepth, we conclude with a fixed-parameter algorithm for this last problem when parameterized by treedepth and the maximum in-degree.

User Satisfaction Estimation is an important task and increasingly being applied in goal-oriented dialogue systems to estimate whether the user is satisfied with the service. It is observed that whether the user’s needs are met often triggers various sentiments, which can be pertinent to the successful estimation of user satisfaction, and vice versa. Thus, User Satisfaction Estimation (USE) and Sentiment Analysis (SA) should be treated as a joint, collaborative effort, considering the strong connections between the sentiment states of speakers and the user satisfaction. Existing joint learning frameworks mainly unify the two highly pertinent tasks over cascade or shared-bottom implementations, however they fail to distinguish task-specific and common features, which will produce sub-optimal utterance representations for downstream tasks. In this paper, we propose a novel Speaker Turn-Aware Multi-Task Adversarial Network (STMAN) for dialogue-level USE and utterance-level SA. Specifically, we first introduce a multi-task adversarial strategy which trains a task discriminator to make utterance representation more task-specific, and then utilize a speaker-turn aware multi-task interaction strategy to extract the common features which are complementary to each task. Extensive experiments conducted on two real-world service dialogue datasets show that our model outperforms several state-of-the-art methods.

Cross-domain named entity recognition (NER), aiming to address the limitation of labeled resources in the target domain, is a challenging yet important task. Most existing studies alleviate the data discrepancy across different domains at the coarse level via combing NER with language modelings or introducing domain-adaptive pre-training (DAPT). Notably, source and target domains tend to share more fine-grained local information within denser subsequences than global information within the whole sequence, such that subsequence features are easier to transfer, which has not been explored well. Besides, compared to token-level representation, subsequence-level information can help the model distinguish different meanings of the same word in different domains. In this paper, we propose to incorporate subsequence-level features for promoting the cross-domain NER. In detail, we first utilize a pre-trained encoder to extract the global information. Then, we re-express each sentence as a group of subsequences and propose a novel bidirectional memory recurrent unit (BMRU) to capture features from the subsequences. Finally, an adaptive coupling unit (ACU) is proposed to combine global information and subsequence features for predicting entity labels. Experimental results on several benchmark datasets illustrate the effectiveness of our model, which achieves considerable improvements.

Conditional independence (CI) tests underlie many approaches to model testing and structure learning in causal inference. Most existing CI tests for categorical and ordinal data stratify the sample by the conditioning variables, perform simple independence tests in each stratum, and combine the results. Unfortunately, the statistical power of this approach degrades rapidly as the number of conditioning variables increases. Here we propose a simple unified CI test for ordinal and categorical data that maintains reasonable calibration and power in high dimensions. We show that our test outperforms existing baselines in model testing and structure learning for dense directed graphical models while being comparable for sparse models. Our approach could be attractive for causal model testing because it is easy to implement, can be used with non-parametric or parametric probability models, has the symmetry property, and has reasonable computational requirements.

3D object detection with surrounding cameras has been a promising direction for autonomous driving. In this paper, we present SimMOD, a Simple baseline for Multi-camera Object Detection, to solve the problem. To incorporate multiview information as well as build upon previous efforts on monocular 3D object detection, the framework is built on sample-wise object proposals and designed to work in a twostage manner. First, we extract multi-scale features and generate the perspective object proposals on each monocular image. Second, the multi-view proposals are aggregated and then iteratively refined with multi-view and multi-scale visual features in the DETR3D-style. The refined proposals are endto-end decoded into the detection results. To further boost the performance, we incorporate the auxiliary branches alongside the proposal generation to enhance the feature learning. Also, we design the methods of target filtering and teacher forcing to promote the consistency of two-stage training. We conduct extensive experiments on the 3D object detection benchmark of nuScenes to demonstrate the effectiveness of SimMOD and achieve competitive performance. Code will be available at https://github.com/zhangyp15/SimMOD.

Predicting the trajectories of pedestrians in crowded conditions is an important task for applications like autonomous navigation systems. Previous studies have tackled this problem using two strategies. They (1) infer all future steps recursively, or (2) predict the potential destinations of pedestrians at once and interpolate the intermediate steps to arrive there. However, these strategies often suffer from the accumulated errors of the recursive inference, or restrictive assumptions about social relations in the intermediate path. In this paper, we present a graph convolutional network-based trajectory prediction. Firstly, we propose a control point prediction that divides the future path into three sections and infers the intermediate destinations of pedestrians to reduce the accumulated error. To do this, we construct multi-relational weighted graphs to account for their physical and complex social relations. We then introduce a trajectory refinement step based on a spatio-temporal and multi-relational graph. By considering the social interactions between neighbors, better prediction results are achievable. In experiments, the proposed network achieves state-of-the-art performance on various real-world trajectory prediction benchmarks.

We introduce a novel method based on semidefinite program (SDP) for the tight and efficient verification of neural networks. The proposed SDP relaxation advances the present state of the art in SDP-based neural network verification by adding a set of linear constraints based on eigenvectors. We extend this novel SDP relaxation by combining it with a branch-and-bound method that can provably close the relaxation gap up to zero. We show formally that the proposed approach leads to a provably tighter solution than the present state of the art. We report experimental results showing that the proposed method outperforms baselines in terms of verified accuracy while retaining an acceptable computational overhead.

Psychometric functions typically characterize binary sensory decisions along a single stimulus dimension. However, real-life sensory tasks vary along a greater variety of dimensions (e.g. color, contrast and luminance for visual stimuli). Approaches to characterizing high-dimensional sensory spaces either require strong parametric assumptions about these additional contextual dimensions, or fail to leverage known properties of classical psychometric curves. We overcome both limitations by introducing a semi-parametric model of sensory discrimination that applies traditional psychophysical models along a stimulus intensity dimension, but puts Gaussian process (GP) priors on the parameters of these models with respect to the remaining dimensions. By combining the flexibility of the GP with the deep literature on parametric psychophysics, our semi-parametric models achieve good performance with much less data than baselines on both synthetic and real-world, high-dimensional psychophysics datasets. We additionally show strong performance in a Bayesian active learning setting, and present a novel active learning paradigm for the semi-parametric model.

Multi-Intent Spoken Language Understanding (SLU), a novel and more complex scenario of SLU, is attracting increasing attention. Unlike traditional SLU, each intent in this scenario has its specific scope. Semantic information outside the scope even hinders the prediction, which tremendously increases the difficulty of intent detection. More seriously, guiding slot filling with these inaccurate intent labels suffers error propagation problems, resulting in unsatisfied overall performance. To solve these challenges, in this paper, we propose a novel Scope-Sensitive Result Attention Network (SSRAN) based on Transformer, which contains a Scope Recognizer (SR) and a Result Attention Network (RAN). SR assignments scope information to each token, reducing the distraction of out-of-scope tokens. RAN effectively utilizes the bidirectional interaction between SF and ID results, mitigating the error propagation problem. Experiments on two public datasets indicate that our model significantly improves SLU performance (5.4% and 2.1% on Overall accuracy) over the state-of-the-art baseline.

Accurate day-ahead nominations of grid losses in electrical distribution networks are important to reduce the societal cost of these losses. We present a modification of the CatBoost ensemble-based system for day-ahead grid loss prediction detailed in Dalal et al. (2020), making four main changes. Base models predict on the log-space of the target, to ensure non-negative predictions. The model ensemble is changed to include different model types, for increased ensemble variance. Feature engineering is applied to consumption and weather forecasts, to improve base model performance. Finally, a non-negative least squares-based stacking method that uses as many available models as possible for each prediction is introduced, to achieve an improved model selection that is robust to missing data. When deployed for over three months in 2022, the resulting system reduced mean absolute error by 10.7% compared to the system from Dalal et al. (2020), a reduction from 5.05 to 4.51 MW. With no tuning of machine learning parameters, the system was also extended to three new grids, where it achieved similar relative error as on the old grids. Our system is robust and easily scalable, and our proposed stacking method could provide improved performance in applications outside grid loss.

Standard deep reinforcement learning (DRL) aims to maximize expected reward, considering collected experiences equally in formulating a policy. This differs from human decision-making, where gains and losses are valued differently and outlying outcomes are given increased consideration. It also fails to capitalize on opportunities to improve safety and/or performance through the incorporation of distributional context. Several approaches to distributional DRL have been investigated, with one popular strategy being to evaluate the projected distribution of returns for possible actions. We propose a more direct approach whereby risk-sensitive objectives, specified in terms of the cumulative distribution function (CDF) of the distribution of full-episode rewards, are optimized. This approach allows for outcomes to be weighed based on relative quality, can be used for both continuous and discrete action spaces, and may naturally be applied in both constrained and unconstrained settings. We show how to compute an asymptotically consistent estimate of the policy gradient for a broad class of risk-sensitive objectives via sampling, subsequently incorporating variance reduction and regularization measures to facilitate effective on-policy learning. We then demonstrate that the use of moderately "pessimistic" risk profiles, which emphasize scenarios where the agent performs poorly, leads to enhanced exploration and a continual focus on addressing deficiencies. We test the approach using different risk profiles in six OpenAI Safety Gym environments, comparing to state of the art on-policy methods. Without cost constraints, we find that pessimistic risk profiles can be used to reduce cost while improving total reward accumulation. With cost constraints, they are seen to provide higher positive rewards than risk-neutral approaches at the prescribed allowable cost.

Recent techniques for analyzing sports precisely has stimulated various approaches to improve player performance and fan engagement. However, existing approaches are only able to evaluate offline performance since testing in real-time matches requires exhaustive costs and cannot be replicated. To test in a safe and reproducible simulator, we focus on turn-based sports and introduce a badminton environment by simulating rallies with different angles of view and designing the states, actions, and training procedures. This benefits not only coaches and players by simulating past matches for tactic investigation, but also researchers from rapidly evaluating their novel algorithms. Our code is available at https://github.com/wywyWang/CoachAI-Projects/tree/main/Strategic%20Environment.