AAAI 2023 Papers

Robustness of machine learning, often referring to securing performance on different data, is always an active field due to the ubiquitous variety and diversity of data in practice. Many studies have been investigated to enhance the learning process robust in recent years. To this end, there is usually a trade-off that results in somewhat extra cost, e.g., more data samples, more complicated objective functions, more iterations to converge in optimization, etc. Then this problem boils down to finding a better trade-off under some conditions. My recent research focuses on robust machine learning with improved efficiency. Particularly, the efficiency here represents learning speed to find a model, and the number of data required to secure the robustness. In the talk, I will survey three pieces of my recent research by elaborating the algorithmic idea and theoretical analysis as technical contributions --- (i) epoch stochastic gradient descent ascent for min-max problems, (ii) stochastic optimization algorithm for non-convex inf-projection problems, and (iii) neighborhood conformal prediction. In the first two pieces of work, the proposed optimization algorithms are general and cover objective functions for robust machine learning. In the third one, I will elaborate an efficient conformal prediction algorithm that guarantee the robustness of prediction after model is trained. Particularly, the efficiency of conformal prediction is measured by its bandwidth.

Pretraining molecular representation models without labels is fundamental to various applications. Conventional methods mainly process 2D molecular graphs and focus solely on 2D tasks, making their pretrained models incapable of characterizing 3D geometry and thus defective for downstream 3D tasks. In this work, we tackle 3D molecular pretraining in a complete and novel sense. In particular, we first propose to adopt an equivariant energy-based model as the backbone for pretraining, which enjoys the merits of fulfilling the symmetry of 3D space. Then we develop a node-level pretraining loss for force prediction, where we further exploit the Riemann-Gaussian distribution to ensure the loss to be E(3)-invariant, enabling more robustness. Moreover, a graph-level noise scale prediction task is also leveraged to further promote the eventual performance. We evaluate our model pretrained from a large-scale 3D dataset GEOM-QM9 on two challenging 3D benchmarks: MD17 and QM9. Experimental results demonstrate the efficacy of our method against current state-of-the-art pretraining approaches, and verify the validity of our design for each proposed component. Code is available at https://github.com/jiaor17/3D-EMGP.

Most existing Human-Object Interaction (HOI) Detection methods rely heavily on full annotations with predefined HOI categories, which is limited in diversity and costly to scale further. We aim at advancing zero-shot HOI detection to detect both seen and unseen HOIs simultaneously. The fundamental challenges are to discover potential human-object pairs and identify novel HOI categories. To overcome the above challenges, we propose a novel End-to-end zero-shot HOI Detection (EoID) framework via vision-language knowledge distillation. We first design an Interactive Score module combined with a Two-stage Bipartite Matching algorithm to achieve interaction distinguishment for human-object pairs in an action-agnostic manner. Then we transfer the distribution of action probability from the pretrained vision-language teacher as well as the seen ground truth to the HOI model to attain zero-shot HOI classification. Extensive experiments on HICO-Det dataset demonstrate that our model discovers potential interactive pairs and enables the recognition of unseen HOIs. Finally, our method outperforms the previous SOTA under various zero-shot settings. Moreover, our method is generalizable to large-scale object detection data to further scale up the action sets. The source code is available at: https://github.com/mrwu-mac/EoID.

There has been a surge of interest in applying deep learning in particle and nuclear physics to replace labor-intensive offline data analysis with automated online machine learning tasks. This paper details a novel AI-enabled triggering solution for physics experiments in Relativistic Heavy Ion Collider and future Electron-Ion Collider. The triggering system consists of a comprehensive end-to-end pipeline based on Graph Neural Networks that classifies trigger events versus background events, makes online decisions to retain signal data, and enables efficient data acquisition. The triggering system first starts with the coordinates of pixel hits lit up by passing particles in the detector, applies three stages of event processing (hits clustering, track reconstruction, and trigger detection), and labels all processed events with the binary tag of trigger versus background events. By switching among different objective functions, we train the Graph Neural Networks in the pipeline to solve multiple tasks: the edge-level track reconstruction problem, the edge-level track adjacency matrix prediction, and the graph-level trigger detection problem. We propose a novel method to treat the events as track-graphs instead of hit-graphs. This method focuses on intertrack relations and is driven by underlying physics processing. As a result, it attains a solid performance (around 72% accuracy) for trigger detection and outperforms the baseline method using hit-graphs by 2% higher accuracy.

In many real-world applications, predictive methods are used to provide inputs for downstream optimization problems. It has been shown that using the downstream task-based objective to learn the intermediate predictive model is often better than using only intermediate task objectives, such as prediction error. The learning task in the former approach is referred to as end-to-end learning. The difficulty in end-to-end learning lies in differentiating through the optimization problem. Therefore, we propose a neural network architecture that can learn to approximately solve these optimization problems, particularly ensuring its output satisfies the feasibility constraints via alternate projections. We show these projections converge at a geometric rate to the exact projection. Our approach is more computationally efficient than existing methods as we do not need to solve the original optimization problem at each iteration. Furthermore, our approach can be applied to a wider range of optimization problems. We apply this to a shortest path problem for which the first stage forecasting problem is a computer vision task of predicting edge costs from terrain maps, a capacitated multi-product newsvendor problem, and a maximum matching problem. We show that this method out-performs existing approaches in terms of final task-based loss and training time.

Entity linking (EL) is the task of linking the text segments to the referring entities in the knowledge graph, typically decomposed into mention detection, and entity disambiguation. Compared to traditional methods treating the two tasks separately, recent end-to-end entity linking methods exploit the mutual dependency between mentions and entities to achieve better performance. However, existing end-to-end EL methods have problems utilizing the dependency of mentions and entities in the task. To this end, we propose to model the EL task as a hierarchical decision-making process and design a hierarchical reinforcement learning algorithm to solve the problem. We conduct extensive experiments to show that the proposed method achieves state-of-the-art performance in several EL benchmark datasets. Our code is publicly available at https://github.com/lhlclhl/he2eel.

Collaborative Communication platforms (e.g., Slack) support multi-party conversations which contain a large number of messages on shared channels. Multiple conversations intermingle within these messages. The task of conversation disentanglement is to cluster these intermingled messages into conversations. Existing approaches are trained using loss functions that optimize only local decisions, i.e. predicting reply-to links for each message and thereby creating clusters of conversations. In this work, we propose an end-to-end reinforcement learning (RL) approach that directly optimizes a global metric. We observe that using existing global metrics such as variation of information and adjusted rand index as a reward for the RL agent deteriorates its performance. This behaviour is because these metrics completely ignore the reply-to links between messages (local decisions) during reward computation. Therefore, we propose a novel thread-level reward function that captures the global metric without ignoring the local decisions. Through experiments on the Ubuntu IRC dataset, we demonstrate that the proposed RL model improves the performance on both link-level and conversation-level metrics.

Recently there are great efforts on leveraging machine learning and logical reasoning. Many approaches start from a given knowledge base, and then try to utilize the knowledge to help machine learning. In real practice, however, the given knowledge base can often be incomplete or even noisy, and thus, it is crucial to develop the ability of knowledge refinement or enhancement. This paper proposes to enable the Abductive learning (ABL) paradigm to have the ability of knowledge refinement/enhancement. In particular, we focus on the problem that, in contrast to closed-environment tasks where a fixed set of symbols are enough to represent the concepts in the domain, in open-environment tasks new concepts may emerge. Ignoring those new concepts can lead to significant performance decay, whereas it is challenging to identify new concepts and add them to the existing knowledge base with potential conflicts resolved. We propose the ABL_nc approach which exploits machine learning in ABL to identify new concepts from data, exploits knowledge graph to match them with entities, and refines existing knowledge base to resolve conflicts. The refined/enhanced knowledge base can then be used in the next loop of ABL and help improve the performance of machine learning. Experiments on three neuro-symbolic learning tasks verified the effectiveness of the proposed approach.

It is common to listen to songs that match one's mood. Thus, an AI music recommendation system that is aware of the user's emotions is likely to provide a superior user experience to one that is unaware. In this paper, we present an emotion-aware music recommendation system. Multiple models are discussed and evaluated for affect identification from a live image of the user. We propose two models: DRViT, which applies dynamic routing to vision transformers, and InvNet50, which uses involution. All considered models are trained and evaluated on the AffectNet dataset. Each model outputs the user's estimated valence and arousal under the circumplex model of affect. These values are compared to the valence and arousal values for songs in a Spotify dataset, and the top-five closest-matching songs are presented to the user. Experimental results of the models and user testing are presented.

The field of emergent communication aims to understand the characteristics of communication as it emerges from artificial agents solving tasks that require information exchange. Communication with discrete messages is considered a desired characteristic, for scientific and applied reasons. However, training a multi-agent system with discrete communication is not straightforward, requiring either reinforcement learning algorithms or relaxing the discreteness requirement via a continuous approximation such as the Gumbel-softmax. Both these solutions result in poor performance compared to fully continuous communication. In this work, we propose an alternative approach to achieve discrete communication -- quantization of communicated message. Using message quantization allows us to train the model end-to-end, achieving superior performance in multiple setups. Moreover, quantization is a natural framework that runs the gamut from continuous to discrete communication. Thus, it sets the ground for a broader view of multi-agent communication in the deep learning era.

Altruistic punishment (or punishment) has been extensively shown as an important mechanism for promoting cooperation in human societies. In AI, the emergence of punishment has received much recent interest. In this paper, we contribute with a novel evolutionary game theoretic model to study the impacts of environmental feedback. Whereas a population of agents plays public goods games, there exists a third-party population whose payoffs depend not only on whether to punish or not, but also on the state of the environment (e.g., how cooperative the agents in a social dilemma are). Focusing on one-shot public goods games, we show that environmental feedback, by itself, can lead to the emergence of punishment. We analyze the co-evolution of punishment and cooperation, and derive conditions for their co-presence, co-dominance and co-extinction. Moreover, we show that the system can exhibit bistability as well as cyclic dynamics. Our findings provide a new explanation for the emergence of punishment. On the other hand, our results also alert the need for careful design of implementing punishment in multi-agent systems, as the resulting evolutionary dynamics can be somewhat complex.

High-performing human teams leverage intelligent and efficient communication and coordination strategies to collaboratively maximize their joint utility. Inspired by teaming behaviors among humans, I seek to develop computational methods for synthesizing intelligent communication and coordination strategies for collaborative multi-robot systems. I leverage both classical model-based control and planning approaches as well as data-driven methods such as Multi-Agent Reinforcement Learning (MARL) to provide several contributions towards enabling emergent cooperative teaming behavior across both homogeneous and heterogeneous (including agents with different capabilities) robot teams.

Increasing global food demand, accompanied by the limited number of expert growers, brings the need for more sustainable and efficient horticulture. The controlled environment of greenhouses enable data collection and precise control. For optimally controlling the greenhouse climate, a grower not only looks at crop production, but rather aims at maximising the profit. However this is a complex, long term optimisation task. In this paper, Constraint Programming (CP) is applied to task of optimal greenhouse economic control, by leveraging a learned greenhouse climate model through a CP embedding. In collaboration with an industrial partner, we demonstrate how to model the greenhouse climate with an LSTM model, embed this LSTM into a CP optimisation framework, and optimise the expected profit of the grower. This data-to-decision pipeline is being integrated into a decision support system for multiple greenhouses in the Netherlands.

The rise of AI methods to make predictions and decisions has led to a pressing need for more explainable artificial intelligence (XAI) methods. One common approach for XAI is to produce a post-hoc explanation, explaining why a black box ML model made a certain prediction. Formal approaches to post-hoc explanations provide succinct reasons for why a prediction was made, as well as why not another prediction was made. But these approaches assume that features are independent and uniformly distributed. While this means that “why” explanations are correct, they may be longer than required. It also means the “why not” explanations may be suspect as the counterexamples they rely on may not be meaningful. In this paper, we show how one can apply background knowledge to give more succinct “why” formal explanations, that are presumably easier to interpret by humans, and give more accurate “why not” explanations. In addition, we show how to use existing rule induction techniques to efficiently extract background information from a dataset.

Graph Contrastive Learning (GCL) has recently drawn much research interest for learning generalizable node representations in a self-supervised manner. In general, the contrastive learning process in GCL is performed on top of the representations learned by a graph neural network (GNN) backbone, which transforms and propagates the node contextual information based on its local neighborhoods. However, nodes sharing similar characteristics may not always be geographically close, which poses a great challenge for unsupervised GCL efforts due to their inherent limitations in capturing such global graph knowledge. In this work, we address their inherent limitations by proposing a simple yet effective framework -- Simple Neural Networks with Structural and Semantic Contrastive Learning} (S^3-CL). Notably, by virtue of the proposed structural and semantic contrastive learning algorithms, even a simple neural network can learn expressive node representations that preserve valuable global structural and semantic patterns. Our experiments demonstrate that the node representations learned by S^3-CL) achieve superior performance on different downstream tasks compared with the state-of-the-art unsupervised GCL methods. Implementation and more experimental details are publicly available at https://github.com/kaize0409/S-3-CL.

Electrophysiological Source Imaging (ESI) refers to reconstructing the underlying brain source activation from non-invasive Electroencephalography (EEG) and Magnetoencephalography (MEG) measurements on the scalp. Estimating the source locations and their extents is a fundamental tool in clinical and neuroscience applications. However, the estimation is challenging because of the ill-posedness and high coherence in the leadfield matrix as well as the noise in the EEG/MEG data. In this work, we proposed a combinatorial search framework to address the ESI problem with a provable optimality guarantee. Specifically, by exploiting the graph neighborhood information in the brain source space, we converted the ESI problem into a graph search problem and designed a combinatorial search algorithm under the framework of A* to solve it. The proposed algorithm is guaranteed to give an optimal solution to the ESI problem. Experimental results on both synthetic data and real epilepsy EEG data demonstrated that the proposed algorithm could faithfully reconstruct the source activation in the brain.

Mixture models of Plackett-Luce (PL), one of the most fundamental ranking models, are an active research area of both theoretical and practical significance. Most previously proposed parameter estimation algorithms instantiate the EM algorithm, often with random initialization. However, such an initialization scheme may not yield a good initial estimate and the algorithms require multiple restarts, incurring a large time complexity. As for the EM procedure, while the E-step can be performed efficiently, maximizing the log-likelihood in the M-step is difficult due to the combinatorial nature of the PL likelihood function. Therefore, previous authors favor algorithms that maximize surrogate likelihood functions. However, the final estimate may deviate from the true maximum likelihood estimate as a consequence. In this paper, we address these known limitations. We propose an initialization algorithm that can provide a provably accurate initial estimate and an EM algorithm that maximizes the true log-likelihood function efficiently. Experiments on both synthetic and real datasets show that our algorithm is competitive in terms of accuracy and speed to baseline algorithms, especially on datasets with a large number of items.

Artificial intelligence (AI)-empowered industrial fault diagnostics is important in ensuring the safe operation of industrial applications. Since complex industrial systems often involve multiple industrial plants (possibly belonging to different companies or subsidiaries) with sensitive data collected and stored in a distributed manner, collaborative fault diagnostic model training often needs to leverage federated learning (FL). As the scale of the industrial fault diagnostic models are often large and communication channels in such systems are often not exclusively used for FL model training, existing deployed FL model training frameworks cannot train such models efficiently across multiple institutions. In this paper, we report our experience developing and deploying the Federated Opportunistic Block Dropout (FedOBD) approach for industrial fault diagnostic model training. By decomposing large-scale models into semantic blocks and enabling FL participants to opportunistically upload selected important blocks in a quantized manner, it significantly reduces the communication overhead while maintaining model performance. Since its deployment in ENN Group in February 2022, FedOBD has served two coal chemical plants across two cities in China to build industrial fault prediction models. It helped the company reduce the training communication overhead by over 70% compared to its previous AI Engine, while maintaining model performance at over 85% test F1 score. To our knowledge, it is the first successfully deployed dropout-based FL approach.

Sparse learning based feature selection has been widely investigated in recent years. In this study, we focus on the l2,0-norm based feature selection, which is effective for exact top-k feature selection but challenging to optimize. To solve the general l2,0-norm constrained problems, we novelly develop a parameter-free optimization framework based on the coordinate descend (CD) method, termed CD-LSR. Specifically, we devise a skillful conversion from the original problem to solving one continuous matrix and one discrete selection matrix. Then the nontrivial l2,0-norm constraint can be solved efficiently by solving the selection matrix with CD method. We impose the l2,0-norm on a vanilla least square regression (LSR) model for feature selection and optimize it with CD-LSR. Extensive experiments exhibit the efficiency of CD-LSR, as well as the discrimination ability of l2,0-norm to identify informative features. More importantly, the versatility of CD-LSR facilitates the applications of the l2,0-norm in more sophisticated models. Based on the competitive performance of l2,0-norm on the baseline LSR model, the satisfactory performance of its applications is reasonably expected. The source MATLAB code are available at: https://github.com/solerxl/Code_For_AAAI_2023.