AAAI 2025 Papers

Split Federated Learning (SFL) splits and collaboratively trains a shared model between clients and server, where clients transmit activations and client-side models to server for updates. Recent SFL studies assume synchronous transmission of activations and client-side models from clients to server. However, due to significant variations in computational and communication capabilities among clients, activations and client-side models arrive at server asynchronously. The delay caused by asynchrony significantly degrades the performance of SFL. To address this issue, we consider an asynchronous SFL framework, where an activation buffer and a model buffer are embedded on the server to manage the asynchronously transmitted activations and client-side models, respectively. Furthermore, as asynchronous activation transmissions cause the buffer to frequently receive activations from resource-rich clients, leading to biased updates of the server-side model, we propose Generative activations-aided Asynchronous SFL (GAS). In GAS, the server maintains an activation distribution for each label based on received activations and generates activations from these distributions according to the degree of bias. These generative activations are then used to assist in updating the server-side model, ensuring more accurate updates. We derive a tighter convergence bound, and our experiments demonstrate the effectiveness of the proposed method.

As urban residents demand higher travel quality, vehicle dispatch has become a critical component of online ride-hailing services. However, current vehicle dispatch systems struggle to navigate the complexities of urban traffic dynamics, including unpredictable traffic conditions, diverse driver behaviors, and fluctuating supply and demand patterns. These challenges have resulted in travel difficulties for passengers in certain areas, while many drivers in other areas are unable to secure orders, leading to a decline in the overall quality of urban transportation services. To address these issues, this paper introduces GARLIC: a framework of GPT-Augmented Reinforcement Learning with Intelligent Control for vehicle dispatching. GARLIC utilizes multiview graphs to capture hierarchical traffic states, and learns a dynamic reward function that accounts for individual driving behaviors. The framework further integrates a GPT model trained with a custom loss function to enable high-precision predictions and optimize dispatching policies in real-world scenarios. Experiments conducted on two real-world datasets demonstrate that GARLIC effectively aligns with driver behaviors while reducing the empty load rate of vehicles.

Hallucination detection has attracted considerable interest due to the tendency of language models to generate texts that contain hallucinations. Most existing methods start with specific local details directly extracted from text, then aggregate to form the final conclusion. However, this direct extraction approach ignores the global context, leading to isolated details, and is prone to missed or over-detections. In this paper, we present a global-to-local approach for hallucination detection (G2LDetect), which considers the global information of the text before identifying local details. We first construct a global representation of the text by transforming it into a hierarchical tree structure. Afterward, we obtain specific local details from the global tree representation using path-wise identification and perform detection on them. This global-to-local detection process ensures that local details are context-aware and complete, thus making more accurate and reliable detection results. Experimental results show that our global-to-local method outperforms existing methods, especially for longer texts.

Evaluation metric of visual captioning is important yet not thoroughly explored. Traditional metrics like BLEU, METEOR, CIDEr, and ROUGE often miss semantic depth, while trained metrics such as CLIP-Score, PAC-S, and Polos are limited in zero-shot scenarios. Advanced Language Model-based metrics also struggle with aligning to nuanced human preferences. To address these issues, we introduce G-VEval, a novel metric inspired by G-Eval and powered by the new GPT-4o. G-VEval uses chain-of-thought reasoning in large multimodal models and supports three modes: reference-free, reference-only, and combined, accommodating both video and image inputs. We also propose MSVD-Eval, a new dataset for video captioning evaluation, to establish a more transparent and consistent framework for both human experts and evaluation metrics. It is designed to address the lack of clear criteria in existing datasets by introducing distinct dimensions of Accuracy, Completeness, Conciseness, and Relevance (ACCR). Extensive results show that G-VEval outperforms existing methods in correlation with human annotations, as measured by Kendall tau-b and Kendall tau-c. This provides a flexible solution for diverse captioning tasks and suggests a straightforward yet effective approach for large language models to understand video content, paving the way for advancements in automated captioning.

Collaborative reasoning enhances recommendation performance by combining the strengths of symbolic learning and deep neural learning. However, current collaborative reasoning models rely on parameterized networks to simulate logical operations within the reasoning process, which (1) do not comply with all axiomatic principles of classical logic and (2) limit the model's generalizability. To address these limitations, a Fuzzy logic approach tailored for Collaborative Reasoning (FuzzCR) is proposed in this work, aiming to augment the recommendation system with cognitive abilities. Specifically, this method redefines the sequential recommendation task as a logical query answering process to facilitate a more structured and logical progression of reasoning. Moreover, learning-free fuzzy logical operations are implemented for the designed reasoning process. Taking advantage of the inherent properties of fuzzy logic, these logical operations satisfy fundamental logical rules and ensure complete reasoning. After training, these operations can be applied to flexible reasoning processes, rather than being confined to fixed computation graphs, thereby exhibiting good generalizability. Extensive experiments conducted on publicly available datasets demonstrate the superiority of this method in solving the sequential recommendation task.

Sequential recommendation (SR) systems predict user preferences by analyzing time-ordered interaction sequences. A common challenge for SR is data sparsity, as users typically interact with only a limited number of items. While contrastive learning has been employed in previous approaches to address the challenges, these methods often adopt binary labels, missing finer patterns and overlooking detailed information in subsequent behaviors of users. Additionally, they rely on random sampling to select negatives in contrastive learning, which may not yield sufficiently hard negatives during later training stages. In this paper, we propose Future data utilization with Enduring Negatives for contrastive learning in sequential Recommendation (FENRec). Our approach aims to leverage future data with time-dependent soft labels and generate enduring hard negatives from existing data, thereby enhancing the effectiveness in tackling data sparsity. Experiment results demonstrate our state-of-the-art performance across four benchmark datasets, with an average improvement of 6.16% across all metrics.

Backdoor attacks present a serious security threat to deep neuron networks (DNNs). Although numerous effective defense techniques have been proposed in recent years, they inevitably rely on the availability of either clean or poisoned data. In contrast, data-free defense techniques have evolved slowly and still lag significantly in performance. To address this issue, different from the traditional approach of pruning followed by fine-tuning, we propose a novel data-free defense method named Optimal Transport-based Backdoor Repairing (OTBR) in this work. This method, based on our findings on neuron weight changes (NWCs) of random unlearning, uses optimal transport (OT)-based model fusion to combine the advantages of both pruned and backdoored models. Specifically, we first demonstrate our findings that the NWCs of random unlearning are positively correlated with those of poison unlearning. Based on this observation, we propose a random-unlearning NWC pruning technique to eliminate the backdoor effect and obtain a backdoor-free pruned model. Then, motivated by the OT-based model fusion, we propose the pruned-to-backdoored OT-based fusion technique, which fuses pruned and backdoored models to combine the advantages of both, resulting in a model that demonstrates high clean accuracy and a low attack success rate. To our knowledge, this is the first work to apply OT and model fusion techniques to backdoor defense. Extensive experiments show that our method successfully defends against all seven backdoor attacks across three benchmark datasets, outperforming both state-of-the-art (SOTA) data-free and data-dependent methods.

The human brain is a complex system, and understanding its mechanisms has been a long-standing challenge in neuroscience. The study of the functional connectome, which maps the functional connections between different brain regions, has provided valuable insights through various advanced analysis techniques developed over the years. Similarly, neural networks, inspired by the brain's architecture, have achieved notable success in diverse applications but are often noted for their lack of interpretability. In this paper, we propose a novel approach that bridges neural networks and human brain functions by leveraging brain-inspired techniques. Our approach, grounded in the insights from the functional connectome, offers scalable ways to characterize topology of large neural networks using stable statistical and machine learning techniques. Our empirical analysis demonstrates its capability to enhance the interpretability of neural networks, providing a deeper understanding of their underlying mechanisms.

Diffusion models have demonstrated outstanding performance in generative tasks, making them ideal candidates for image editing. Recent studies highlight their ability to apply desired edits effectively by following textual instructions, yet with two key challenges remaining. First, these models struggle to apply multiple edits simultaneously, resulting in computational inefficiencies due to their reliance on sequential processing. Second, relying on textual prompts to determine the editing region can lead to unintended alterations to the image. We introduce FunEditor, an efficient diffusion model designed to learn atomic editing functions and perform complex edits by aggregating simpler functions. This approach enables complex editing tasks, such as object movement, by aggregating multiple functions and applying them simultaneously to specific areas. Our experiments demonstrate that FunEditor significantly outperforms recent inference-time optimization methods and fine-tuned models, either quantitatively across various metrics or through visual comparisons or both, on complex tasks like object movement and object pasting. In the meantime, with only 4 steps of inference, FunEditor achieves 5--24 times inference speedups over existing popular methods.

In this paper, we consider the k-center problem with outliers (the (k, z)-center problem) in the context of Massively Parallel Computation (MPC). Existing MPC algorithms for the (k, z)-center problem typically require Ω(k) local space per machine. While this may be feasible when k is small, these algorithms become impractical for large k, where each machine may lack sufficient space for computation. This motivates the study of fully-scalable algorithms with sublinear local space. We propose the first fully-scalable MPC algorithm for the (k, z)-center problem. The main challenge is to design an MPC algorithm that operates with sublinear local space for finding the inliers close to the optimal clustering centers, and ensuring the approximation loss remains bounded. To address this issue, we propose an iterative sampling-based algorithm with sublinear local space in the data size. A key component of our approach is an outliers-removal algorithm that adjusts the sample size in each iteration to select inliers as clustering centers. However, the number of discarded inliers increases with the iteration of the outliers-removal algorithm, making it difficult to bound. To address this, we propose a self-adaptive method that can automatically adjust sample size to account for different data distributions on each machine, ensuring a lower bound on the sampling success probability. With these techniques, we present an O(log^*n)-approximation MPC algorithm for the (k, z)-center problem in constant-dimensional Euclidean space. The algorithm discards at most (1 + ε)z outliers, completing in O(log log n) computation rounds while using Θ(n^δ) local space per machine.

Tabular data plays a vital role in various real-world scenarios and finds extensive applications. Although recent deep tabular models have shown remarkable success, they still struggle to handle data distribution shifts, leading to performance degradation when testing distributions change. To remedy this, a robust tabular model must adapt to generalize to unknown distributions during testing. In this paper, we investigate the problem of fully test-time adaptation (FTTA) for tabular data, where the model is adapted using only the testing data. We identify three key challenges: the existence of label and covariate distribution shifts, the lack of effective data augmentation, and the sensitivity of adaptation, which render existing FTTA methods ineffective for tabular data. To this end, we propose the Fully Test-time Adaptation for Tabular data, namely FTAT, which enables FTTA methods to robustly optimize the label distribution of predictions, adapt to shifted covariate distributions, and dynamically adapt the model for various tasks and models. We conduct comprehensive experiments on six benchmark datasets, which are evaluated using three metrics. The experimental results demonstrate that FTAT outperforms state-of-the-art methods by a margin.

Large language models (LLMs) have demonstrated remarkable performance in multimodal tasks even with frozen LLM Block and only a few trainable parameters. However, the underlying mechanisms of how LLMs enhance multimodal performance remains unclear. In this work, we focus on the phenomenon that ``Merely concatenating a frozen LLM block to the Vision Transformer (ViT) encoder can yield significant performance enhancements. Moreover, the choice of LLM block and insertion position can have a substantial impact, leading to varying degrees of improvement''. We analyze the optimization of the training process from the perspective of gradient dynamics and find that frozen LLM blocks act as gradient coherence rectifiers, aligning the gradients of different samples more closely during training. Furthermore, we demonstrate that the representation similarity between the inserted LLM block and the adjacent ViT block influences performance, with greater similarity tending to yield larger positive gains. Through these findings, we can justify the selection of suitable LLM blocks to be inserted at appropriate positions, and introduce additional gradient backpropagation paths by incorporating LLM blocks, could improve the performance of vanilla ViT through the rectification effect of gradient consistency during the training process, without the need to add LLM blocks during inference. Our experiments demonstrate the effectiveness of this strategy, making the practical application of the gradient rectification effect feasible.

Stochastic Block Models (SBMs) are a popular approach to modeling single real-world graphs. The key idea of SBMs is to partition the vertices of the graph into blocks with similar edge densities within, as well as between different blocks. However, what if we are given not one but multiple graphs that are unaligned and of different sizes? How can we find out if these graphs share blocks with similar connectivity structures? In this paper, we propose the shared stochastic block modeling (SSBM) problem, in which we model n graphs using SBMs that share parameters of s blocks. We show that fitting an SSBM is NP-hard, and consider two approaches to fit good models in practice. In the first, we directly maximize the likelihood of the shared model using a Markov chain Monte Carlo algorithm. In the second, we first fit an SBM for each graph and then select which blocks to share. We propose an integer linear program to find the optimal shared blocks and to scale to large numbers of blocks, we propose a fast greedy algorithm. Through extensive empirical evaluation on synthetic and real-world data, we show that our methods work well in practice.

Predicting the future impact of newly published articles is pivotal for advancing scientific discovery in an era of unprecedented scholarly expansion. This paper introduces a promising approach, leveraging the capabilities of LLMs to predict the future impact of newborn articles solely based on titles and abstracts. Breaking away from traditional methods heavily reliant on external data, we propose fine-tuning the LLM to uncover the intrinsic semantic patterns shared by highly impactful articles from a vast collection of text-score pairs. These semantic features are further utilized to predict the proposed indicator, TNCSIsp, which incorporates favorable normalization properties across value, field, and time. To facilitate parameter-efficient fine-tuning of the LLM, we have also meticulously curated a dataset containing over 12,000 entries, each annotated with titles, abstracts, and their corresponding TNCSIsp values. Experimental results reveal an MAE of 0.216 and an NDCG@20 of 0.901, setting new benchmarks in predicting the impact of newborn articles. Finally, we present a real-world application example for predicting the impact of newborn journal articles to demonstrate its noteworthy practical value. Overall, our findings challenge existing paradigms and propose a shift towards a more content-focused prediction of academic impact, offering new insights for article impact prediction.

Estimation-of-distribution algorithms (EDAs) are optimization algorithms that learn a distribution from which good solutions can be sampled easily. A key parameter of most EDAs is the sample size (population size). Too small values lead to the undesired effect of genetic drift, while larger values slow down the process. Building on a quantitative analysis of how the population size leads to genetic drift, we design a smart-restart mechanism for EDAs. By stopping runs when the risk for genetic drift is high, it automatically runs the EDA in good parameter regimes. Via a mathematical runtime analysis, we prove a general performance guarantee for this smart-restart scheme. For many situations where the optimal parameter values are known, this shows that the restart scheme automatically finds these optimal values, leading to the asymptotically optimal performance. We also conduct an extensive experimental analysis. On four classic benchmarks, the smart-restart scheme leads to a performance close to the one obtainable with optimal parameter values. We also conduct experiments with PBIL (cross-entropy algorithm) on the max-cut problem and the bipartition problem. Again, the smart-restart mechanism finds much better values for the population size than those suggested in the literature, leading to a much better performance.

Deep learning has significantly enhanced survival prediction using whole slide images (WSIs) by adopting a two-stage learning paradigm: WSI preparation and patient-level prediction. While existing research generally concentrates on developing advanced patient-level prediction modules, the critical importance of WSI preparation has been largely overlooked. In practice, WSI preparation is influenced by numerous factors, including tissue heterogeneity, sampling strategies, and technical considerations. These uncontrollable external factors incur variability in the number of WSIs among patients, introducing significant bias and resulting in inferior performance for patients with few WSIs. To address this challenge, we propose a novel approach named WSI-Diffusion. Unlike existing WSI generation models that produce augmented versions of input WSIs, our method generates entirely new WSIs in representation space to serve as complementary data. WSIDiffusion employs a two-stage hierarchical diffusion process. Two novel modules, WSI-level and patch-level Diffusers are designed to capture complex correlations between WSIs and patches. The generated WSIs are integrated as supplementary data, and a light patient-level prediction module is then trained for survival prediction. Experimental results across five datasets demonstrate the superiority of our proposal.

Intuitively, an ideal collaborative filtering (CF) model should learn from users' full rankings over all items to make optimal top-K recommendations. Due to the absence of such full rankings in practice, most CF models rely on pairwise loss functions to approximate full rankings, resulting in an immense performance gap. In this paper, we provide a novel analysis using the multiple ordinal classification concept to reveal the inevitable gap between a pairwise approximation and the ideal case. However, bridging the gap in practice encounters two formidable challenges: (1) none of the real-world datasets contains full ranking information; (2) there does not exist a loss function that is capable of consuming ranking information. To overcome these challenges, we propose a pseudo-ranking paradigm (PRP) that addresses the lack of ranking information by introducing pseudo-rankings supervised by an original noise injection mechanism. Additionally, we put forward a new ranking loss function designed to handle ranking information effectively. To ensure our method's robustness against potential inaccuracies in pseudo-rankings, we equip the ranking loss function with a gradient-based confidence mechanism to detect and mitigate abnormal gradients. Extensive experiments on four real-world datasets demonstrate that PRP significantly outperforms state-of-the-art methods.

Pretrained Language Models (PLMs) have become the de facto starting point for fine-tuning on downstream tasks. However, as model sizes continue to increase, traditional fine-tuning of all parameters becomes challenging. To address this, parameter-efficient fine-tuning (PEFT) methods have gained popularity as a means to adapt PLMs effectively. In parallel, recent studies have revealed the presence of activation sparsity within the intermediate outputs of the multilayer perceptron (MLP) blocks in transformers. Low activation density enables efficient model inference on sparsity-aware hardware. Building upon this insight, in this work, we propose a novel density loss that encourages higher activation sparsity (equivalently, lower activation density) in the pre-trained models. We demonstrate the effectiveness of our approach by utilizing mainstream PEFT techniques, including QLoRA, LoRA, Adapter, and Prompt/Prefix Tuning, to facilitate efficient model adaptation across diverse downstream tasks. Experiments show that our proposed method, DEFT (Density-Efficient Fine-Tuning), can consistently reduce activation density by up to 44.94% on RoBERTa (Large) and by 53.19 (encoder density) and 90.60% (decoder density) on Flan-T5-XXL (11B) compared to PEFT, using GLUE and QA (SQuAD) benchmarks respectively while maintaining competitive performance on downstream tasks. We also introduce ADA-DEFT, an adaptive variant of our DEFT approach, which achieves significant memory and runtime savings during inference for large models. For instance, ADA-DEFT reduces runtime by 8.75% and memory usage by 16.78% in Flan-T5-XL and by 2.79% and 2.54%, respectively, in Flan-T5- XXL. Additionally, we showcase that DEFT works complementarily with quantized and pruned models.

We focus on the classification problem with a separable dataset, one of the most important and classical problems from machine learning. The standard approach to this task is logistic regression with gradient descent (LR+GD). Recent studies have observed that LR+GD can find a solution with arbitrarily large step sizes, defying conventional optimization theory. Our work investigates this phenomenon and makes three interconnected key observations about LR+GD with large step sizes. First, we find a remarkably simple explanation of why LR+GD with large step sizes solves the classification problem: LR+GD reduces to a batch version of the celebrated perceptron algorithm when the step size tends to infinity. Second, we observe that larger step sizes lead LR+GD to higher logistic losses when it tends to the perceptron algorithm, but larger step sizes also lead to faster convergence to a solution for the classification problem, meaning that logistic loss is an unreliable metric of the proximity to a solution. Surprisingly, high loss values can actually indicate faster convergence. Third, since the convergence rate in terms of loss function values of LR+GD is unreliable, we examine the iteration complexity required by LR+GD with large step sizes to solve the classification problem and prove that this complexity is suboptimal. To address this, we propose a new method, Normalized LR+GD – based on the connection between LR+GD and the perceptron algorithm – with much better theoretical guarantees.