AAAI 2024 Papers

Cost-guided bottom-up search (BUS) algorithms use a cost function to guide the search to solve program synthesis tasks. In this paper, we show that current state-of-the-art cost-guided BUS algorithms suffer from a common problem: they can lose useful information given by the model and fail to perform the search in a best-first order according to a cost function. We introduce a novel best-first bottom-up search algorithm, which we call Bee Search, that does not suffer information loss and is able to perform cost-guided bottom-up synthesis in a best-first manner. Importantly, Bee Search performs best-first search with respect to the generation of programs, i.e., it does not even create in memory programs that are more expensive than the solution program. It attains best-first ordering with respect to generation by performing a search in an abstract space of program costs. We also introduce a new cost function that better uses the information provided by an existing cost model. Empirical results on string manipulation and bit-vector tasks show that Bee Search can outperform existing cost-guided BUS approaches when employing more complex domain-specific languages (DSLs); Bee Search and previous approaches perform equally well with simpler DSLs. Furthermore, our new cost function with Bee Search outperforms previous cost functions on string manipulation tasks.

Level generation is a central focus of Procedural Content Generation (PCG), yet deep learning-based approaches are limited by scarce training data, i.e., human-designed levels. Despite being a dominant framework, Generative Adversarial Networks (GANs) exhibit a substantial quality gap between generated and human-authored levels, alongside rising training costs, particularly with increasing token complexity. In this paper, we introduce a diffusion-based generative model that learns from just one example. Our approach involves two core components: 1) an efficient yet expressive level representation, and 2) a latent denoising network with constrained receptive fields. To start with, our method utilizes token semantic labels, similar to word embeddings, to provide dense representations. This strategy not only surpasses one-hot encoding in representing larger game levels but also improves stability and accelerates convergence in latent diffusion. In addition, we adapt the denoising network architecture to confine the receptive field to localized patches of the data, aiming to facilitate single-example learning. Extensive experiments demonstrate that our model is capable of generating stylistically congruent samples of arbitrary sizes compared to manually designed levels. It suits a wide range of level structures with fewer artifacts than GAN-based approaches. The source code is available at https://github.com/shiqi-dai/diffusioncraft.

Optimal transport (OT) is an important methodology to measure distribution discrepancy, which has achieved promising performance in artificial intelligence applications, e.g., unsupervised domain adaptation. However, from the view of transportation, there are still limitations: 1) the local discriminative structures for downstream tasks, e.g., cluster structure for classification, cannot be explicitly admitted by the learned OT plan; 2) the entropy regularization induces a dense OT plan with increasing uncertainty. To tackle these issues, we propose a novel Probability-Polarized OT (PPOT) framework, which can characterize the structure of OT plan explicitly. Specifically, the probability polarization mechanism is proposed to guide the optimization direction of OT plan, which generates a clear margin between similar and dissimilar transport pairs and reduces the uncertainty. Further, a dynamic mechanism for margin is developed by incorporating task-related information into the polarization, which directly captures the intra/inter class correspondence for knowledge transportation. A mathematical understanding for PPOT is provided from the view of gradient, which ensures interpretability. Extensive experiments on several datasets validate the effectiveness and empirical efficiency of PPOT.

Probabilities of causation are proven to be critical in modern decision-making. This paper deals with the problem of estimating the probabilities of causation when treatment and effect are not binary. Pearl defined the binary probabilities of causation, such as the probability of necessity and sufficiency (PNS), the probability of sufficiency (PS), and the probability of necessity (PN). Tian and Pearl then derived sharp bounds for these probabilities of causation using experimental and observational data. In this paper, we define and provide theoretical bounds for all types of probabilities of causation with multivalued treatments and effects. We further discuss examples where our bounds guide practical decisions and use simulation studies to evaluate how informative the bounds are for various data combinations.

In practice, it is essential to compare and rank candidate policies offline before real-world deployment for safety and reliability. Prior work seeks to solve this offline policy ranking (OPR) problem through value-based methods, such as Off-policy evaluation (OPE). However, they fail to analyze special case performance (e.g., worst or best cases), due to the lack of holistic characterization of policies’ performance. It is even more difficult to estimate precise policy values when the reward is not fully accessible under sparse settings. In this paper, we present Probabilistic Offline Policy Ranking (POPR), a framework to address OPR problems by leveraging expert data to characterize the probability of a candidate policy behaving like experts, and approximating its entire performance posterior distribution to help with ranking. POPR does not rely on value estimation, and the derived performance posterior can be used to distinguish candidates in worst-, best-, and average-cases. To estimate the posterior, we propose POPR-EABC, an Energy-based Approximate Bayesian Computation (ABC) method conducting likelihood-free inference. POPR-EABC reduces the heuristic nature of ABC by a smooth energy function, and improves the sampling efficiency by a pseudo-likelihood. We empirically demonstrate that POPR-EABC is adequate for evaluating policies in both discrete and continuous action spaces across various experiment environments, and facilitates probabilistic comparisons of candidate policies before deployment.

Probabilistic circuits (PCs) have gained prominence in recent years as a versatile framework for discussing probabilistic models that support tractable queries and are yet expressive enough to model complex probability distributions. Nevertheless, tractability comes at a cost: PCs are less expressive than neural networks. In this paper we introduce probabilistic neural circuits (PNCs), which strike a balance between PCs and neural nets in terms of tractability and expressive power. Theoretically, we show that PNCs can be interpreted as deep mixtures of Bayesian networks. Experimentally, we demonstrate that PNCs constitute powerful function approximators.

Open-World Compositional Zero-shot Learning (OW-CZSL) aims to recognize novel compositions of state and object primitives in images with no priors on the compositional space, which induces a tremendously large output space containing all possible state-object compositions. Existing works either learn the joint compositional state-object embedding or predict simple primitives with separate classifiers. However, the former method heavily relies on external word embedding methods, and the latter ignores the interactions of interdependent primitives, respectively. In this paper, we revisit the primitive prediction approach and propose a novel method, termed Progressive Cross-primitive Compatibility (ProCC), to mimic the human learning process for OW-CZSL tasks. Specifically, the cross-primitive compatibility module explicitly learns to model the interactions of state and object features with the trainable memory units, which efficiently acquires cross-primitive visual attention to reason high-feasibility compositions, without the aid of external knowledge. Moreover, to alleviate the invalid cross-primitive interactions, especially for partial-supervision conditions (pCZSL), we design a progressive training paradigm to optimize the primitive classifiers conditioned on pre-trained features in an easy-to-hard manner. Extensive experiments on three widely used benchmark datasets demonstrate that our method outperforms other representative methods on both OW-CZSL and pCZSL settings by large margins.

Building agents with adaptive behavior in cooperative tasks stands as a paramount goal in the realm of multi-agent systems. Current approaches to developing cooperative agents rely primarily on learning-based methods, whose policy generalization depends heavily on the diversity of teammates they interact with during the training phase. Such reliance, however, constrains the agents' capacity for strategic adaptation when cooperating with unfamiliar teammates, which becomes a significant challenge in zero-shot coordination scenarios. To address this challenge, we propose ProAgent, a novel framework that harnesses large language models (LLMs) to create proactive agents capable of dynamically adapting their behavior to enhance cooperation with teammates. ProAgent can analyze the present state, and infer the intentions of teammates from observations. It then updates its beliefs in alignment with the teammates' subsequent actual behaviors. Moreover, ProAgent exhibits a high degree of modularity and interpretability, making it easily integrated into various of coordination scenarios. Experimental evaluations conducted within the Overcooked-AI environment unveil the remarkable performance superiority of ProAgent, outperforming five methods based on self-play and population-based training when cooperating with AI agents. Furthermore, in partnered with human proxy models, its performance exhibits an average improvement exceeding 10% compared to the current state-of-the-art method. For more information about our project, please visit https://pku-proagent.github.io.

Performance of trimap-free image matting methods is limited when trying to decouple the deterministic and undetermined regions, especially in the scenes where foregrounds are semantically ambiguous, chromaless, or high transmittance. In this paper, we propose a novel framework named Privileged Prior Information Distillation for Image Matting (PPID-IM) that can effectively transfer privileged prior environment-aware information to improve the performance of trimap-free students in solving hard foregrounds. The prior information of trimap regulates only the teacher model during the training stage, while not being fed into the student network during actual inference. To achieve effective privileged cross-modality (i.e. trimap and RGB) information distillation, we introduce a Cross-Level Semantic Distillation (CLSD) module that reinforces the students with more knowledgeable semantic representations and environment-aware information. We also propose an Attention-Guided Local Distillation module that efficiently transfers privileged local attributes from the trimap-based teacher to trimap-free students for the guidance of local-region optimization. Extensive experiments demonstrate the effectiveness and superiority of our PPID on image matting. The code will be released soon.

We examine a private ADMM variant for (strongly) convex objectives which is a primal-dual iterative method. Each iteration has a user with a private function used to update the primal variable, masked by Gaussian noise for local privacy, without directly adding noise to the dual variable. Privacy amplification by iteration explores if noises from later iterations can enhance the privacy guarantee when releasing final variables after the last iteration. Cyffers et al. explored privacy amplification by iteration for the proximal ADMM variant, where a user's entire private function is accessed and noise is added to the primal variable. In contrast, we examine a private ADMM variant requiring just one gradient access to a user's function, but both primal and dual variables must be passed between successive iterations. To apply Balle et al.'s coupling framework to the gradient ADMM variant, we tackle technical challenges with novel ideas. First, we address the non-expansive mapping issue in ADMM iterations by using a customized norm. Second, because the dual variables are not masked with any noise directly, their privacy guarantees are achieved by treating two consecutive noisy ADMM iterations as a Markov operator. Our main result is that the privacy guarantee for the gradient ADMM variant can be amplified proportionally to the number of iterations. For strongly convex objective functions, this amplification exponentially increases with the number of iterations. These amplification results align with the previously studied special case of stochastic gradient descent.

Defocus blur, due to spatially-varying sizes and shapes, is hard to remove. Existing methods either are unable to effectively handle irregular defocus blur or fail to generalize well on other datasets. In this work, we propose a divide-and-conquer approach to tackling this issue, which gives rise to a novel end-to-end deep learning method, called prior-and-prediction inverse kernel transformer (P2IKT), for single image defocus deblurring. Since most defocus blur can be approximated as Gaussian blur or its variants, we construct an inverse Gaussian kernel module in our method to enhance its generalization ability. At the same time, an inverse kernel prediction module is introduced in order to flexibly address the irregular blur that cannot be approximated by Gaussian blur. We further design a scale recurrent transformer, which estimates mixing coefficients for adaptively combining the results from the two modules and runs the scale recurrent ``coarse-to-fine" procedure for progressive defocus deblurring. Extensive experimental results demonstrate that our P2IKT outperforms previous methods in terms of PSNR on multiple defocus deblurring datasets.

Low dimensional models like PCA are often used to simplify complex datasets by learning a single approximating subspace. This paradigm has expanded to union of subspaces models, like those learned by subspace clustering. In this paper, we present Principal Component Trees (PCTs), a graph structure that generalizes these ideas to identify mixtures of components that together describe the subspace structure of high-dimensional datasets. Each node in a PCT corresponds to a principal component of the data, and the edges between nodes indicate the components that must be mixed to produce a subspace that approximates a portion of the data. In order to construct PCTs, we propose two angle-distribution hypothesis tests to detect subspace clusters in the data. To analyze, compare, and select the best PCT model, we define two persistent homology measures that describe their shape. We show our construction yields two key properties of PCTs, namely ancestral orthogonality and non-decreasing singular values. Our main theoretical results show that learning PCTs reduces to PCA under multivariate normality, and that PCTs are efficient parameterizations of intersecting union of subspaces. Finally, we use PCTs to analyze neural network latent space, word embeddings, and reference image datasets.

Principal-agent problems arise when one party acts on behalf of another, leading to conflicts of interest. The economic literature has extensively studied principal-agent problems, and recent work has extended this to more complex scenarios such as Markov Decision Processes (MDPs). In this paper, we further explore this line of research by investigating how reward shaping under budget constraints can improve the principal's utility. We study a two-player Stackelberg game where the principal and the agent have different reward functions, and the agent chooses an MDP policy for both players. The principal offers an additional reward to the agent, and the agent picks their policy selfishly to maximize their reward, which is the sum of the original and the offered reward. Our results establish the NP-hardness of the problem and offer polynomial approximation algorithms for two classes of instances: Stochastic trees and deterministic decision processes with a finite horizon.

Embedding Human and Articulated Object Interaction (HAOI) in 3D is an important direction for a deeper human activity understanding. Different from previous works that use parametric and CAD models to represent humans and objects, in this work, we propose a novel 3D geometric primitive-based language to encode both humans and objects. Given our new paradigm, humans and objects are all compositions of primitives instead of heterogeneous entities. Thus, mutual information learning may be achieved between the limited 3D data of humans and different object categories. Moreover, considering the simplicity of the expression and the richness of the information it contains, we choose the superquadric as the primitive representation. To explore an effective embedding of HAOI for the machine, we build a new benchmark on 3D HAOI consisting of primitives together with their images and propose a task requiring machines to recover 3D HAOI using primitives from images. Moreover, we propose a baseline of single-view 3D reconstruction on HAOI. We believe this primitive-based 3D HAOI representation would pave the way for 3D HAOI studies. Our code and data are available at https://mvig-rhos.com/p3haoi.

Rental assistance programs provide individuals with financial assistance to prevent housing instabilities caused by evictions and avert homelessness. Since these programs operate under resource constraints, they must decide who to prioritize. Typically, funding is distributed by a reactive allocation process that does not systematically consider risk of future homelessness. We partnered with Anonymous County (PA) to explore a proactive and preventative allocation approach that prioritizes individuals facing eviction based on their risk of future homelessness. Our ML models, trained on state and county administrative data accurately identify at-risk individuals, outperforming simpler prioritization approaches by at least 20% while meeting our equity and fairness goals across race and gender. Furthermore, our approach would reach 28% of individuals who are overlooked by the current process and end up homeless. Beyond improvements to the rental assistance program in Anonymous County, this study can inform the development of evidence-based decision support tools in similar contexts, including lessons about data needs, model design, evaluation, and field validation.

The rapid progress of Transformers in artificial intelligence has come at the cost of increased resource consumption and greenhouse gas emissions due to growing model sizes. Prior work suggests using pretrained small models to improve training efficiency, but this approach may not be suitable for new model structures. On the other hand, training from scratch can be slow, and progressively stacking layers often fails to achieve significant acceleration. To address these challenges, we propose a novel method called Apollo, which prepares lessons for expanding operations by learning high-layer functionality during training of low layers. Our approach involves low-value-prioritized sampling (LVPS) to train different depths and weight sharing to facilitate efficient expansion. We also introduce an interpolation method for stable model depth extension. Experiments demonstrate that Apollo achieves state-of-the-art acceleration ratios, even rivaling methods using pretrained models, making it a universal and efficient solution for training deep models while reducing time, financial, and environmental costs.

We consider the problem of constrained multi-objective optimization over black-box objectives, with user-defined preferences, with a largely infeasible input space. Our goal is to approximate the optimal Pareto set from the small fraction of feasible inputs. The main challenges include huge design space, multiple objectives, numerous constraints, and rare feasible inputs identified only through expensive experiments. We present PAC-MOO, a novel preference-aware multi-objective Bayesian optimization algorithm to solve this problem. It leverages surrogate models for objectives and constraints to intelligently select the sequence of inputs for evaluation to achieve the target goal.

Large language models (LLMs) often contain misleading content, emphasizing the need to align them with human values to ensure secure AI systems. Reinforcement learning from human feedback (RLHF) has been employed to achieve this alignment. However, it encompasses two main drawbacks: (1) RLHF exhibits complexity, instability, and sensitivity to hyperparameters in contrast to SFT. (2) Despite massive trial-and-error, multiple sampling is reduced to pair-wise contrast, thus lacking contrasts from a macro perspective. In this paper, we propose Preference Ranking Optimization (PRO) as an efficient SFT algorithm to directly fine-tune LLMs for human alignment. PRO extends the pair-wise contrast to accommodate preference rankings of any length. By iteratively contrasting candidates, PRO instructs the LLM to prioritize the best response while progressively ranking the rest responses. In this manner, PRO effectively transforms human alignment into aligning the probability ranking of n responses generated by LLM with the preference ranking of humans towards these responses. Experiments have shown that PRO outperforms baseline algorithms, achieving comparable results to ChatGPT and human responses through automatic-based, reward-based, GPT-4, and human evaluations.

Existing cold-start recommendation methods often adopt item-level alignment strategies to align the content feature and the collaborative feature of warm items for model training, however, cold items in the test stage have no historical interactions with users to obtain the collaborative feature. These existing models ignore the aforementioned condition of cold items in the training stage, resulting in the performance limitation. In this paper, we propose a preference aware dual contrastive learning based recommendation model (PAD-CLRec), where the user preference is explored to take into account the condition of cold items for feature alignment. Here, the user preference is obtained by aggregating a group of collaborative feature of the warm items in the user's purchase records. Then, a group-level alignment between the user preference and the item's content feature can be realized via a proposed preference aware contrastive function for enhancing cold-item recommendation. In addition, a joint objective function is introduced to achieve a better trade-off between the recommendation performance of warm items and cold items from both item-level and group-level perspectives, yielding better overall recommendation performance. Extensive experiments are conducted to demonstrate the effectiveness of the proposed method, and the results show the superiority of our method, as compared with the state-of-the-arts.