IJCAI 2024 Papers

Overfitting in adversarial training has attracted the interest of researchers in the community of artificial intelligence and machine learning in recent years. To address this issue, in this paper we begin by evaluating the defense performances of several calibration methods on various robust models. Our analysis and experiments reveal two intriguing properties: 1) a well-calibrated robust model is decreasing the confidence of robust model; 2) there is a trade-off between the confidences of natural and adversarial images. These new properties offer a straightforward insight into designing a simple but effective regularization, called Self-Residual-Calibration (SRC). The proposed SRC calculates the absolute residual between adversarial and natural logit features corresponding to the ground-truth labels. Furthermore, we utilize the pinball loss to minimize the quantile residual between them, resulting in more robust regularization. Extensive experiments indicate that our SRC can effectively mitigate the overfitting problem while improving the robustness of state-of-the-art models. Importantly, SRC is complementary to various regularization methods. When combined with them, we are capable of achieving the top-rank performance on the AutoAttack benchmark leaderboard.

Counterfactual regret minimization (CFR) is a family of algorithms for effectively solving imperfect-information games. It decomposes the total regret into counterfactual regrets, utilizing local regret minimization algorithms, such as Regret Matching (RM) or RM+, to minimize them. Recent research establishes a connection between Online Mirror Descent (OMD) and RM+, paving the way for an optimistic variant PRM+ and its extension PCFR+. However, PCFR+ assigns uniform weights for each iteration when determining regrets, leading to substantial regrets when facing dominated actions. This work explores minimizing weighted counterfactual regret with optimistic OMD, resulting in a novel CFR variant PDCFR+. It integrates PCFR+ and Discounted CFR (DCFR) in a principled manner, swiftly mitigating negative effects of dominated actions and consistently leveraging predictions to accelerate convergence. Theoretical analyses prove that PDCFR+ converges to a Nash equilibrium, particularly under distinct weighting schemes for regrets and average strategies. Experimental results demonstrate PDCFR+'s fast convergence in common imperfect-information games. The code is available at https://github.com/rpSebastian/PDCFRPlus.

In many social choice applications, information about individuals' preferences can only be elicited using a limited number of pairwise comparisons. In these cases, the task is twofold: we must first choose the queries, and then second, we must aggregate the responses to choose an outcome. We study the problem of designing algorithms for this setting. To compare the effectiveness of different outcomes, we use the metric distortion framework. In addition, we consider various constraints on the query algorithms, namely, placing restrictions on how the choice of the next query may depend on previous answers. Our main contributions are nearly optimal algorithms for all settings considered.

Conditional Independence (CI) graphs are a type of probabilistic graphical models that are primarily used to gain insights about feature relationships. Each edge represents the partial correlation between the connected features which gives information about their direct dependence. In this survey, we list out different methods and study the advances in techniques developed to recover CI graphs. We cover traditional optimization methods as well as recently developed deep learning architectures along with their recommended implementations . To facilitate wider adoption, we include preliminaries that consolidate associated operations, for example techniques to obtain covariance matrix for mixed datatypes. Keywords: Conditional Independence Graphs, Probabilistic Graphical Models, Graphical Lasso, Deep Learning, Optimization

The modeling of just noticeable difference (JND) in supervised learning for visual signals has made significant progress. However, existing JND models often suffer from limited generalization due to the need for large-scale training data and their constraints to certain image types. Moreover, these models primarily focus on a single RGB modality, ignoring the potential complementary impacts of multiple modalities. To address these challenges, we propose a new meta-learning approach for the JND modeling, called MetaJND. We introduce two key visual-sensitive modalities like saliency and depth, and leverage a self-attention mechanism for effective interdependence of multi-modal features. Additionally, we incorporate meta-learning for the modality alignment, facilitating dynamic weight generation. Furthermore, we perform hierarchical fusion through multi-layer channel and spatial feature rectification. Experimental results on four benchmark datasets demonstrate the effectiveness of our MetaJND. Moreover, we have also evaluated its performance in compression and watermarking applications, observing higher bit-rate savings and better watermark hiding capabilities.

State-of-the-art deep ISP models alleviate the dilemma of limited generalization capabilities across heterogeneous inputs by increasing the size and complexity of the network, which inevitably leads to considerable growth in parameter counts and FLOPs. To address this challenge, this paper presents MetaISP - a streamlined model that achieves superior reconstruction quality by adaptively modulating its parameters and architecture in response to diverse inputs. Our rationale revolves around obtaining corresponding spatial and channel-wise correction matrices for various inputs within distinct feature spaces, which assists in assigning optimal attention. This is achieved by predicting dynamic weights for each input image and combining these weights with multiple learnable basis matrices to construct the correction matrices. The proposed MetaISP makes it possible to obtain best performance while being computationally efficient. SOTA results are achieved on two large-scale datasets, e.g. 23.80dB PSNR on ZRR, exceeding the previous SOTA 0.19dB with only 9.2% of its parameter count and 10.6% of its FLOPs; 25.06dB PSNR on MAI21, exceeding the previous SOTA 0.17dB with only 0.9% of its parameter count and 2.7% of its FLOPs.

Meta-learning accelerates the learning process on unseen learning tasks by acquiring prior knowledge through previous related tasks. The PAC-Bayesian theory provides a theoretical framework to analyze the generalization of meta-learning to unseen tasks. However, previous works still encounter two notable limitations: (1) they merely focus on the data-free priors, which often result in inappropriate regularization and loose generalization bounds; (2) more importantly, their optimization process usually involves nested optimization problems, incurring significant computational costs. To address these issues, we derive new generalization bounds and introduce a novel PAC-Bayesian framework for meta-learning that integrates data-dependent priors. This framework enables the extraction of optimal posteriors for each task in closed form, thereby allowing us to minimize generalization bounds incorporated data-dependent priors with only a simple local entropy. The resulting algorithm, which employs SGLD for sampling from the optimal posteriors, is stable, efficient, and computationally lightweight, eliminating the need for nested optimization. Extensive experimental results demonstrate that our proposed method outperforms the other baselines.

Relation extraction (RE) is an important task that aims to identify the relationships between entities in texts. While large language models (LLMs) have revealed remarkable in-context learning (ICL) capability for general zero and few-shot learning, recent studies indicate that current LLMs still struggle with zero and few-shot RE. Previous studies are mainly dedicated to design prompt formats and select good examples for improving ICL-based RE. Although both factors are vital for ICL, if one can fundamentally boost the ICL capability of LLMs in RE, the zero and few-shot RE performance via ICL would be significantly improved. To this end, we introduce Micre (Meta In-Context learning of LLMs for Relation Extraction), a new meta-training framework for zero and few-shot RE where an LLM is tuned to do ICL on a diverse collection of RE datasets (i.e., learning to learn in context for RE). Through meta-training, the model becomes more effectively to learn a new RE task in context by conditioning on a few training examples with no parameter updates or task-specific templates at inference time, enabling better zero and few-shot task generalization. We experiment Micre on various LLMs with different model scales and 12 public RE datasets, and then evaluate it on unseen RE benchmarks under zero and few-shot settings. Micre delivers comparable or superior performance compared to a range of baselines including supervised fine-tuning and typical in-context learning methods. We find that the gains are particular significant for larger model scales, and using a diverse set of the meta-training RE datasets is key to improvements. Empirically, we show that Micre can transfer the relation semantic knowledge via relation label name during inference on target RE datasets.

Training data plays a pivotal role in AI models. Large language models (LLMs) are trained with massive amounts of documents, and their parameters hold document-related contents. Recently, several studies identified content-specific locations in LLMs by examining the parameters. Instead of the post hoc interpretation, we propose another approach. We propose document-wise memory architecture to track document memories in training. The proposed architecture maps document representations to memory entries, which softly mask memories in the forward process of LLMs. Additionally, we propose document guidance loss, which increases the likelihood of text with document memories and reduces the likelihood of the text with the memories of other documents. Experimental results on Wikitext-103-v1 with Pythia-1B show that the proposed methods provide different memory entries for documents and high recall of document-related content in generation with trained document-wise memories.

This paper presents a comprehensive survey of Medical Neural Architecture Search (MedNAS), a burgeoning field at the confluence of deep learning and medical imaging. With the increasing prevalence of FDA-approved medical deep learning models, MedNAS emerges as a key area in leveraging computational innovations for healthcare advancements. Our survey examines the paradigm shift introduced by Neural Architecture Search (NAS), which automates neural network design, replacing traditional, manual designs. We explore the unique search spaces tailored for medical tasks on different types of data from images to EEG, the methodologies of MedNAS, and their impact on medical applications.

Tabular data prediction has been employed in medical applications such as patient health risk prediction. However, existing methods usually revolve around the algorithm design while overlooking the significance of data engineering. Medical tabular datasets frequently exhibit significant heterogeneity across different sources, with limited sample sizes per source. As such, previous predictors are often trained on manually curated small datasets that struggle to generalize across different tabular datasets during inference. This paper proposes to scale medical tabular data predictors (MediTab) to various tabular inputs with varying features. The method uses a data engine that leverages large language models (LLMs) to consolidate tabular samples to overcome the barrier across tables with distinct schema. It also aligns out-domain data with the target task using a "learn, annotate, and refinement'' pipeline. The expanded training data then enables the pre-trained MediTab to infer for arbitrary tabular input in the domain without fine-tuning, resulting in significant improvements over supervised baselines: it reaches an average ranking of 1.57 and 1.00 on 7 patient outcome prediction datasets and 3 trial outcome prediction datasets, respectively. In addition, MediTab exhibits impressive zero-shot performances: it outperforms supervised XGBoost models by 8.9% and 17.2% on average in two prediction tasks, respectively.

Randomized mechanisms can have good normative properties compared to their deterministic counter-parts. However, randomized mechanisms are problematic in several ways such as in their verifiability. We propose here to de-randomize such mechanisms by having agents play a game instead of tossing a coin. The game is designed so agents play randomly, and this play injects “randomness” into the mechanism. Surprisingly this de-randomization retains many of the good normative properties of the original randomized mechanism but gives a mechanism that is deterministic and easy, for instance, to audit. We consider three general purpose methods to de-randomize mechanisms, and apply these to six different domains: voting, facility location, task allocation, school choice, peer selection, and resource allocation. We propose a number of novel de-randomized mechanisms for these six domains with good normative properties (such as equilibria in which agents sincerely report preferences over the original problem). In one domain, we additionally show that a new and desirable normative property emerges as a result of de-randomization.property emerges as a result of de-randomization.

Robustness to malicious attacks is of paramount importance for distributed learning. Existing works often consider the classical Byzantine attacks model, which assumes that some workers can send arbitrarily malicious messages to the server and disturb the aggregation steps of the distributed learning process. To defend against such worst-case Byzantine attacks, various robust aggregators have been proven effective and much superior to the often-used mean aggregator. In this paper, we show that robust aggregators are too conservative for a class of weak but practical malicious attacks, as known as label poisoning attacks, where the sample labels of some workers are poisoned. Surprisingly, we are able to show that the mean aggregator is more robust than the state-of-the-art robust aggregators in theory, given that the distributed data are sufficiently heterogeneous. In fact, the learning error of the mean aggregator is proven to be optimal in order. Experimental results corroborate our theoretical findings, demonstrating the superiority of the mean aggregator under label poisoning attacks.

SimRank is one of the most fundamental measures that evaluate the structural similarity between two nodes in a graph and has been applied in a plethora of data mining and machine learning tasks. These tasks often involve single-source SimRank computation that evaluates the SimRank values between a source node u and all other nodes. Due to its high computation complexity, single-source SimRank computation for large graphs is notoriously challenging, and hence recent studies resort to distributed processing. To our surprise, although SimRank has been widely adopted for two decades, theoretical aspects of distributed SimRanks with provable results have rarely been studied. In this paper, we conduct a theoretical study on single-source SimRank computation in the Massive Parallel Computation (MPC) model, which is the standard theoretical framework modeling distributed systems. Existing distributed SimRank algorithms enforce either Ω(log n) communication round complexity or Ω(n) machine space for a graph of n nodes. We overcome this barrier. Particularly, given a graph of n nodes, for any query node v and constant error ϵ>3/n, we show that using O(log² log n) rounds of communication among machines is enough to compute single-source SimRank values with at most ϵ absolute errors, while each machine only needs a space sub-linear to n. To the best of our knowledge, this is the first single-source SimRank algorithm in MPC that can overcome the Θ(log n) round complexity barrier with provable result accuracy.

This paper presents NgramMarkov, a variant of the Markov constraints. It is dedicated to text generation in constraint programming (CP). It involves a set of n-grams (i.e., sequence of n words) associated with probabilities given by a large language model (LLM). It limits the product of the probabilities of the n-gram of a sentence. The propagator of this constraint can be seen as an extension of the ElementaryMarkov constraint propagator, incorporating the LLM distribution instead of the maximum likelihood estimation of n-grams. It uses a gliding threshold, i.e., it rejects n-grams whose local probabilities are too low, to guarantee balanced solutions. It can also be combined with a "look-ahead" approach to remove n-grams that are very unlikely to lead to acceptable sentences for a fixed-length horizon. This idea is based on the MDDMarkovProcess constraint propagator, but without explicitly using an MDD (Multi-Valued Decision Diagram). The experimental results show that the generated text is valued in a similar way to the LLM perplexity function. Using this new constraint dramatically reduces the number of candidate sentences produced, improves computation times, and allows larger corpora or smaller n-grams to be used. A real-world problem has been solved for the first time using 4-grams instead of 5-grams.

Prompt engineering is still the primary way for users of generative text-to-image models to manipulate generated images in a targeted way. Based on treating the model as a continuous function and by passing gradients between the image space and the prompt embedding space, we propose and analyze a new method to directly manipulate the embedding of a prompt instead of the prompt text. We then derive three practical interaction tools to support users with image generation: (1) Optimization of a metric defined in the image space that measures, for example, the image style. (2) Supporting a user in creative tasks by allowing them to navigate in the image space along a selection of directions of "near" prompt embeddings. (3) Changing the embedding of the prompt to include information that a user has seen in a particular seed but has difficulty describing in the prompt. Compared to prompt engineering, user-driven prompt embedding manipulation enables a more fine-grained, targeted control that integrates a user's intentions. Our user study shows that our methods are considered less tedious and that the resulting images are often preferred.

Relation Extraction (RE) identifies relations between entities in text, typically relying on supervised models that demand abundant high-quality data. Various approaches, including Data Augmentation (DA), have been proposed as promising solutions for addressing low-resource challenges in RE. However, existing DA methods in RE often struggle to ensure consistency and contextual diversity in generated data due to the fine-grained nature of RE. Inspired by the extensive generative capabilities of large language models (LLMs), we introduce a novel framework named ConsistRE, aiming to maintain context consistency in RE. ConsistRE initiates by collecting a substantial corpus from external resources and employing statistical algorithms and semantics to identify keyword hints closely related to relation instances. These keyword hints are subsequently integrated as contextual constraints in sentence generation, ensuring the preservation of relation dependence and diversity with LLMs. Additionally, we implement syntactic dependency selection to enhance the syntactic structure of the generated sentences. Experimental results from the evaluation of SemEval, TACRED, and TACREV datasets unequivocally demonstrate that ConsistRE outperforms other baselines in F1 values by 1.76%, 3.92%, and 2.53%, respectively, particularly when operating under low-resource experimental conditions.

Urban traffic speed prediction aims to estimate the future traffic speed for improving urban transportation services. Enormous efforts have been made to exploit Graph Neural Networks (GNNs) for modeling spatial correlations and temporal dependencies of traffic speed evolving patterns, regularized by graph topology. While achieving promising results, current traffic speed prediction methods still suffer from ignoring topology-free patterns, which cannot be captured by GNNs. To tackle this challenge, we propose a generic model for enabling the current GNN-based methods to preserve topology-free patterns. Specifically, we first develop a Dual Cross-Scale Transformer (DCST) architecture, including a Spatial Transformer and a Temporal Transformer, to preserve the cross-scale topology-free patterns and associated dynamics, respectively. Then, to further integrate both topology-regularized/-free patterns, we propose a distillation-style learning framework, in which the existing GNN-based methods are considered as the teacher model, and the proposed DCST architecture is considered as the student model. The teacher model would inject the learned topology-regularized patterns into the student model for integrating topology-free patterns. The extensive experimental results demonstrated the effectiveness of our methods.

Traffic prediction plays a key role in various smart city applications, which can help traffic managers make traffic plans in advance, assist online ride-hailing companies in deploying vehicles reasonably, and provide early warning of congestion for safety authorities. While increasingly complex models achieve impressive prediction performance, there are concerns about the effectiveness of these models in handling large-scale road networks. Especially for researchers who don't have access to powerful GPU devices, the expensive memory burden limits the usefulness of these models. In this paper, we take the first step of learning on the large-scale spatio-temporal graph and propose a divide-and-conquer training strategy for Large Spatio-Temporal Graph Learning, namely LarSTL. The core idea behind this strategy is to divide the large graph into multiple subgraphs, which are treated as task streams to sequentially train the model to conquer each subgraph one by one. We introduce a novel perspective based on the continuous learning paradigm to achieve this goal. In order to overcome forgetting the knowledge learned from previous subgraphs, an experience-replay strategy consolidates the learned knowledge by replaying nodes sampled from previous subgraphs. Moreover, we configure specific feature adaptors for each subgraph to extract personalized features, and it is also beneficial to consolidate the learned knowledge from the perspective of parameters. We conduct experiments using multiple large-scale traffic network datasets on a V100 GPU with only 16GB memory, and the results demonstrate that our LarSTL can achieve competitive performance and high efficiency.