AAAI 2023 Papers

For many applications of Markov Decision Processes (MDPs), the transition function cannot be specified exactly. Bayes-Adaptive MDPs (BAMDPs) extend MDPs to consider transition probabilities governed by latent parameters. To act optimally in BAMDPs, one must maintain a belief distribution over the latent parameters. Typically, this distribution is described by a set of sample (particle) MDPs, and associated weights which represent the likelihood of a sample MDP being the true underlying MDP. However, as the number of dimensions of the latent parameter space increases, the number of sample MDPs required to sufficiently represent the belief distribution grows exponentially. Thus, maintaining an accurate belief in the form of a set of sample MDPs over complex latent spaces is computationally intensive, which in turn affects the performance of planning for these models. In this paper, we propose an alternative approach for maintaining the belief over the latent parameters. We consider a class of BAMDPs where the transition probabilities can be expressed in closed form as a polynomial of the latent parameters, and outline a method to maintain a closed-form belief distribution for the latent parameters which results in an accurate belief representation. Furthermore, the closed-form representation does away with the need to tune the number of sample MDPs required to represent the belief. We evaluate two domains and empirically show that the polynomial, closed-form, belief representation results in better plans than a sampling-based belief representation.

Autonomous agents embedded in a physical environment need the ability to recognize objects and their properties from sensory data. Such a perceptual ability is often implemented by supervised machine learning models, which are pre-trained using a set of labelled data. In real-world, open-ended deployments, however, it is unrealistic to assume to have a pre-trained model for all possible environments. Therefore, agents need to dynamically learn/adapt/extend their perceptual abilities online, in an autonomous way, by exploring and interacting with the environment where they operate. This paper describes a way to do so, by exploiting symbolic planning. Specifically, we formalize the problem of automatically training a neural network to recognize object properties as a symbolic planning problem (using PDDL). We use planning techniques to produce a strategy for automating the training dataset creation and the learning process. Finally, we provide an experimental evaluation in both a simulated and a real environment, which shows that the proposed approach is able to successfully learn how to recognize new object properties.

Some of the most powerful reinforcement learning frameworks use planning for action selection. Interestingly, their planning horizon is either fixed or determined arbitrarily by the state visitation history. Here, we expand beyond the naive fixed horizon and propose a theoretically justified strategy for adaptive selection of the planning horizon as a function of the state-dependent value estimate. We propose two variants for lookahead selection and analyze the trade-off between iteration count and computational complexity per iteration. We then devise a corresponding deep Q-network algorithm with an adaptive tree search horizon. We separate the value estimation per depth to compensate for the off-policy discrepancy between depths. Lastly, we demonstrate the efficacy of our adaptive lookahead method in a maze environment and Atari.

Safe and reliable decision-making is critical for long-term deployment of autonomous systems. Despite the recent advances in artificial intelligence, ensuring safe and reliable operation of human-aligned autonomous systems in open-world environments remains a challenge. My research focuses on developing planning and learning algorithms that support reliable autonomy in fully and partially observable environments, in the presence of uncertainty, limited information, and limited resources. This talk covers a summary of my recent research towards reliable autonomy.

Autonomous systems are often deployed in the open world where it is hard to obtain complete specifications of objectives and constraints. Operating based on an incomplete model can produce negative side effects (NSEs), which affect the safety and reliability of the system. We focus on mitigating NSEs in environments modeled as Markov decision processes (MDPs). First, we learn a model of NSEs using observed data that contains state-action trajectories and severity of associated NSEs. Unlike previous works that associate NSEs with state-action pairs, our framework associates NSEs with entire trajectories, which is more general and captures non-Markovian dependence on states and actions. Second, we learn finite state controllers (FSCs) that predict NSE severity for a given trajectory and generalize well to unseen data. Finally, we develop a constrained MDP model that uses information from the underlying MDP and the learned FSC for planning while avoiding NSEs. Our empirical evaluation demonstrates the effectiveness of our approach in learning and mitigating Markovian and non-Markovian NSEs.

Existing image stitching approaches based on global or local homography estimation are not free from the parallax problem and suffer from undesired artifacts. In this paper, instead of relying on the homography-based warp, we propose a novel deep image stitching framework exploiting the pixel-wise warp field to handle the large-parallax problem. The proposed deep image stitching framework consists of a Pixel-wise Warping Module (PWM) and a Stitched Image Generating Module (SIGMo). For PWM, we obtain pixel-wise warp in a similar manner as estimating an optical flow (OF). In the stitching scenario, the input images usually include non-overlap (NOV) regions of which warp cannot be directly estimated, unlike the overlap (OV) regions. To help the PWM predict a reasonable warp on the NOV region, we impose two geometrical constraints: an epipolar loss and a line-preservation loss. With the obtained warp field, we relocate the pixels of the target image using forward warping. Finally, the SIGMo is trained by the proposed multi-branch training framework to generate a stitched image from a reference image and a warped target image. For training and evaluating the proposed framework, we build and publish a novel dataset including image pairs with corresponding pixel-wise ground truth warp and stitched result images. We show that the results of the proposed framework are quantitatively and qualitatively superior to those of the conventional methods.

Although weakly-supervised techniques can reduce the labeling effort, it is unclear whether a saliency model trained with weakly-supervised data (e.g., point annotation) can achieve the equivalent performance of its fully-supervised version. This paper attempts to answer this unexplored question by proving a hypothesis: there is a point-labeled dataset where saliency models trained on it can achieve equivalent performance when trained on the densely annotated dataset. To prove this conjecture, we proposed a novel yet effective adversarial trajectory-ensemble active learning (ATAL). Our contributions are three-fold: 1) Our proposed adversarial attack triggering uncertainty can conquer the overconfidence of existing active learning methods and accurately locate these uncertain pixels. 2) Our proposed trajectory-ensemble uncertainty estimation method maintains the advantages of the ensemble networks while significantly reducing the computational cost. 3) Our proposed relationship-aware diversity sampling algorithm can conquer oversampling while boosting performance. Experimental results show that our ATAL can find such a point-labeled dataset, where a saliency model trained on it obtained 97%-99% performance of its fully-supervised version with only 10 annotated points per image.

Multi-task deep reinforcement learning (DRL) ambitiously aims to train a general agent that masters multiple tasks simultaneously. However, varying learning speeds of different tasks compounding with negative gradients interference makes policy learning inefficient. In this work, we propose PiCor, an efficient multi-task DRL framework that splits learning into policy optimization and policy correction phases. The policy optimization phase improves the policy by any DRL algothrim on the sampled single task without considering other tasks. The policy correction phase first constructs an adaptive adjusted performance constraint set. Then the intermediate policy learned by the first phase is constrained to the set, which controls the negative interference and balances the learning speeds across tasks. Empirically, we demonstrate that PiCor outperforms previous methods and significantly improves sample efficiency on simulated robotic manipulation and continuous control tasks. We additionally show that adaptive weight adjusting can further improve data efficiency and performance.

This paper proposes a physics-guided neural network model to predict crop yield and maintain the fairness over space. Failures to preserve the spatial fairness in predicted maps of crop yields can result in biased policies and intervention strategies in the distribution of assistance or subsidies in supporting individuals at risk. Existing methods for fairness enforcement are not designed for capturing the complex physical processes that underlie the crop growing process, and thus are unable to produce good predictions over large regions under different weather conditions and soil properties. More importantly, the fairness is often degraded when existing methods are applied to different years due to the change of weather conditions and farming practices. To address these issues, we propose a physics-guided neural network model, which leverages the physical knowledge from existing physics-based models to guide the extraction of representative physical information and discover the temporal data shift across years. In particular, we use a reweighting strategy to discover the relationship between training years and testing years using the physics-aware representation. Then the physics-guided neural network will be refined via a bi-level optimization process based on the reweighted fairness objective. The proposed method has been evaluated using real county-level crop yield data and simulated data produced by a physics-based model. The results demonstrate that this method can significantly improve the predictive performance and preserve the spatial fairness when generalized to different years.

In this paper, we address the problem of video temporal sentence localization, which aims to localize a target moment from videos according to a given language query. We observe that existing models suffer from a sheer performance drop when dealing with simple phrases contained in the sentence. It reveals the limitation that existing models only capture the annotation bias of the datasets but lack sufficient understanding of the semantic phrases in the query. To address this problem, we propose a phrase-level Temporal Relationship Mining (TRM) framework employing the temporal relationship relevant to the phrase and the whole sentence to have a better understanding of each semantic entity in the sentence. Specifically, we use phrase-level predictions to refine the sentence-level prediction, and use Multiple Instance Learning to improve the quality of phrase-level predictions. We also exploit the consistency and exclusiveness constraints of phrase-level and sentence-level predictions to regularize the training process, thus alleviating the ambiguity of each phrase prediction. The proposed approach sheds light on how machines can understand detailed phrases in a sentence and their compositions in their generality rather than learning the annotation biases. Experiments on the ActivityNet Captions and Charades-STA datasets show the effectiveness of our method on both phrase and sentence temporal localization and enable better model interpretability and generalization when dealing with unseen compositions of seen concepts. Code can be found at https://github.com/minghangz/TRM.

The overarching goal of this work is to enable the collection of language describing a wide variety of objects viewed in virtual reality. We aim to create full 3D models from a small number of ‘keyframe’ images of objects found in the publicly available Grounded Language Dataset (GoLD) using photogrammetry. We will then collect linguistic descriptions by placing our models in virtual reality and having volunteers describe them. To evaluate the impact of virtual reality immersion on linguistic descriptions of the objects, we intend to apply contrastive learning to perform grounded language learning, then compare the descriptions collected from images (in GoLD) versus our models.

Despite hundreds of methods published in the literature, forecasting epidemic dynamics remains challenging yet important. The challenges stem from multiple sources, including: the need for timely data, co-evolution of epidemic dynamics with behavioral and immunological adaptations, and the evolution of new pathogen strains. The ongoing COVID-19 pandemic highlighted these challenges; in an important article, Reich et al. did a comprehensive analysis highlighting many of these challenges. In this paper, we take another step in critically evaluating existing epidemic forecasting methods. Our methods are based on a simple yet crucial observation - epidemic dynamics go through a number of phases (waves). Armed with this understanding, we propose a modification to our deployed Bayesian ensembling case time series forecasting framework. We show that ensembling methods employing the phase information and using different weighting schemes for each phase can produce improved forecasts. We evaluate our proposed method with both the currently deployed model and the COVID-19 forecasthub models. The overall performance of the proposed model is consistent across the pandemic but more importantly, it is ranked third and first during two critical rapid growth phases in cases, regimes where the performance of most models from the CDC forecasting hub dropped significantly.

Modeling what makes an advertisement persuasive, i.e., eliciting the desired response from consumer, is critical to the study of propaganda, social psychology, and marketing. Despite its importance, computational modeling of persuasion in computer vision is still in its infancy, primarily due to the lack of benchmark datasets that can provide persuasion-strategy labels associated with ads. Motivated by persuasion literature in social psychology and marketing, we introduce an extensive vocabulary of persuasion strategies and build the first ad image corpus annotated with persuasion strategies. We then formulate the task of persuasion strategy prediction with multi-modal learning, where we design a multi-task attention fusion model that can leverage other ad-understanding tasks to predict persuasion strategies. The dataset also provides image segmentation masks, which labels persuasion strategies in the corresponding ad images on the test split. We publicly release our code and dataset at https://midas-research.github.io/persuasion-advertisements/.

Persona-based dialogue systems aim to generate consistent responses based on historical context and predefined persona. Unlike conventional dialogue generation, the persona-based dialogue needs to consider both dialogue context and persona, posing a challenge for coherent training. Specifically, this requires a delicate weight balance between context and persona. To achieve that, in this paper, we propose an effective framework with Persona-Adaptive Attention (PAA), which adaptively integrates the weights from the persona and context information via our designed attention. In addition, a dynamic masking mechanism is applied to the PAA to not only drop redundant information in context and persona but also serve as a regularization mechanism to avoid overfitting. Experimental results demonstrate the superiority of the proposed PAA framework compared to the strong baselines in both automatic and human evaluation. Moreover, the proposed PAA approach can perform equivalently well in a low-resource regime compared to models trained in a full-data setting, which achieve a similar result with only 20% to 30% of data compared to the larger models trained in the full-data setting. To fully exploit the effectiveness of our design, we designed several variants for handling the weighted information in different ways, showing the necessity and sufficiency of our weighting and masking designs.

The efficacy of topological data analysis (TDA) has been demonstrated in many different machine learning pipelines, particularly those in which structural characteristics of data are highly relevant. However, TDA's usability in large scale machine learning applications is hindered by the significant computational cost of generating persistence diagrams. In this work, a method that allows this computationally expensive process to be approximated by deep neural networks is proposed. Moreover, the method's practicality in estimating 0-dimensional persistence diagrams across a diverse range of images is shown.

Multi-agent path planning (MAPP) is the problem of planning collision-free trajectories from start to goal locations for a team of agents. This work explores a relatively unexplored setting of MAPP where streams of agents have to go through the starts and goals with high throughput. We tackle this problem by formulating a new variant of MAPP called periodic MAPP in which the timing of agent appearances is periodic. The objective with periodic MAPP is to find a periodic plan, a set of collision-free trajectories that the agent streams can use repeatedly over periods, with periods that are as small as possible. To meet this objective, we propose a solution method that is based on constraint relaxation and optimization. We show that the periodic plans once found can be used for a more practical case in which agents in a stream can appear at random times. We confirm the effectiveness of our method compared with baseline methods in terms of throughput in several scenarios that abstract autonomous intersection management tasks.

In this work, we expand the discussion of bias in Automatic Speech Recognition (ASR) through a large-scale audit. Using a large and global data set of speech, we perform an audit of some of the most popular English ASR services. We show that, even when controlling for multiple linguistic covariates, ASR service performance has a statistically significant relationship to the political alignment of the speaker's birth country with respect to the United States' geopolitical power.

To autonomously perform tasks, a robot should continually perceive the state of its environment, reason with the task at hand, plan and execute appropriate actions. In this pipeline, perception is largely unsolved and one of the more challenging problems. Common indoor environments typically pose two main problems: 1) inherent occlusions leading to unreliable observations of objects, and 2) the presence and involvement of a wide range of objects with varying physical and visual attributes (i.e., rigid, articulated, deformable, granular, transparent, etc.). Thus, we need algorithms that can accommodate perceptual uncertainty in the state estimation and generalize to a wide range of objects. Probabilistic inference methods have been highly suitable for modeling perceptual uncertainty, and data-driven approaches using deep learning techniques have shown promising advancements toward generalization. Perception for manipulation is a more intricate setting requiring the best from both worlds. My research aims to develop robot perception algorithms that can generalize over objects and tasks while accommodating perceptual uncertainty to support robust task execution in the real world. In this presentation, I will briefly highlight my research in these two research threads.

The soaring number of personal mobile devices and public cameras poses a threat to fundamental human rights and ethical principles. For example, the stolen of private information such as face image by malicious third parties will lead to catastrophic consequences. By manipulating appearance of face in the image, most of existing protection algorithms are effective but irreversible. Here, we propose a practical and systematic solution to invertiblely protect face information in the full-process pipeline from camera to final users. Specifically, We design a novel lightweight Flow-based Face Encryption Method (FFEM) on the local embedded system privately connected to the camera, minimizing the risk of eavesdropping during data transmission. FFEM uses a flow-based face encoder to encode each face to a Gaussian distribution and encrypts the encoded face feature by random rotating the Gaussian distribution with the rotation matrix is as the password. While encrypted latent-variable face images are sent to users through public but less reliable channels, password will be protected through more secure channels through technologies such as asymmetric encryption, blockchain, or other sophisticated security schemes. User could select to decode an image with fake faces from the encrypted image on the public channel. Only trusted users are able to recover the original face using the encrypted matrix transmitted in secure channel. More interestingly, by tuning Gaussian ball in latent space, we could control the fairness of the replaced face on attributes such as gender and race. Extensive experiments demonstrate that our solution could protect privacy and enhance fairness with minimal effect on high-level downstream task.