AAAI 2024 Papers

In this work, we design an Artificially Intelligent Task Allocator (AITA) that proposes a task allocation for a team of humans. A key property of this allocation is that when an agent with imperfect knowledge (about their teammate's costs and/or the team's performance metric) contests the allocation with a counterfactual, a contrastive explanation can always be provided to showcase why the proposed allocation is better than the proposed counterfactual. For this, we consider a negotiation process that produces a negotiation-aware task allocation and, when contested, leverages a negotiation tree to provide a contrastive explanation. With human subject studies, we show that the proposed allocation indeed appears fair to a majority of participants and, when not, the explanations generated are judged as convincing and easy to comprehend.

The recent advancements in Deep Reinforcement Learning (DRL) have significantly enhanced the performance of adaptive Traffic Signal Control (TSC). However, DRL policies are typically represented by neural networks, which are over-parameterized black-box models. As a result, the learned policies often lack interpretability, and cannot be deployed directly in the real-world edge hardware due to resource constraints. In addition, the DRL methods often exhibit limited generalization performance, struggling to generalize the learned policy to other geographical regions. These factors limit the practical application of learning-based approaches. To address these issues, we suggest the use of an inherently interpretable program for representing the control policy. We present a new approach, Programmatic Interpretable reinforcement learning for traffic signal control (π-light), designed to autonomously discover non-differentiable programs. Specifically, we define a Domain Specific Language (DSL) and transformation rules for constructing programs, and utilize Monte Carlo Tree Search (MCTS) to find the optimal program in a discrete space. Extensive experiments demonstrate that our method consistently outperforms baseline approaches. Moreover, π-Light exhibits superior generalization capabilities compared to DRL, enabling training and evaluation across intersections from different cities. Finally, we analyze how the learned program policies can directly deploy on edge devices with extremely limited resources.

Federated Learning (FL) is a promising privacy-preserving distributed learning paradigm but suffers from high communi- cation cost when training large-scale machine learning models. Sign-based methods, such as SignSGD, have been proposed as a biased gradient compression technique for reducing the communication cost. However, sign-based algorithms could diverge under heterogeneous data, which thus motivated the de- velopment of advanced techniques, such as the error-feedback method and stochastic sign-based compression, to fix this issue. Nevertheless, these methods still suffer from slower convergence rates, and none of them allows multiple local SGD updates like FedAvg. In this paper, we propose a novel noisy perturbation scheme with a general symmetric noise distribution for sign-based compression, which not only al- lows one to flexibly control the bias-variance tradeoff for the compressed gradient, but also provides a unified viewpoint to existing stochastic sign-based methods. More importantly, the proposed scheme enables the development of the very first sign-based FedAvg algorithm (z-SignFedAvg) to accelerate the convergence. Theoretically, we show that z-SignFedAvg achieves a faster convergence rate than existing sign-based methods and, under the uniformly distributed noise, can enjoy the same convergence rate as its uncompressed counterpart. Extensive experiments are conducted to demonstrate that the z-SignFedAvg can achieve competitive empirical performance on real datasets and outperforms existing schemes.

Causal inference in a sub-population involves identifying the causal effect of an intervention on a specific subgroup, which is distinguished from the whole population through the influence of systematic biases in the sampling process. However, ignoring the subtleties introduced by sub-populations can either lead to erroneous inference or limit the applicability of existing methods. We introduce and advocate for a causal inference problem in sub-populations (henceforth called s-ID), in which we merely have access to observational data of the targeted sub-population (as opposed to the entire population). Existing inference problems in sub-populations operate on the premise that the given data distributions originate from the entire population, thus, cannot tackle the s-ID problem. To address this gap, we provide necessary and sufficient conditions that must hold in the causal graph for a causal effect in a sub-population to be identifiable from the observational distribution of that sub-population. Given these conditions, we present a sound and complete algorithm for the s-ID problem.

When compressing point clouds, point-based deep learning models operate points in a continuous space, which has a chance to minimize the geometric fidelity loss introduced by voxelization in preprocessing. But these methods could hardly scale to inputs with arbitrary points. Furthermore, the point cloud frames are individually compressed, failing the conventional wisdom of leveraging inter-frame similarity. In this work, we propose a patchwise compression framework called patchDPCC, which consists of a patch group generation module and a point-based compression model. Algorithms are developed to generate patches from different frames representing the same object, and more importantly, these patches are regulated to have the same number of points. We also incorporate a feature transfer module in the compression model, which refines the feature quality by exploiting the inter-frame similarity. Our model generates point-wise features for entropy coding, which guarantees the reconstruction speed. The evaluation on the MPEG 8i dataset shows that our method improves the compression ratio by 47.01% and 85.22% when compared to PCGCv2 and V-PCC with the same reconstruction quality, which is 9% and 16% better than that D-DPCC does. Our method also achieves the fastest decoding speed among the learning-based compression models.

Vision-Language models (VLMs) pre-trained on large corpora have demonstrated notable success across a range of downstream tasks. In light of the rapidly increasing size of pre-trained VLMs, parameter-efficient transfer learning (PETL) has garnered attention as a viable alternative to full fine-tuning. One such approach is the adapter, which introduces a few trainable parameters into the pre-trained models while preserving the original parameters during adaptation. In this paper, we present a novel modeling framework that recasts adapter tuning after attention as a graph message passing process on attention graphs, where the projected query and value features and attention matrix constitute the node features and the graph adjacency matrix, respectively. Within this framework, tuning adapters in VLMs necessitates handling heterophilic graphs, owing to the disparity between the projected query and value space. To address this challenge, we propose a new adapter architecture, p-adapter, which employs p-Laplacian message passing in Graph Neural Networks (GNNs). Specifically, the attention weights are re-normalized based on the features, and the features are then aggregated using the calibrated attention matrix, enabling the dynamic exploitation of information with varying frequencies in the heterophilic attention graphs. We conduct extensive experiments on different pre-trained VLMs and multi-modal tasks, including visual question answering, visual entailment, and image captioning. The experimental results validate our method's significant superiority over other PETL methods. Our code is available at https://github.com/wuhy68/p-Adapter/.

LiDAR sensors are widely used in autonomous driving, and the growing storage and transmission demands have made LiDAR point cloud compression (LPCC) a hot research topic. To address the challenges posed by the large-scale and uneven-distribution (spatial and categorical) of LiDAR point data, this paper presents a new multimodal-driven scalable LPCC framework. For the large-scale challenge, we decouple the original LiDAR data into multi-layer point subsets, compress and transmit each layer separately, so as to ensure the reconstruction quality requirement under different scenarios. For the uneven-distribution challenge, we extract, align, and fuse heterologous feature representations, including point modality with position information, depth modality with spatial distance information, and segmentation modality with category information. Extensive experimental results on the benchmark SemanticKITTI database validate that our method outperforms 14 recent representative LPCC methods.

Pre-trained language models have shown their high performance of text processing in intelligent customer service platforms. However, these models do not leverage domain specific information. In this paper, we propose icsPLMs optimized for intelligent customer service on both word and sentence levels. Our experimental results represent that using targeted strategies can further improve the performance of pre-trained language models in this field.

Accurate prediction of meteorological elements, such as temperature and relative humidity, is important to human livelihood, early warning of extreme weather, and urban governance. Recently, neural network-based methods have shown impressive performance in this field. However, most of them are overcomplicated and impenetrable. In this paper, we propose a straightforward and interpretable differential framework, where the key lies in explicitly estimating the evolutionary trends. Specifically, three types of trends are exploited. (1) The proximity trend simply uses the most recent changes. It works well for approximately linear evolution. (2) The sequential trend explores the global information, aiming to capture the nonlinear dynamics. Here, we develop an attention-based trend unit to help memorize long-term features. (3) The flow trend is motivated by the nature of evolution, i.e., the heat or substance flows from one region to another. Here, we design a flow-aware attention unit. It can reflect the interactions via performing spatial attention over flow maps. Finally, we develop a trend fusion module to adaptively fuse the above three trends. Extensive experiments on two datasets demonstrate the effectiveness of our method.

Accurately annotating multiple 3D objects in LiDAR scenes is laborious and challenging. While a few previous studies have attempted to leverage semi-automatic methods for cost-effective bounding box annotation, such methods have limitations in efficiently handling numerous multi-class objects. To effectively accelerate 3D annotation pipelines, we propose iDet3D, an efficient interactive 3D object detector. Supporting a user-friendly 2D interface, which can ease the cognitive burden of exploring 3D space to provide click interactions, iDet3D enables users to annotate the entire objects in each scene with minimal interactions. Taking the sparse nature of 3D point clouds into account, we design a negative click simulation (NCS) to improve accuracy by reducing false-positive predictions. In addition, iDet3D incorporates two click propagation techniques to take full advantage of user interactions: (1) dense click guidance (DCG) for keeping user-provided information throughout the network and (2) spatial click propagation (SCP) for detecting other instances of the same class based on the user-specified objects. Through our extensive experiments, we present that our method can construct precise annotations in a few clicks, which shows the practicality as an efficient annotation tool for 3D object detection.

Ride-hailing platforms have been facing the challenge of balancing demand and supply. Existing vehicle reposition techniques often treat drivers as homogeneous agents and relocate them deterministically, assuming compliance with the reposition. In this paper, we consider a more realistic and driver-centric scenario where drivers have unique cruising preferences and can decide whether to take the recommendation or not on their own. We propose i-Rebalance, a personalized vehicle reposition technique with deep reinforcement learning (DRL). i-Rebalance estimates drivers' decisions on accepting reposition recommendations through an on-field user study involving 99 real drivers. To optimize supply-demand balance and enhance preference satisfaction simultaneously, i-Rebalance has a sequential reposition strategy with dual DRL agents: Grid Agent to determine the reposition order of idle vehicles, and Vehicle Agent to provide personalized recommendations to each vehicle in the pre-defined order. This sequential learning strategy facilitates more effective policy training within a smaller action space compared to traditional joint-action methods. Evaluation of real-world trajectory data shows that i-Rebalance improves driver acceptance rate by 38.07% and total driver income by 9.97%.

Class incremental learning (CIL) aims to solve the notorious forgetting problem, which refers to the fact that once the network is updated on a new task, its performance on previously-learned tasks degenerates catastrophically. Most successful CIL methods store exemplars (samples of learned tasks) to train a feature extractor incrementally, or store prototypes (features of learned tasks) to estimate the incremental feature distribution. However, the stored exemplars would violate the data privacy concerns, while the fixed prototypes might not reasonably be consistent with the incremental feature distribution, hindering the exploration of real-world CIL applications. In this paper, we propose a data-free CIL method with embedding distillation and Task-oriented generation (eTag), which requires neither exemplar nor prototype. Embedding distillation prevents the feature extractor from forgetting by distilling the outputs from the networks' intermediate blocks. Task-oriented generation enables a lightweight generator to produce dynamic features, fitting the needs of the top incremental classifier. Experimental results confirm that the proposed eTag considerably outperforms state-of-the-art methods on several benchmark datasets.

Recent advances in Large Language Models (LLMs) have achieved remarkable breakthroughs in understanding and responding to user intents. However, their performance lag behind general use cases in some expertise domains, such as Chinese medicine. Existing efforts to incorporate Chinese medicine into LLMs rely on Supervised Fine-Tuning (SFT) with single-turn and distilled dialogue data. These models lack the ability for doctor-like proactive inquiry and multi-turn comprehension and cannot align responses with experts' intentions. In this work, we introduce Zhongjing, the first Chinese medical LLaMA-based LLM that implements an entire training pipeline from continuous pre-training, SFT, to Reinforcement Learning from Human Feedback (RLHF). Additionally, we construct a Chinese multi-turn medical dialogue dataset of 70,000 authentic doctor-patient dialogues, CMtMedQA, which significantly enhances the model's capability for complex dialogue and proactive inquiry initiation. We also define a refined annotation rule and evaluation criteria given the unique characteristics of the biomedical domain. Extensive experimental results show that Zhongjing outperforms baselines in various capacities and matches the performance of ChatGPT in some abilities, despite the 100x parameters. Ablation studies also demonstrate the contributions of each component: pre-training enhances medical knowledge, and RLHF further improves instruction-following ability and safety. Our code, datasets, and models are available at https://github.com/SupritYoung/Zhongjing.

This work studies the behaviors of two large-population teams competing in a discrete environment. The team-level interactions are modeled as a zero-sum game while the agent dynamics within each team is formulated as a collaborative mean-field team problem. Drawing inspiration from the mean-field literature, we first approximate the large-population team game with its infinite-population limit. Subsequently, we construct a fictitious centralized system and transform the infinite-population game to an equivalent zero-sum game between two coordinators. Via a novel reachability analysis, we study the optimality of coordination strategies, which induce decentralized strategies under the original information structure. The optimality of the resulting strategies is established in the original finite-population game, and the theoretical guarantees are verified by numerical examples.

We propose a method to learn prediction models such as classifiers for unseen target tasks where labeled and unlabeled data are absent but a few relevant input features for solving the tasks are given. Although machine learning requires data for training, data are often difficult to collect in practice. On the other hand, for many applications, a few relevant features would be more easily obtained. Although zero-shot learning or zero-shot domain adaptation use external knowledge to adapt to unseen classes or tasks without data, relevant features have not been used in existing studies. The proposed method improves the generalization performance on the target tasks, where there are no data but a few relevant features are given, by meta-learning from labeled data in related tasks. In the meta-learning phase, it is essential to simulate test phases on target tasks where prediction model learning is required without data. To this end, our neural network-based prediction model is meta-learned such that it correctly responds to perturbations of the relevant features on randomly generated synthetic data. By this modeling, the prediction model can explicitly learn the discriminability of the relevant features without real target data. When unlabeled training data are available in the target tasks, the proposed method can incorporate such data to boost the performance in a unified framework. Our experiments demonstrate that the proposed method outperforms various existing methods with four real-world datasets.

Existing object detection models are mainly trained on large-scale labeled datasets. However, annotating data for novel aerial object classes is expensive since it is time-consuming and may require expert knowledge. Thus, it is desirable to study label-efficient object detection methods on aerial images. In this work, we propose a zero-shot method for aerial object detection named visual Description Regularization, or DescReg. Concretely, we identify the weak semantic-visual correlation of the aerial objects and aim to address the challenge with prior descriptions of their visual appearance. Instead of directly encoding the descriptions into class embedding space which suffers from the representation gap problem, we propose to infuse the prior inter-class visual similarity conveyed in the descriptions into the embedding learning. The infusion process is accomplished with a newly designed similarity-aware triplet loss which incorporates structured regularization on the representation space. We conduct extensive experiments with three challenging aerial object detection datasets, including DIOR, xView, and DOTA. The results demonstrate that DescReg significantly outperforms the state-of-the-art ZSD methods with complex projection designs and generative frameworks, e.g., DescReg outperforms best reported ZSD method on DIOR by 4.5 mAP on unseen classes and 8.1 in HM. We further show the generalizability of DescReg by integrating it into generative ZSD methods as well as varying the detection architecture. Codes will be released at https://github.com/zq-zang/DescReg.

Cognitive diagnosis seeks to estimate the cognitive states of students by exploring their logged practice quiz data. It plays a pivotal role in personalized learning guidance within intelligent education systems. In this paper, we focus on an important, practical, yet often underexplored task: domain-level zero-shot cognitive diagnosis (DZCD), which arises due to the absence of student practice logs in newly launched domains. Recent cross-domain diagnostic models have been demonstrated to be a promising strategy for DZCD. These methods primarily focus on how to transfer student states across domains. However, they might inadvertently incorporate non-transferable information into student representations, thereby limiting the efficacy of knowledge transfer. To tackle this, we propose Zero-1-to-3, a domain-level zero-shot cognitive diagnosis framework via one batch of early-bird students towards three diagnostic objectives. Our approach initiates with pre-training a diagnosis model with dual regularizers, which decouples student states into domain-shared and domain-specific parts. The shared cognitive signals can be transferred to the target domain, enriching the cognitive priors for the new domain, which ensures the cognitive state propagation objective. Subsequently, we devise a strategy to generate simulated practice logs for cold-start students through analyzing the behavioral patterns from early-bird students, fulfilling the domain-adaption goal. Consequently, we refine the cognitive states of cold-start students as diagnostic outcomes via virtual data, aligning with the diagnosis-oriented goal. Finally, extensive experiments on six real-world datasets highlight the efficacy of our model for DZCD and its practical application in question recommendation. The code is publicly available at https://github.com/bigdata-ustc/Zero-1-to-3.

Mirror detection is an active research topic in computer vision. However, all existing mirror detectors learn mirror representations from large-scale pixel-wise datasets, which are tedious and expensive to obtain. Although weakly-supervised learning has been widely explored in related topics, we note that popular weak supervision signals (e.g., bounding boxes, scribbles, points) still require some efforts from the user to locate the target objects, with a strong assumption that the images to annotate always contain the target objects. Such an assumption may result in the over-segmentation of mirrors. Our key idea of this work is that the existence of mirrors over a time period may serve as a weak supervision to train a mirror detector, for two reasons. First, if a network can predict the existence of mirrors, it can essentially locate the mirrors. Second, we observe that the reflected contents of a mirror tend to be similar to those in adjacent frames, but exhibit considerable contrast to regions in far-away frames (e.g., non-mirror frames). To this end, in this paper, we propose ZOOM, the first method to learn robust mirror representations from extremely-weak annotations of per-frame ZerO-One Mirror indicators in videos. The key insight of ZOOM is to model the similarity and contrast (between mirror and non-mirror regions) in temporal variations to locate and segment the mirrors. To this end, we propose a novel fusion strategy to leverage temporal consistency information for mirror localization, and a novel temporal similarity-contrast modeling module for mirror segmentation. We construct a new video mirror dataset for training and evaluation. Experimental results under new and standard metrics show that ZOOM performs favorably against existing fully-supervised mirror detection methods.

Lowering the memory requirement in full-parameter training on large models has become a hot research area. MeZO fine-tunes the large language models (LLMs) by just forward passes in a zeroth-order SGD optimizer (ZO-SGD), demonstrating excellent performance with the same GPU memory usage as inference. However, the simulated perturbation stochastic approximation for gradient estimate in MeZO leads to severe oscillations and incurs a substantial time overhead. Moreover, without momentum regularization, MeZO shows severe over-fitting problems. Lastly, the perturbation-irrelevant momentum on ZO-SGD does not improve the convergence rate. This study proposes ZO-AdaMU to resolve the above problems by adapting the simulated perturbation with momentum in its stochastic approximation. Unlike existing adaptive momentum methods, we relocate momentum on simulated perturbation in stochastic gradient approximation. Our convergence analysis and experiments prove this is a better way to improve convergence stability and rate in ZO-SGD. Extensive experiments demonstrate that ZO-AdaMU yields better generalization for LLMs fine-tuning across various NLP tasks than MeZO and its momentum variants.