AAAI 2023 论文列表

Language models (LMs) have demonstrated their capability in possessing commonsense knowledge of the physical world, a crucial aspect of performing tasks in everyday life. However, it remains unclear whether they have the capacity to generate grounded, executable plans for embodied tasks. This is a challenging task as LMs lack the ability to perceive the environment through vision and feedback from the physical environment. In this paper, we address this important research question and present the first investigation into the topic. Our novel problem formulation, named G-PlanET, inputs a high-level goal and a data table about objects in a specific environment, and then outputs a step-by-step actionable plan for a robotic agent to follow. To facilitate the study, we establish an evaluation protocol and design a dedicated metric, KAS, to assess the quality of the plans. Our experiments demonstrate that the use of tables for encoding the environment and an iterative decoding strategy can significantly enhance the LMs' ability in grounded planning. Our analysis also reveals interesting and non-trivial findings.

Conventional graph neural networks (GNNs) are often confronted with fairness issues that may stem from their input, including node attributes and neighbors surrounding a node. While several recent approaches have been proposed to eliminate the bias rooted in sensitive attributes, they ignore the other key input of GNNs, namely the neighbors of a node, which can introduce bias since GNNs hinge on neighborhood structures to generate node representations. In particular, the varying neighborhood structures across nodes, manifesting themselves in drastically different node degrees, give rise to the diverse behaviors of nodes and biased outcomes. In this paper, we first define and generalize the degree bias using a generalized definition of node degree as a manifestation and quantification of different multi-hop structures around different nodes. To address the bias in the context of node classification, we propose a novel GNN framework called Generalized Degree Fairness-centric Graph Neural Network (DegFairGNN). Specifically, in each GNN layer, we employ a learnable debiasing function to generate debiasing contexts, which modulate the layer-wise neighborhood aggregation to eliminate the degree bias originating from the diverse degrees among nodes. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our model on both accuracy and fairness metrics.

We study the close interplay between error and compression in the non-parametric multiclass classification setting in terms of prototype learning rules. We focus in particular on a recently proposed compression-based learning rule termed OptiNet. Beyond its computational merits, this rule has been recently shown to be universally consistent in any metric instance space that admits a universally consistent rule---the first learning algorithm known to enjoy this property. However, its error and compression rates have been left open. Here we derive such rates in the case where instances reside in Euclidean space under commonly posed smoothness and tail conditions on the data distribution. We first show that OptiNet achieves non-trivial compression rates while enjoying near minimax-optimal error rates. We then proceed to study a novel general compression scheme for further compressing prototype rules that locally adapts to the noise level without sacrificing accuracy. Applying it to OptiNet, we show that under a geometric margin condition further gain in the compression rate is achieved. Experimental results comparing the performance of the various methods are presented.

Multimodal relation extraction is an essential task for knowledge graph construction. In this paper, we take an in-depth empirical analysis that indicates the inaccurate information in the visual scene graph leads to poor modal alignment weights, further degrading performance. Moreover, the visual shuffle experiments illustrate that the current approaches may not take full advantage of visual information. Based on the above observation, we further propose a strong baseline with an implicit fine-grained multimodal alignment based on Transformer for multimodal relation extraction. Experimental results demonstrate the better performance of our method. Codes are available at https://github.com/zjunlp/DeepKE/tree/main/example/re/multimodal.

Recently, to reap the quantum advantage, empowering reinforcement learning (RL) with quantum computing has attracted much attention, which is dubbed as quantum RL (QRL). However, current QRL algorithms employ an online learning scheme, i.e., the policy that is run on a quantum computer needs to interact with the environment to collect experiences, which could be expensive and dangerous for practical applications. In this paper, we aim to solve this problem in an offline learning manner. To be more specific, we develop the first offline quantum RL (offline QRL) algorithm named CQ2L (Conservative Quantum Q-learning), which learns from offline samples and does not require any interaction with the environment. CQ2L utilizes variational quantum circuits (VQCs), which are improved with data re-uploading and scaling parameters, to represent Q-value functions of agents. To suppress the overestimation of Q-values resulting from offline data, we first employ a double Q-learning framework to reduce the overestimation bias; then a penalty term that encourages generating conservative Q-values is designed. We conduct abundant experiments to demonstrate that the proposed method CQ2L can successfully solve offline QRL tasks that the online counterpart could not.

Offline imitation learning (IL) promises the ability to learn performant policies from pre-collected demonstrations without interactions with the environment. However, imitating behaviors fully offline typically requires numerous expert data. To tackle this issue, we study the setting where we have limited expert data and supplementary suboptimal data. In this case, a well-known issue is the distribution shift between the learned policy and the behavior policy that collects the offline data. Prior works mitigate this issue by regularizing the KL divergence between the stationary state-action distributions of the learned policy and the behavior policy. We argue that such constraints based on exact distribution matching can be overly conservative and hamper policy learning, especially when the imperfect offline data is highly suboptimal. To resolve this issue, we present RelaxDICE, which employs an asymmetrically-relaxed f-divergence for explicit support regularization. Specifically, instead of driving the learned policy to exactly match the behavior policy, we impose little penalty whenever the density ratio between their stationary state-action distributions is upper bounded by a constant. Note that such formulation leads to a nested min-max optimization problem, which causes instability in practice. RelaxDICE addresses this challenge by supporting a closed-form solution for the inner maximization problem. Extensive empirical study shows that our method significantly outperforms the best prior offline IL method in six standard continuous control environments with over 30% performance gain on average, across 22 settings where the imperfect dataset is highly suboptimal.

Proximal Policy Optimization (PPO) is an important reinforcement learning method, which has achieved great success in sequential decision-making problems. However, PPO faces the issue of sample inefficiency, which is due to the PPO cannot make use of off-policy data. In this paper, we propose an Off-Policy Proximal Policy Optimization method (Off-Policy PPO) that improves the sample efficiency of PPO by utilizing off-policy data. Specifically, we first propose a clipped surrogate objective function that can utilize off-policy data and avoid excessively large policy updates. Next, we theoretically clarify the stability of the optimization process of the proposed surrogate objective by demonstrating the degree of policy update distance is consistent with that in the PPO. We then describe the implementation details of the proposed Off-Policy PPO which iteratively updates policies by optimizing the proposed clipped surrogate objective. Finally, the experimental results on representative continuous control tasks validate that our method outperforms the state-of-the-art methods on most tasks.

Point cloud compression with a higher compression ratio and tiny loss is essential for efficient data transportation. However, previous methods that depend on 3D convolution or frequent multi-head self-attention operations bring huge computations. To address this problem, we propose an octree-based Transformer compression method called OctFormer, which does not rely on the occupancy information of sibling nodes. Our method uses non-overlapped context windows to construct octree node sequences and share the result of a multi-head self-attention operation among a sequence of nodes. Besides, we introduce a locally-enhance module for exploiting the sibling features and a positional encoding generator for enhancing the translation invariance of the octree node sequence. Compared to the previous state-of-the-art works, our method obtains up to 17% Bpp savings compared to the voxel-context-based baseline and saves an overall 99% coding time compared to the attention-based baseline.

Single-view RGB-D human reconstruction with implicit functions is often formulated as per-point classification. Specifically, a set of 3D locations within the view-frustum of the camera are first projected independently onto the image and a corresponding feature is subsequently extracted for each 3D location. The feature of each 3D location is then used to classify independently whether the corresponding 3D point is inside or outside the observed object. This procedure leads to sub-optimal results because correlations between predictions for neighboring locations are only taken into account implicitly via the extracted features. For more accurate results we propose the occupancy planes (OPlanes) representation, which enables to formulate single-view RGB-D human reconstruction as occupancy prediction on planes which slice through the camera's view frustum. Such a representation provides more flexibility than voxel grids and enables to better leverage correlations than per-point classification. On the challenging S3D data we observe a simple classifier based on the OPlanes representation to yield compelling results, especially in difficult situations with partial occlusions due to other objects and partial visibility, which haven't been addressed by prior work.

Black-box optimization (BBO) algorithms are concerned with finding the best solutions for problems with missing analytical details. Most classical methods for such problems are based on strong and fixed a priori assumptions, such as Gaussianity. However, the complex real-world problems, especially when the global optimum is desired, could be very far from the a priori assumptions because of their diversities, causing unexpected obstacles. In this study, we propose a generative adversarial net-based broad-spectrum global optimizer (OPT-GAN) which estimates the distribution of optimum gradually, with strategies to balance exploration-exploitation trade-off. It has potential to better adapt to the regularity and structure of diversified landscapes than other methods with fixed prior, e.g., Gaussian assumption or separability. Experiments on diverse BBO benchmarks and high dimensional real world applications exhibit that OPT-GAN outperforms other traditional and neural net-based BBO algorithms. The code and Appendix are available at https://github.com/NBICLAB/OPT-GAN

Driving under the influence (DUI) is one of the main causes of traffic accidents, often leading to severe life and property losses. Setting up sobriety checkpoints on certain roads is the most commonly used practice to identify DUI-drivers in many countries worldwide. However, setting up checkpoints according to the police's experiences may not be effective for ignoring the strategic interactions between the police and DUI-drivers, particularly when inspecting resources are limited. To remedy this situation, we adapt the classic Stackelberg security game (SSG) to a new SSG-DUI game to describe the strategic interactions in catching DUI-drivers. SSG-DUI features drivers' bounded rationality and social knowledge sharing among them, thus realizing improved real-world fidelity. With SSG-DUI, we propose OPRADI, a systematic approach for advising better strategies in setting up checkpoints. We perform extensive experiments to evaluate it in both simulated environments and real-world contexts, in collaborating with a Chinese city's police bureau. The results reveal its effectiveness in improving police's real-world operations, thus having significant practical potentials.

To bridge the ever-increasing gap between deep neural networks' complexity and hardware capability, network quantization has attracted more and more research attention. The latest trend of mixed precision quantization takes advantage of hardware's multiple bit-width arithmetic operations to unleash the full potential of network quantization. However, existing approaches rely heavily on an extremely time-consuming search process and various relaxations when seeking the optimal bit configuration. To address this issue, we propose to optimize a proxy metric of network orthogonality that can be efficiently solved with linear programming, which proves to be highly correlated with quantized model accuracy and bit-width. Our approach significantly reduces the search time and the required data amount by orders of magnitude, but without a compromise on quantization accuracy. Specifically, we achieve 72.08% Top-1 accuracy on ResNet-18 with 6.7Mb parameters, which does not require any searching iterations. Given the high efficiency and low data dependency of our algorithm, we use it for the post-training quantization, which achieves 71.27% Top-1 accuracy on MobileNetV2 with only 1.5Mb parameters.

For the complicated input-output systems with nonlinearity and stochasticity, Deep State Space Models (SSMs) are effective for identifying systems in the latent state space, which are of great significance for representation, forecasting, and planning in online scenarios. However, most SSMs are designed for discrete-time sequences and inapplicable when the observations are irregular in time. To solve the problem, we propose a novel continuous-time SSM named Ordinary Differential Equation Recurrent State Space Model (ODE-RSSM). ODE-RSSM incorporates an ordinary differential equation (ODE) network (ODE-Net) to model the continuous-time evolution of latent states between adjacent time points. Inspired from the equivalent linear transformation on integration limits, we propose an efficient reparameterization method for solving batched ODEs with non-uniform time spans in parallel for efficiently training the ODE-RSSM with irregularly sampled sequences. We also conduct extensive experiments to evaluate the proposed ODE-RSSM and the baselines on three input-output datasets, one of which is a rollout of a private industrial dataset with strong long-term delay and stochasticity. The results demonstrate that the ODE-RSSM achieves better performance than other baselines in open loop prediction even if the time spans of predicted points are uneven and the distribution of length is changeable. Code is availiable at https://github.com/yuanzhaolin/ODE-RSSM.

Maximum Satisfiability (MaxSAT) is a prototypical constraint optimization problem, and its generalized version is the (Weighted) Partial MaxSAT problem, denoted as (W)PMS, which deals with hard and soft clauses. Considerable progress has been made on stochastic local search (SLS) algorithms for solving (W)PMS, which mainly focus on clause weighting techniques. In this work, we identify two issues of existing clause weighting techniques for (W)PMS, and propose two ideas correspondingly. First, we observe that the initial values of soft clause weights have a big effect on the performance of the SLS solver for solving (W)PMS, and propose a weight initialization method. Second, we propose a new clause weighting scheme that for the first time employs different conditions for updating hard and soft clause weights. Based on these two ideas, we develop a new SLS solver for (W)PMS named NuWLS. Through extensive experiments, NuWLS performs much better than existing SLS solvers on all 6 benchmarks from the incomplete tracks of MaxSAT Evaluations (MSEs) 2019, 2020, and 2021. In terms of the number of winning instances, NuWLS outperforms state-of-the-art SAT-based incomplete solvers on all the 6 benchmarks. More encouragingly, a hybrid solver that combines NuWLS and an SAT-based solver won all four categories in the incomplete track of the MaxSAT Evaluation 2022.

Citizens’ assemblies are groups of randomly selected constituents who are tasked with providing recommendations on policy questions. Assembly members form their recommendations through a sequence of discussions in small groups (deliberation), in which group members exchange arguments and experiences. We seek to support this process through optimization, by studying how to assign participants to discussion groups over multiple sessions, in a way that maximizes interaction between participants and satisfies diversity constraints within each group. Since repeated meetings between a given pair of participants have diminishing marginal returns, we capture interaction through a submodular function, which is approximately optimized by a greedy algorithm making calls to an ILP solver. This framework supports different submodular objective functions, and we identify sensible options, but we also show it is not necessary to commit to a particular choice: Our main theoretical result is a (practically efficient) algorithm that simultaneously approximates every possible objective function of the form we are interested in. Experiments with data from real citizens' assemblies demonstrate that our approach substantially outperforms the heuristic algorithm currently used by practitioners.

We propose ordering-based approaches for learning the maximal ancestral graph (MAG) of a structural equation model (SEM) up to its Markov equivalence class (MEC) in the presence of unobserved variables. Existing ordering-based methods in the literature recover a graph through learning a causal order (c-order). We advocate for a novel order called removable order (r-order) as they are advantageous over c-orders for structure learning. This is because r-orders are the minimizers of an appropriately defined optimization problem that could be either solved exactly (using a reinforcement learning approach) or approximately (using a hill-climbing search). Moreover, the r-orders (unlike c-orders) are invariant among all the graphs in a MEC and include c-orders as a subset. Given that set of r-orders is often significantly larger than the set of c-orders, it is easier for the optimization problem to find an r-order instead of a c-order. We evaluate the performance and the scalability of our proposed approaches on both real-world and randomly generated networks.

Most skeleton-based action recognition methods assume that the same type of action samples in the training set and the test set share similar motion patterns. However, action samples in real scenarios usually contain novel motion patterns which are not involved in the training set. As it is laborious to collect sufficient training samples to enumerate various types of novel motion patterns, this paper presents a practical skeleton-based action recognition task where the training set contains common motion patterns of action samples and the test set contains action samples that suffer from novel motion patterns. For this task, we present a Mask Graph Convolutional Network (Mask-GCN) to focus on learning action-specific skeleton joints that mainly convey action information meanwhile masking action-agnostic skeleton joints that convey rare action information and suffer more from novel motion patterns. Specifically, we design a policy network to learn layer-wise body masks to construct masked adjacency matrices, which guide a GCN-based backbone to learn stable yet informative action features from dynamic graph structure. Extensive experiments on our newly collected dataset verify that Mask-GCN outperforms most GCN-based methods when testing with various novel motion patterns.

Novel intent class detection is an important problem in real world scenario for conversational agents for continuous interaction. Several research works have been done to detect novel intents in a mono-lingual (primarily English) texts and images. But, current systems lack an end-to-end universal framework to detect novel intents across various different languages with less human annotation effort for mis-classified and system rejected samples. This paper proposes NIDAL (Novel Intent Detection and Active Learning based classification), a semi-supervised framework to detect novel intents while reducing human annotation cost. Empirical results on various benchmark datasets demonstrate that this system outperforms the baseline methods by more than 10% margin for accuracy and macro-F1. The system achieves this while maintaining overall annotation cost to be just ~6-10% of the unlabeled data available to the system.

Applying deep neural networks to 3D point cloud processing has demonstrated a rapid pace of advancement in those domains where 3D geometry information can greatly boost task performance, such as AR/VR, robotics, and autonomous driving. However, as the size of both the neural network model and 3D point cloud continues to scale, reducing the entailed computation and memory access overhead is a primary challenge to meet strict latency and energy constraints of practical applications. This paper proposes a new weight pruning technique for 3D point cloud based on spatial point distribution. We identify that particular groups of neighborhood voxels in 3D point cloud contribute more frequently to actual output features than others. Based on this observation, we propose to selectively prune less contributing groups of neighborhood voxels first to reduce the computation overhead while minimizing the impact on model accuracy. We apply our proposal to three representative sparse 3D convolution libraries. Our proposal reduces the inference latency by 1.60× on average and energy consumption by 1.74× on NVIDIA GV100 GPU with no loss in accuracy metric