These 1064 ICLR 2024 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ICLR 2024 paper, free trial on arXivSub.
$\alpha$TC-VAE: On the relationship between Disentanglement and Diversity
Cristian Meo (Delft University of Technology), Justin Dauwels (Delft University of Technology)
π― What it does: A new variational autoencoder, Ξ±-TCVAE, is proposed, which achieves better separation and information retention by optimizing the lower bound of the joint total correlation.
π― What it does: This paper proposes a diffusion model defined on infinite-dimensional Hilbert space, β-Diff, and trains it by learning continuous functions that can generate high-quality images at any resolution through random sampling of subset coordinates.
π― What it does: This paper proposes a core set (coreset) selection method based on graph message passing called D2PRUNING, which aims to balance data diversity and difficulty.
π― What it does: In the self-supervised pre-training of 3D point clouds, MaskFeat3D is proposed, which predicts the surface normal vectors and curvature variations of masked points under the Masked Autoencoder (MAE) framework, rather than recovering point coordinates.
π― What it does: A branching decoder is proposed, capable of generating multiple text sequences in parallel during a single decoding process, avoiding the concatenation order bias of traditional sequential decoders.
Mehdi Fatemi (Microsoft Research), Sindhu C. M. Gowda
CodeReinforcement LearningTime SeriesSequential
π― What it does: A causal inference framework based on dynamic processes is proposed, which uses reinforcement learning to directly estimate the expected grit and reachability of events from observation sequences, and quantifies the causal contributions of each state/action component through decomposition formulas.
A Framework for Inference Inspired by Human Memory Mechanisms
Xiangyu Zeng (University of Electronic Science and Technology of China), Zhicheng Zhang (University of Electronic Science and Technology of China)
CodeTransformerText
π― What it does: Proposes the PMI framework: a three-module structure of perception, memory, and reasoning, constructing a dual-layer structure of working memory (WM) and long-term memory (LTM), and achieving information updating and integration through competitive writing, outer product association, and content retrieval.
π― What it does: A Bayesian Optimization framework named ColaBO is proposed, allowing users to inject prior knowledge about function properties (such as optimal location, optimal value, or preference relations) into the posterior distribution of the Gaussian Process, thereby directly reflecting user experience in sampling and acquisition functions.
π― What it does: A Meta-Policy knowledge distillation framework called MPDistil has been developed, allowing teachers and students to collaborate and compete during the distillation process, thereby enhancing the performance of both.
A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation
Zhengbo Wang (University of Science and Technology of China), Tieniu Tan (Chinese Academy of Sciences)
CodeClassificationDomain AdaptationVision Language ModelContrastive LearningImage
π― What it does: This paper proposes a training-free adaptation method for CLIP, which directly estimates class means and covariances from a small number of samples using Gaussian Discriminant Analysis (GDA), constructs a linear classifier, and integrates it with the original zero-shot classifier of CLIP.
A Lie Group Approach to Riemannian Batch Normalization
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
CodeTime SeriesSequentialBenchmark
π― What it does: A unified Lie group batch normalization framework, LieBN, is proposed, which can control the mean and variance on any left-invariant Lie group, and implement batch normalization of various metrics on SPD manifolds through parameterized deformed Lie groups.
A Multi-Level Framework for Accelerating Training Transformer Models
Longwei Zou (Tsinghua University), Yangdong Deng (Tsinghua University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A multi-level training framework is proposed, which significantly reduces the training cost of Transformer models by utilizing the coarse and fine tuning of model width and depth.
A Neural Framework for Generalized Causal Sensitivity Analysis
Dennis Frauen (LMU Munich), Mihaela van der Schaar (University of Cambridge)
CodeFlow-based ModelTabularBiomedical DataElectronic Health Records
π― What it does: This paper proposes NEURALCSA, a general causal sensitivity analysis framework based on neural networks, which can provide effective upper and lower bounds under different sensitivity models, types of treatments (discrete/continuous), and causal queries (such as CATE, distribution effects, multiple outcomes).
A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models
Haoran Xu (Johns Hopkins University), Hany Hassan Awadalla (Microsoft)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Two-stage fine-tuning of LLaMA-2: first, fine-tune with multilingual monolingual data to enhance cross-lingual capabilities, and then fine-tune with a small amount of high-quality parallel data to obtain the ALMA model, which significantly improves performance on multilingual translation tasks.
π― What it does: This paper proposes a plugin-based prior (PnP) deformable image registration network called PIRATE, and further fine-tunes it using a deep equilibrium model (DEQ) to create PIRATE+, achieving a collaborative optimization of data consistency constraints in the registration field and learned CNN priors.
π― What it does: A new primal-dual method (ACVI) is proposed and analyzed for solving variational inequalities (VI) with general constraints, and it is improved into two variants: I-ACVI for approximately solving subproblems and P-ACVI for simple inequality constraints.
A Probabilistic Framework for Modular Continual Learning
Lazar Valkov (MIT IBM Watson AI Lab), Charles Sutton (University of Edinburgh)
CodeSequentialBenchmark
π― What it does: A modular continual learning framework PICLE is proposed, which uses probabilistic models for rapid evaluation of module combinations.
A Semantic Invariant Robust Watermark for Large Language Models
Aiwei Liu (Tsinghua University), Lijie Wen (Tsinghua University)
CodeAdversarial AttackTransformerLarge Language ModelText
π― What it does: A robust watermarking algorithm based on contextual semantic invariance is proposed, which uses an embedding model to generate watermark logits, maintaining detection accuracy under semantic-preserving attacks such as text rewriting and synonym replacement.
A Simple and Effective Pruning Approach for Large Language Models
Mingjie Sun (Carnegie Mellon University), J Zico Kolter
CodeTransformerLarge Language ModelText
π― What it does: A pruning method for large language models (LLM) called Wanda is proposed, which evaluates importance by multiplying the absolute value of weights with the L2 norm of the corresponding input activations for each output dimension, directly setting low-importance weights to zero without the need for retraining.
A Simple Interpretable Transformer for Fine-Grained Image Classification and Analysis
DIPANJYOTI PAUL, Wei-Lun Chao
CodeClassificationExplainability and InterpretabilityTransformerImage
π― What it does: An interpretable Transformer classifier called INTR is proposed, which allows each category to actively search for itself in the image using class-specific queries, thereby naturally generating attention heatmaps during inference.
π― What it does: The statistical properties of the Wasserstein Autoencoder (WAE) on low-dimensional structured data are theoretically analyzed, deriving an upper bound for the expected excess risk and proving that this upper bound is related to the intrinsic dimension of the data (Minkowski dimension), rather than the dimension of the high-dimensional feature space.
A Topological Perspective on Demystifying GNN-Based Link Prediction Performance
Yu Wang (Vanderbilt University), Tyler Derr (Vanderbilt University)
CodeGraph Neural NetworkGraph
π― What it does: This study investigates the performance differences of different nodes in the link prediction (LP) task using GNNs, proposing the Topological Concentration (TC) metric to measure the relationship between local topology of nodes and LP performance, and explores the computation, approximation, and impact of TC on performance.
A Unified Framework for Bayesian Optimization under Contextual Uncertainty
Sebastian Shenghong Tay (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeOptimizationTabularTime Series
π― What it does: This paper proposes a unified framework for handling Bayesian optimization under context uncertainty (BOCU) and presents a general Thompson sampling algorithm capable of optimizing any uncertain objective.
π― What it does: A Unified Sampling Framework (USF) is designed, and the sampling quality of the diffusion model is optimized through an automatic search method (S3) with very few NFE.
A Versatile Causal Discovery Framework to Allow Causally-Related Hidden Variables
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
CodeTabular
π― What it does: A variable causal discovery framework RLCD is proposed, which utilizes the rank information of the covariance matrix to identify causal structures containing interrelated hidden variables.
π― What it does: An extended module of Transformer called Abstractor is proposed, which achieves explicit relational reasoning through relational cross-attention, enabling decoupling between object features and relational information, thereby improving sample efficiency in relational reasoning tasks.
Accelerating Data Generation for Neural Operators via Krylov Subspace Recycling
Hong Wang (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
CodeOptimizationComputational EfficiencyTabularPhysics Related
π― What it does: The SKR algorithm is proposed, utilizing sorting and Krylov subspace reuse to accelerate the solution of linear equations in the neural operator training data generation process.
π― What it does: In a federated continual learning environment with distribution heterogeneity and potentially unrelated tasks, an adaptive knowledge filtering mechanism using 'accurate forgetting' is proposed to maintain performance on previous tasks.
Accurate Retraining-free Pruning for Pretrained Encoder-based Language Models
Seungcheol Park (Seoul National University), U Kang (Seoul National University)
CodeCompressionTransformerLarge Language ModelText
π― What it does: This paper proposes K-prune, a structured pruning method that does not require retraining, for compressing pre-trained encoder-based language models.
Active Test-Time Adaptation: Theoretical Analyses and An Algorithm
Shurui Gui (Texas A and M University), Shuiwang Ji (Texas A and M University)
CodeDomain AdaptationImage
π― What it does: A new Active Testing Time Adaptation (ATTA) framework is proposed, which can adapt in real-time to the continuously changing test distribution under the condition of obtaining only a limited number of labeled samples, avoiding catastrophic forgetting.
Adapting Large Language Models via Reading Comprehension
Daixuan Cheng (Beijing Institute for General Artificial Intelligence), Furu Wei (Microsoft Research)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataFinance Related
π― What it does: Investigate the impact of continuous pre-training on domain-specific corpora for large language models and propose an adaptation method to convert raw corpora into reading comprehension texts.
Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts
Jian Xie (Fudan University), Yu Su (Ohio State University)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: The system evaluates the behavior of large language models in knowledge conflict scenarios and proposes a complete framework for constructing counter-memory.
Adaptive Federated Learning with Auto-Tuned Clients
Junhyung Lyle Kim (Rice University), Anastasios Kyrillidis (Rice University)
CodeFederated LearningImageText
π― What it does: This paper proposes Ξ-SGD, a distributed SGD algorithm that adaptively adjusts the learning rate for each client in a federated learning environment;
Adaptive Regularization of Representation Rank as an Implicit Constraint of Bellman Equation
Qiang He (Ruhr University Bochum), Setareh Maghsudi (Ruhr University Bochum)
CodeReinforcement LearningSequential
π― What it does: This paper proposes BEER (Bellman Equation-based automatic rank Regularizer) to automatically adjust the rank of representation layers in deep reinforcement learning, preventing the model from becoming overly complex or underfitting.
π― What it does: This paper proposes a self-training framework (ST-SGG) that generates pseudo-labels for unlabeled relational triples, utilizing large-scale unlabeled data to enhance the performance of scene graph generation models, particularly for fine-grained predicates.
π― What it does: A Vision Transformer based on adaptive window cropping (LMD-ViT) is proposed for local motion deblurring, focusing attention calculations only on blurred areas while preserving clear regions.
π― What it does: This paper proposes an Alternating Denoising Diffusion Process (ADDP) that learns a unified representation for both image generation and image recognition within a single model.
Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning
Mohamed Elsayed (University of Alberta), A. Rupam Mahmood (University of Alberta)
CodeOptimizationReinforcement LearningImage
π― What it does: Proposes Utility-based Perturbed Gradient Descent (UPGD), which addresses catastrophic forgetting and loss of plasticity in continual learning by utilizing weight importance gated gradients and injecting noise into unimportant weights.
π― What it does: A gradient backpropagation framework for diffusion generative models based on the adjoint sensitivity method, AdjointDPM, is proposed, which can simultaneously optimize model weights, conditional text, and noise states.
Advancing the Lower Bounds: an Accelerated, Stochastic, Second-order Method with Optimal Adaptation to Inexactness
Artem Agafonov (Moscow Institute of Physics and Technology), Martin TakΓ‘Δ (Laboratory for Information and Inference Systems IEM STI EPFL)
CodeOptimizationTabular
π― What it does: An accelerated stochastic second-order optimization method is proposed, which achieves optimal convergence speed even when the gradient and Hessian matrix are noisy (imprecise);
π― What it does: This paper proposes AdAutoMixup, an automatic mixup data augmentation framework that simultaneously trains a mixed sample generator and classifier through adversarial learning to enhance the generalization, calibration, and robustness of image classification.
Scott Sussex (ETH ZΓΌrich), Andreas Krause (ETH ZΓΌrich)
CodeOptimizationReinforcement LearningTabular
π― What it does: This paper proposes a framework for causal Bayesian optimization with adversarial intervention (ACBO) and presents the CBO-MW algorithm, which combines multiplicative weights with self-calibrating causal models. It derives an asymptotic lossless upper bound based on graph structures and designs a scalable distributed version, D-CBO-MW, on this basis.
π― What it does: A backdoor defense method based on feature map pruning, FMP, is proposed, which can detect and remove backdoor feature maps in the model without relying on trigger reproduction, and restore model performance through fine-tuning.
AffineQuant: Affine Transformation Quantization for Large Language Models
Yuexiao Ma (ByteDance Inc), Rongrong Ji (Xiamen University)
CodeTransformerLarge Language ModelText
π― What it does: A post-training quantization (AffineQuant) scheme based on reversible affine transformations is proposed for low-bit quantization of large-scale language models, significantly reducing quantization error and maintaining output consistency.
Xiao Liu (Tsinghua University), Jie Tang (Tsinghua University)
CodeLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
π― What it does: Constructed and released AgentBenchβa large-scale LLM-as-Agent evaluation benchmark covering 8 different environments (code, games, web), and systematically evaluated 29 types of LLMs (API and OSS).
π― What it does: A hybrid physics-data-driven model called AirPhyNet is proposed, which integrates diffusion and transport equations into a graph neural network for predicting urban PM2.5 air quality.
ALAM: Averaged Low-Precision Activation for Memory-Efficient Training of Transformer Models
Sunghyeon Woo (Seoul National University), Dongsuk Jeon (Seoul National University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A framework called ALAM based on average quantization and gradient norm variance is proposed to significantly compress the activation memory during Transformer training.
Algorithms for Caching and MTS with reduced number of predictions
Karim Ahmed Abdel Sadek, Marek Elias (Bocconi University)
CodeRecommendation SystemOptimizationTabularTime Series
π― What it does: Designed to enhance the performance of caching and Metric Task System (MTS) algorithms in scenarios with limited query counts or restricted query frequencies, using action predictions, and proposed the Follower-Robust combination algorithm.
π― What it does: This paper proposes the Alice Benchmark, providing a domain adaptation benchmark for object re-identification (person/vehicle) tasks, which includes editable synthetic source data and newly collected real target data.
π― What it does: This paper proposes to constrain the model output using the Lipschitz upper bound of GNNs to enhance individual fairness based on ranking in graph neural networks.
Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps
Mingxiao Li (KU Leuven), Marie-Francine Moens (KU Leuven)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: This study investigates the exposure bias problem that arises during the inference process of diffusion models (DPM) and proposes a Time-Shift sampler that alleviates this bias without the need to retrain the model.
AlpaGasus: Training a Better Alpaca with Fewer Data
Lichang Chen (University of Maryland), Hongxia Jin (Samsung Research America)
CodeLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: By using an automated LLM evaluator to filter out low-quality instruction/response pairs, only high-quality data is retained, and then instruction fine-tuning is performed on a smaller dataset to obtain a better LLM;
Alt-Text with Context: Improving Accessibility for Images on Twitter
Nikita Srivatsan (Carnegie Mellon University), Taylor Berg-Kirkpatrick (University of California San Diego)
CodeGenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: A model for generating alt-text based on multimodal input (image features + tweet text) is proposed, and a Twitter dataset consisting of 371k images, tweet texts, and alt-text is publicly released.
AmortizedPeriod: Attention-based Amortized Inference for Periodicity Identification
Hang Yu (Ant Group), Xinzhe Wang (Ant Group)
CodeAnomaly DetectionTransformerAuto EncoderTime SeriesFinance Related
π― What it does: A periodic detection method based on variational inference and attention networks, AmortizedPeriod, is proposed, which can simultaneously identify cycles, trends, noise, anomalies, and missing data.
Amortizing intractable inference in large language models
Edward J Hu, Nikolay Malkin (Mila - Quebec AI Institute, Universite de Montreal)
CodeTransformerLarge Language ModelReinforcement LearningFlow-based ModelText
π― What it does: This paper proposes the use of Generative Flow Networks (GFlowNet) to fine-tune large language models, enabling them to sample intractable posterior distributions, thereby achieving complex tasks such as chain reasoning and text completion.
An Analytical Solution to Gauss-Newton Loss for Direct Image Alignment
Sergei Solonets (Technical University of Munich), Daniel Cremers (Technical University of Munich)
CodeOptimizationSimultaneous Localization and MappingImage
π― What it does: Derived and implemented a closed-form solution for the Gauss-Newton loss to dynamically control the convergence domain of direct image alignment, enhancing robustness to initialization.
π― What it does: An efficient membership inference attack (PIA/PIAN) for diffusion models is proposed, which only requires two queries to determine whether a sample belongs to the training set.
An Emulator for Fine-tuning Large Language Models using Small Language Models
Eric Mitchell (Stanford University), Christopher D Manning
CodeKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: A framework called 'Emulated Fine-Tuning' (EFT) is proposed, which combines the benchmark log probabilities of the pre-trained model with the behavior delta of the fine-tuned model to achieve an independent combination of pre-training and fine-tuning knowledge at different scales; at the same time, the framework enables model up-scaling and reward interpolation during testing.
CodeObject DetectionAutonomous DrivingSafty and PrivacyComputational EfficiencyMultimodality
π― What it does: The HEAL framework is proposed to achieve open heterogeneous collaborative perception, supporting the low-cost integration of new types of sensors or models.
An Image Is Worth 1000 Lies: Transferability of Adversarial Images across Prompts on Vision-Language Models
Haochen Luo (University of Oxford), Philip Torr (University of Oxford)
CodeAdversarial AttackPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: This paper proposes the Cross-Prompt Attack (CroPA) and studies the adversarial transferability of visual-language models under different text prompts.
An interpretable error correction method for enhancing code-to-code translation
Min Xue (Heidelberg University), Marla Leuther (Heidelberg University)
CodeExplainability and InterpretabilityAI Code AssistantTransformerTextRetrieval-Augmented Generation
π― What it does: A kNN-based error correction method kNNβECD and kNNβECSm is proposed to improve the code translation of TransCoderβST and provide interpretable correction paths.
π― What it does: This study investigates how the under-representation of minority groups in Contrastive Learning (CL) leads to their feature representations clustering with similar majority groups, resulting in representation harm and allocation harm.
An Unforgeable Publicly Verifiable Watermark for Large Language Models
Aiwei Liu (Tsinghua University), Philip S. Yu (University of Illinois at Chicago)
CodeGenerationData SynthesisRecurrent Neural NetworkLarge Language ModelText
π― What it does: This paper proposes a publicly verifiable and unforgeable text watermarking algorithm for large language models, called UPV, which enables watermark detection without exposing the watermark generation key.
Analyzing and Mitigating Object Hallucination in Large Vision-Language Models
Yiyang Zhou (University of North Carolina at Chapel Hill), Huaxiu Yao (University of North Carolina at Chapel Hill)
CodeObject DetectionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: The LURE method is proposed, utilizing a post-processing mechanism to correct object misreporting in the text generated by large visual-language models.
π― What it does: This paper studies the impact of the feed-forward (FF) block and its surrounding residuals and layer normalization on input contextualization in Transformer layers by incorporating FF blocks into a refined analysis of attention maps, and proposes a method to decompose the nonlinear part of FF using Integrated Gradients (IG).
π― What it does: Transform the existing personalized text-to-image diffusion model into an animation generator, supporting model-agnostic fine-tuning.
π― What it does: This paper proposes ArchLock, which utilizes Neural Architecture Search (NAS) to reduce the transferability of DNNs at the architectural level, maintaining high performance on the source task while weakening the model's performance on potential target tasks.
Are Bert Family Good Instruction Followers? A Study on Their Potential And Limitations
yisheng xiao, Min Zhang (Soochow University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Using the encoder-only structure of the BERT family, a non-autoregressive generation scheme with dynamic mixed attention and Mask-Predict was designed for instruction fine-tuning of XML-RXL, resulting in Instruct-XMLR, which was validated in evaluations across multiple tasks, languages, and translations.
Are Human-generated Demonstrations Necessary for In-context Learning?
Rui Li (University of Science and Technology of China), Jiwei Li (Zhejiang University)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Proposes a Self-Reflection Prompt (SEC) framework that allows large language models to first generate demonstration examples suitable for each test sample without the need for human examples, and then use these demonstrations for reasoning.
Maxim Khanov (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: During the inference phase, a reward model is used to schedule the outputs of large language models to align with human preferences, avoiding the costly RLHF training.
π― What it does: The study quantifies uncertainty in the ROC curve of similarity scores and proposes a re-centered bootstrap method to construct confidence bands for assessing fairness in facial recognition.
At Which Training Stage Does Code Data Help LLMs Reasoning?
YINGWEI MA, Shanshan Li (National University of Defense Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: This study investigates the impact of incorporating code data at different training stages (pre-training and instruction fine-tuning) on the reasoning capabilities of large language models, evaluating its enhancement effects on general reasoning and code reasoning tasks.
Attention-based Iterative Decomposition for Tensor Product Representation
Taewon Park (Kyungpook National University), Minho Lee (Kyungpook National University)
CodeRecurrent Neural NetworkTransformerText
π― What it does: Proposes the AID module to improve the decomposition and binding process of Tensor Product Representation (TPR), enhancing the system's generalization ability.
π― What it does: This paper proposes the ACORM framework, which achieves behavioral heterogeneity and cooperation optimization in multi-agent reinforcement learning through role representation.
AttEXplore: Attribution for Explanation with model parameters eXploration
Zhiyu Zhu (University of Sydney), Flora D. Salim (University of New South Wales)
CodeExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerImage
π― What it does: A new interpretability method called AttEXplore is proposed, which explores model parameters during the interpretability process and utilizes transferable adversarial attacks to generate more representative samples, thereby more accurately locating the importance of input features.
π― What it does: Proposes the AuG-KD method to complete knowledge distillation in the absence of training data for the teacher model and in the presence of domain shift.
π― What it does: A Bayesian optimization method (ABS) that uses the advantage function as the mean of the Gaussian process is designed, combined with an adaptive aggregation of multiple Q-function estimators to search for deterministic policy parameters.
Augmenting Transformers with Recursively Composed Multi-grained Representations
Xiang Hu (Ant Group), Wei Wu (Ant Group)
CodeTransformerText
π― What it does: A recursive combination-enhanced Transformer model named ReCAT is proposed, which can explicitly model the hierarchical syntactic structure of the original text during learning and reasoning processes without relying on gold-standard trees.
AutomaTikZ: Text-Guided Synthesis of Scientific Vector Graphics with TikZ
Jonas Belouadi (Bielefeld University), Steffen Eger (University of Mannheim)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageText
π― What it does: A system called AutomaTiZ for automatically generating scientific graphics is proposed, utilizing TikZ as an intermediate language for generating text to vector graphics.
π― What it does: This paper presents AutoVP, an end-to-end automated visual prompting framework designed to select the optimal input scaling, visual prompts, pre-trained models, and label mapping strategies to adapt to downstream image classification tasks.
π― What it does: Utilize auxiliary task labels to enhance the performance of the main task while maintaining the computational and parameter costs of single tasks during inference.
Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency
Soumyadeep Pal (University of Alberta), Sijia Liu (Michigan State University)
CodeAnomaly DetectionOptimizationImage
π― What it does: This paper proposes a Mask-Aware SPC method based on two-layer optimization for automatically identifying data contaminated by backdoors without the need for clean samples and thresholds.
BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models
Zhen Xiang (University of Illinois Urbana-Champaign), Bo Li (University of Illinois Urbana-Champaign)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: We propose BadChain, a backdoor attack targeting LLMs using chain-of-thought (COT) prompts, which can induce the model to produce malicious answers upon the appearance of trigger words without accessing the training set or model parameters.
π― What it does: A backend defense method is proposed, which fine-tunes the attacked model on a small number of mislabelled clean samples to extract its backdoor functionality and construct a backdoor expert model for detecting and filtering backdoor inputs.
Balancing Act: Constraining Disparate Impact in Sparse Models
Meraj Hashemizadeh (Mila), Jose Gallego-Posada (Mila)
CodeOptimizationExplainability and InterpretabilityImage
π― What it does: A constraint optimization-based sparsification method CEAG is proposed, which directly controls the accuracy difference between the dense model and the sparse model across various subgroups, thereby alleviating the impact of differences caused by pruning.
BarLeRIa: An Efficient Tuning Framework for Referring Image Segmentation
Yaoming Wang (Shanghai Jiao Tong University), Qi Tian (Huawei Cloud)
CodeSegmentationTransformerVision Language ModelImage
π― What it does: A parameter-efficient fine-tuning framework named BarLeRIa is proposed to accomplish reference image segmentation tasks while keeping the pre-trained model frozen.
BatteryML: An Open-source Platform for Machine Learning on Battery Degradation
Han Zhang (Tsinghua University), Jiang Bian (Microsoft Research)
CodeTransformerTabularTime SeriesBenchmark
π― What it does: Built the BatteryML open-source platform, unifying data formats, feature engineering, and multi-model training, supporting battery degradation prediction.
Bayes Conditional Distribution Estimation for Knowledge Distillation Based on Conditional Mutual Information
Linfeng Ye (University of Waterloo), EN-HUI YANG
CodeKnowledge DistillationImage
π― What it does: By simultaneously maximizing the log-likelihood and conditional mutual information in the training of the teacher model, the estimation of the Bayesian conditional distribution is improved, thereby enhancing the effect of knowledge distillation.
π― What it does: This paper presents BayesDiff, a method for pixel-level uncertainty estimation of images generated by diffusion models through Bayesian inference. It utilizes uncertainty metrics to achieve low-quality image filtering, enhance generation diversity, and repair defects.
π― What it does: A Bayesian Neural Controlled Differential Equation (BNCDE) is proposed for estimating individualized treatment effects in continuous time, capable of outputting the posterior predictive distribution of potential outcomes.