AAAI 2025 Papers

In this paper, we propose an efficient, fast, and versatile distillation method to accelerate the generation of pre-trained diffusion models. The method reaches state-of-the-art performances in terms of FID and CLIP-Score for few steps image generation on the COCO2014 and COCO2017 datasets, while requiring only several GPU hours of training and fewer trainable parameters than existing methods. In addition to its efficiency, the versatility of the method is also exposed across several tasks such as *text-to-image*, *inpainting*, *face-swapping*, *super-resolution* and using different backbones such as UNet-based denoisers (SD1.5, SDXL), DiT (Pixart) and MMDiT (SD3), as well as adapters. In all cases, the method allowed to reduce drastically the number of sampling steps while maintaining very high-quality image generation.

Recent progress in Multimodal Large Language Models (MLLMs) often use large image tokens to compensate the visual shortcoming of MLLMs, which not only exhibits obvious redundancy but also greatly exacerbates the already high computation. Token pruning is an effective solution for speeding up MLLMs, but when and how to drop tokens still remains a challenge. In this paper, we propose a novel and training-free approach for the effective visual token pruning of MLLMs, termed FitPrune, which can quickly produce a complete pruning recipe for MLLMs according to a pre-defined budget. Specifically, FitPrune considers token pruning as a statistical problem of MLLM and its objective is to find out an optimal pruning scheme that can minimize the divergence of the attention distributions before and after pruning. In practice, FitPrune can be quickly accomplished based on the attention statistics from a small batch of inference data, avoiding the expensive trials of MLLMs. According to the pruning recipe, an MLLM can directly remove the redundant visual tokens of different examples during inference. To validate FitPrune, we apply it to a set of recent MLLMs, including LLaVA-1.5, LLaVA-HR and LLaVA-NEXT, and conduct extensive experiments on a set of benchmarks. The experimental results show that our FitPrune can not only reduce the computational complexity to a large extent, while retaining high performance, e.g., -54.9% FLOPs for LLaVA-NEXT with only 0.5% accuracy drop. Notably, the pruning recipe can be obtained in about 5 minutes.

Federated bilevel optimization (FBO) has garnered significant attention lately, driven by its promising applications in meta-learning and hyperparameter optimization. Existing algorithms generally aim to approximate the gradient of the upper-level objective function (hypergradient) in the federated setting. However, because of the nonlinearity of the hypergradient and client drift, they often involve complicated computations. These computations, like multiple optimization sub-loops and second-order derivative evaluations, end up with significant memory consumption and high computational costs. In this paper, we propose a computationally and memory-efficient FBO algorithm named MemFBO. MemFBO features a fully single-loop structure with all involved variables updated simultaneously, and uses only first-order gradient information for all local updates. We show that MemFBO exhibits a linear convergence speedup with milder assumptions in both partial and full client participation scenarios. We further implement MemFBO in a novel FBO application for federated data cleaning. Our experiments, conducted on this application and federated hyper-representation, demonstrate the effectiveness of the proposed algorithm.

First-order linear temporal logic (FOLTL) is a flexible and expressive formalism capable of naturally describing complex behaviors and properties. Although the logic is in general highly undecidable, the idea of using it as a specification language for the verification of complex infinite-state systems is appealing. However, a missing piece, which has proved to be an invaluable tool in dealing with other temporal logics, is an automaton model capable of capturing the logic. In this paper we address this issue, by defining and studying such a model, which we call first-order automaton. We define this very general class of automata, and the corresponding notion of regular first-order language (of finite words), showing their closure under most language-theoretic operations. We show how they can capture any FOLTL formula over finite words, over any signature and theory, and provide sufficient conditions for the semi-decidability of their non-emptiness problem. Then, to show the usefulness of the formalism, we prove the decidability of monodic FOLTL, a classic result known in the literature, with a simpler and direct proof.

Adversarial training (AT) incurs significant computational overhead, leading to growing interest in designing inherently robust architectures. We demonstrate that a carefully designed first layer of the neural network can serve as an implicit adversarial noise filter (ANF). This filter is created using a combination of large kernel size, increased convolution filters, and a maxpool operation. We show that integrating this filter as the first layer in architectures such as ResNet, VGG, and EfficientNet results in adversarially robust networks. Our approach achieves higher adversarial accuracies than existing natively robust architectures without AT and is competitive with adversarial-trained architectures across a wide range of datasets. Supporting our findings, we show that (a) the decision regions for our method have better margins, (b) the visualized loss surfaces are smoother, (c) the modified peak signal-to-noise ratio (mPSNR) values at the output of the ANF are higher, (d) high-frequency components are more attenuated, and (e) architectures incorporating ANF exhibit better denoising in Gaussian noise compared to baseline architectures.

Medical Visual Question Answering (MedVQA) serves as an automated medical assistant, capable of answering patient queries and aiding physician diagnoses based on medical images and questions. Recent advancements have shown that incorporating Large Language Models (LLMs) into MedVQA tasks significantly enhances the capability for answer generation. However, for tasks requiring fine-grained organ-level precise localization, relying solely on language prompts struggles to accurately locate relevant regions within medical images due to substantial background noise. To address this challenge, we explore the use of visual prompts in MedVQA tasks for the first time and propose fine-grained adaptive visual prompts to enhance generative MedVQA. Specifically, we introduce an Adaptive Visual Prompt Creator that adaptively generates region-level visual prompts based on image characteristics of various organs, providing fine-grained references for LLMs during answer retrieval and generation from the medical domain, thereby improving the model's precise cross-modal localization capabilities on original images. Furthermore, we incorporate a Hierarchical Answer Generator with Parameter-Efficient Fine-Tuning (PEFT) techniques, significantly enhancing the model's understanding of spatial and contextual information with minimal parameter increase, promoting the alignment of representation learning with the medical space. Extensive experiments on VQA-RAD, SLAKE, and DME datasets validate the effectiveness of our proposed method, demonstrating its potential in generative MedVQA.

Recent advancements in large language models (LLMs) have broadened their application scope but revealed challenges in balancing capabilities across general knowledge, coding, and mathematics. To address this, we introduce a Collaborative and Semantic Experts (CoE) approach for supervised fine-tuning (SFT), which employs a two-phase training strategy. Initially, expert training fine-tunes the feed-forward network on specialized datasets, developing distinct experts in targeted domains. Subsequently, expert leveraging synthesizes these trained experts into a structured model with semantic guidance to activate specific experts, enhancing performance and interpretability. Evaluations on comprehensive benchmarks across MMLU, HumanEval, GSM8K, MT-Bench, and AlpacaEval confirm CoE's efficacy, demonstrating improved performance and expert collaboration in diverse tasks, significantly outperforming traditional SFT methods.

Existing Salient Object Ranking (SOR) aims to infer ranking of salient objects based on their saliency degree. However, it tends to only focus on salient objects while neglecting non-salient ones. This coarse-grained ranking limits the performance of downstream tasks. For instance, in image retrieval tasks, focusing solely on the relationship between salient objects is insufficient for achieving fine-grained scene analysis, which may result in retrieved results that do not satisfy user requirements. High-quality retrieval requires fine-grained analysis, making it essential to rank non-salient objects. Based on this need, we propose a new task: Fine-grained Object Importance Ranking in 360 Scenes (FOIR-360), which focus on predicting the relative importance of "ALL objects'' at the instance-level. Our task takes into account all objects, allowing us to refine the original "coarse-grained'' to a "fine-grained'' level. Currently, the main challenge for this new task is the lack of supervised data for model training or even for model testing. Therefore, we propose a novel weakly supervised method to address the shortage of datasets. Furthermore, to the best of our knowledge, there is no existing suitable annotation protocol for this new task. The main reason is that annotating fine-grained rankings is extremely difficult, especially in panoramic scenes that contain numerous instances where even humans are unable to determine which one is more important than others. As the first attempt, we introduce a new annotation protocol designed to highlight the ranking of objects that are non-salient yet still important. Based on this protocol, we construct the first fine-grained 360Rank dataset. In summary, all these new task, weakly supervised method, annotation protocol, and dataset have the potential to drive advancements in the field.

AI becomes increasingly vital for telecom industry, as the burgeoning complexity of upcoming mobile communication networks places immense pressure on network operators. While there is a growing consensus that intelligent network self-driving holds the key, it heavily relies on expert experience and knowledge extracted from network data. In an effort to facilitate convenient analytics and utilization of wireless big data, we introduce the concept of knowledge graphs into the field of mobile networks, giving rise to what we term as wireless data knowledge graphs (WDKGs). However, the heterogeneous and dynamic nature of communication networks renders manual WDKG construction both prohibitively costly and error-prone, presenting a fundamental challenge. In this context, we propose an unsupervised data-and-model driven graph structure learning (DMGSL) framework, aimed at automating WDKG refinement and updating. Tackling WDKG heterogeneity involves stratifying the network into homogeneous layers and refining it at a finer granularity. Furthermore, to capture WDKG dynamics effectively, we segment the network into static snapshots based on the coherence time and harness the power of recurrent neural networks to incorporate historical information. Extensive experiments conducted on the established WDKG demonstrate the superiority of the DMGSL over the baselines, particularly in terms of node classification accuracy.

Multivariate time series forecasting is crucial across various industries, where accurate extraction of complex periodic and trend components can significantly enhance prediction performance. However, existing models often struggle to capture these intricate patterns. To address these challenges, we propose FilterTS, a novel forecasting model that utilizes specialized filtering techniques based on the frequency domain. FilterTS introduces a Dynamic Cross-Variable Filtering Module, a key innovation that dynamically leverages other variables as filters to extract and reinforce shared variable frequency components across variables in multivariate time series. Additionally, a Static Global Filtering Module captures stable frequency components, identified throughout the entire training set. Moreover, the model is built in the frequency domain, converting time-domain convolutions into frequency-domain multiplicative operations to enhance computational efficiency. Extensive experimental results on eight real-world datasets have demonstrated that FilterTS significantly outperforms existing methods in terms of prediction accuracy and computational efficiency.

Recent Anomaly Detection (AD) methods have achieved great success with In-Distribution (ID) data. However, real-world data often exhibits distribution shift, causing huge performance decay on traditional AD methods. From this perspective, few previous work has explored AD with distribution shift, and the distribution-invariant normality learning has been proposed based on the Reverse Distillation (RD) framework. However, we observe the misalignment issue between the teacher and the student network that causes detection failure, thereby propose FiCo, Filter or Compensate, to address the distribution shift issue in AD. FiCo firstly compensates the distribution-specific information to reduce the misalignment between the teacher and student network via the Distribution-Specific Compensation (DiSCo) module, and secondly filters all abnormal information to capture distribution-invariant normality with the Distribution-Invariant Filter (DiIFi) module. Extensive experiments on three different AD benchmarks demonstrate the effectiveness of FiCo, which outperforms all existing state-of-the-art (SOTA) methods, and even achieves better results on the ID scenario compared with RD-based methods.

Continual Semantic Parsing (CSP) aims to train parsers to convert natural language questions into SQL across tasks with limited annotated examples, adapting to dynamically updated databases in real-world scenarios. Previous studies mitigate this challenge by replaying historical data or employing parameter-efficient tuning (PET), but they often violate data privacy or rely on ideal continual learning settings. To address these issues, we propose a new Large Language Model (LLM)-Enhanced Continuous Semantic Parsing method, named LECSP, which alleviates forgetting while encouraging generalization, without requiring real data replay or ideal settings. Specifically, it first analyzes the commonalities and differences between tasks from the SQL syntax perspective to guide LLMs in reconstructing key memories and improving memory accuracy through calibration. Then, it uses a task-aware dual-teacher distillation framework to promote the accumulation and transfer of knowledge during sequential training. Experimental results on two CSP benchmarks show that our method significantly outperforms existing methods, even those utilizing data replay or ideal settings. Additionally, we achieve generalization performance beyond upper limits, better adapting to unseen tasks.

Large Vision-Language Models (LVLMs) signify a groundbreaking paradigm shift within the Artificial Intelligence (AI) community, extending beyond the capabilities of Large Language Models (LLMs) by assimilating additional modalities (e.g., images). Despite this advancement, the safety of LVLMs remains adequately underexplored, with a potential overreliance on the safety assurances purported by their underlying LLMs. In this paper, we propose FigStep, a straightforward yet effective black-box jailbreak algorithm against LVLMs. Instead of feeding textual harmful instructions directly, FigStep converts the prohibited content into images through typography to bypass the safety alignment. The experimental results indicate that FigStep can achieve an average attack success rate of 82.50% on six promising open-source LVLMs. Not merely to demonstrate the efficacy of FigStep, we conduct comprehensive ablation studies and analyze the distribution of the semantic embeddings to uncover that the reason behind the success of FigStep is the deficiency of safety alignment for visual embeddings. Moreover, we compare FigStep with five text-only jailbreaks and four image-based jailbreaks to demonstrate the superiority of FigStep, i.e., negligible attack costs and better attack performance. Above all, our work reveals that current LVLMs are vulnerable to jailbreak attacks, which highlights the necessity of novel cross-modality safety alignment techniques.

Crop field boundaries are foundational datasets for agricultural monitoring and assessments but are expensive to collect manually. Machine learning (ML) methods for automatically extracting field boundaries from remotely sensed images could help realize the demand for these datasets at a global scale. However, current ML methods for field instance segmentation lack sufficient geographic coverage, accuracy, and generalization capabilities. Further, research on improving ML methods is restricted by the lack of labeled datasets representing the diversity of global agricultural fields. We present Fields of The World (FTW)---a novel ML benchmark dataset for agricultural field instance segmentation spanning 24 countries on four continents (Europe, Africa, Asia, and South America). FTW is an order of magnitude larger than previous datasets with 70,462 samples, each containing instance and semantic segmentation masks paired with multi-date, multi-spectral Sentinel-2 satellite images. We provide results from baseline models for the new FTW benchmark, show that models trained on FTW have better zero-shot and fine-tuning performance in held-out countries than models that aren't pre-trained with diverse datasets, and show positive qualitative zero-shot results of FTW models in a real-world scenario -- running on Sentinel-2 scenes over Ethiopia.

Few-shot Continual Relation Extraction is a crucial challenge for enabling AI systems to identify and adapt to evolving relationships in dynamic real-world domains. Traditional memory-based approaches often overfit to limited samples, failing to reinforce old knowledge, with the scarcity of data in few-shot scenarios further exacerbating these issues by hindering effective data augmentation in the latent space. In this paper, we propose a novel retrieval-based solution, starting with a large language model to generate descriptions for each relation. From these descriptions, we introduce a bi-encoder retrieval training paradigm to enrich both sample and class representation learning. Leveraging these enhanced representations, we design a retrieval-based prediction method where each sample "retrieves" the best fitting relation via a reciprocal rank fusion score that integrates both relation description vectors and class prototypes. Extensive experiments on multiple datasets demonstrate that our method significantly advances the state-of-the-art by maintaining robust performance across sequential tasks, effectively addressing catastrophic forgetting.

Audio-Visual Semantic Segmentation (AVSS) has gained significant attention in the multi-modal domain, aiming to segment video objects that produce specific sounds in the corresponding audio. Despite notable progress, existing methods still struggle to handle new classes not included in the original training set. To this end, we introduce Few-Shot Incremental Learning (FSIL) to the AVSS task, which seeks to seamlessly integrate new classes with limited incremental samples while preserving the knowledge of old classes. Two challenges arise in this new setting: (1) To reduce labeling costs, old classes within the incremental samples are treated as background, similar to silent objects. Training the model directly with background annotations may worsen the loss of distinctive knowledge about old classes, such as their outlines and sounds. (2) Most existing models adopt early cross-modal fusion with a single-tower design, incorporating more characteristics into class representations, which impedes knowledge transfer between classes based on similarity. To address these issues, we propose a Few-shot Incremental learning framework via class-centric foregrouNd aggreGation and dual-tower knowlEdge tRansfer (FINGER) for the AVSS task, which comprises two targeted modules: (1) The class-centric foreground aggregation gathers class-specific features for each foreground class while disregarding background features. The background class is excluded during training and inferred from the foreground predictions. (2) The dual-tower knowledge transfer postpones cross-modal fusion to separately conduct knowledge transfer for each modality. Extensive experiments validate the effectiveness of the FINGER model, significantly surpassing state-of-the-art methods.

Recently, a number of effective methods have been proposed to tackle the challenging task of Few-Shot Fine-Grained Image Classification (FS-FGIC). However, how to fully leverage the backbone network to discover and extract detailed features to generate more discriminative class prototypes, as well as how to accurately model the similarity relationship between query samples and the class prototypes, are still issues to be further considered. Therefore, we propose a novel progreSsively featUre refInement and conTinuous rElationship moDeling method, SUITED for short, to address these two issues existing in the State-of-the-Art FS-FGIC methods. Specifically, we design the Progressive Feature Refinement Module (PFRM) to fully exploit the backbone network's progressive feature extraction capabilities, forming multi-scale feature representations to further enhance discriminative features. Then, the Continuous Relationship Modeling Module (CRMM) is proposed to capture the dependencies between query samples and the corresponding class prototypes, achieving precise optimization of the distances among corresponding sample points in the feature space. We conducted extensive experiments on five fine-grained benchmark datasets, and the experimental results demonstrate that the proposed method is comprehensively ahead of the existing State-of-the-Art methods.

Learned image compression (LIC) has achieved state-of-the-art rate-distortion performance, deemed promising for next-generation image compression techniques. However, pre-trained LIC models usually suffer from significant performance degradation when applied to out-of-training-domain images, implying their poor generalization capabilities. To tackle this problem, we propose a few-shot domain adaptation method for LIC by integrating plug-and-play adapters into pre-trained models. Drawing inspiration from the analogy between latent channels and frequency components, we examine domain gaps in LIC and observe that out-of-training-domain images disrupt pre-trained channel-wise decomposition. Consequently, we introduce a method for channel-wise re-allocation using convolution-based adapters and low-rank adapters, which are lightweight and compatible to mainstream LIC schemes. Extensive experiments across multiple domains and multiple representative LIC schemes demonstrate that our method significantly enhances pre-trained models, achieving comparable performance to H.266/VVC intra coding with merely 25 target-domain samples. Additionally, our method matches the performance of full-model finetune while transmitting fewer than 2% of the parameters.

Audio-Visual Learning (AVL) aims at the audio-visual perception with both audio and vision modalities. AVL also suffers from data insufficiency in many applications as with other unimodal tasks. Concurrently, AVL often needs to continuously learn over time rather than all knowledge simultaneously. Considering the above two perspectives, our work mainly focuses on benchmarking the unexplored Few-Shot Audio-Visual Class-Incremental Learning (FS-AVCIL), i.e., continually perceiving novel categories described by a limited number of labeled examples with audio and visual modalities. Firstly, we provide the detailed task configuration together with a thorough analysis of the challenges in FS-AVCIL: (1) how to efficiently learn and fuse multimodal information with limited labeled examples; and (2) how to alleviate catastrophic forgetting cross-modal semantic correlations with limited data. Then, we propose an efficient framework based on Vision Transformer to solve FS-AVCIL. This framework contains two parts: temporal-residual prompting for audio-visual synergy adapter and temporal prompt regularization. Specifically, temporal-residual prompting is incorporated into the audio-visual adapter to efficiently finetune the pre-trained foundation model with limited data and capture audio-visual correlation by learning temporal-relevant prompts. Besides, we regularize temporal-relevant prompts to memorize previous knowledge by fully using the temporal knowledge from various perspectives. This framework is validated in audio-visual classification tasks under the FS-AVCIL scenario, and extensive experiments demonstrate its superior performance.