arXivSub Start free trial

ICLR 2024 Papers — Page 16

International Conference on Learning Representations · 2260 papers

Optimal transport based adversarial patch to leverage large scale attack transferability

Pol Labarbarie (IRT SystemX), Milad Leyli-abadi

Adversarial AttackConvolutional Neural NetworkTransformerImage

🎯 What it does: This paper proposes an optimal transport-based adversarial patch attack method that generates patches capable of inducing misclassification across various networks in a black-box transfer attack scenario.

Optimistic Bayesian Optimization with Unknown Constraints

Quoc Phong Nguyen (Massachusetts Institute of Technology), Patrick Jaillet (Massachusetts Institute of Technology)

OptimizationTabular

🎯 What it does: A Bayesian optimization algorithm under known constraints (UCB-C, UCB-D) is proposed, providing theoretical guarantees in a decoupled query setting and achieving adaptive input and function query selection.

Oracle Efficient Algorithms for Groupwise Regret

Krishna Acharya (Georgia Institute of Technology), Juba Ziani (Georgia Institute of Technology)

OptimizationReinforcement LearningTabularBiomedical Data

🎯 What it does: In the online prediction problem, a new oracle-efficient algorithm is proposed for scenarios with intersecting subsequences (i.e., individuals belonging to multiple groups) to simultaneously ensure sublinear regret (groupwise regret) relative to the best model for all groups.

Orbit-Equivariant Graph Neural Networks

Matthew Morris (Oxford), Ian Horrocks (Oxford)

Drug DiscoveryGraph Neural NetworkGraphBiomedical Data

🎯 What it does: This paper proposes a more relaxed orbit-equivariance property than traditional equivariant functions to address the issue of generating different outputs for similar nodes. Based on this, four graph neural network architectures that can achieve orbit-equivariance are designed, and their effectiveness is validated on drug molecule design and synthesis datasets.

Order-Preserving GFlowNets

Yihang Chen (École Polytechnique Fédérale de Lausanne), Lukas Mauch (Sony Europe)

OptimizationReinforcement LearningTabularSequential

🎯 What it does: This paper proposes Order-Preserving GFlowNets (OP-GFNs), a method that samples candidate sets by learning a reward function consistent with a given (partial) ranking without the need for explicit scalar rewards, applicable to both single-objective and multi-objective optimization.

Out-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothness

Fran Jelenić (University of Zagreb), Jan Snajder

Anomaly DetectionTransformerText

🎯 What it does: A method for OOD detection using the smoothness of inter-layer representations in Transformer, named BLOOD, is proposed.

Out-of-Distribution Detection with Negative Prompts

Jun Nie (University of Science and Technology of China), Xinmei Tian (University of Science and Technology of China)

Anomaly DetectionTransformerPrompt EngineeringContrastive LearningImage

🎯 What it does: Using a pre-trained CLIP model, we learn positive and negative prompt words to construct a positive-negative classifier for more accurately detecting out-of-distribution (OOD) samples.

Out-Of-Domain Unlabeled Data Improves Generalization

seyed amir hossein saberi, Babak Khalaj

ClassificationDomain AdaptationOptimizationImage

🎯 What it does: A framework that combines Distributed Robust Optimization (DRO) with self-supervised learning (RSS training) is proposed to enhance the robustness and generalization performance of classifiers in the presence of a small number of labeled samples and a large number of unlabeled samples with distribution shifts.

Out-of-Variable Generalisation for Discriminative Models

Siyuan Guo (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)

Domain AdaptationTabularBiomedical Data

🎯 What it does: The study investigates the generalization of discriminative models when the variable sets of the source and target environments do not completely overlap, and proposes using the higher-order moments of residuals to identify the target prediction function.

Outliers with Opposing Signals Have an Outsized Effect on Neural Network Optimization

Elan Rosenfeld (Carnegie Mellon University), Andrej Risteski (Carnegie Mellon University)

OptimizationConvolutional Neural NetworkImage

🎯 What it does: By analyzing the changes in loss during the training process, the significant impact of anomalous samples with 'opposite signals' on the optimization of deep networks has been identified and quantified.

Overcoming the Pitfalls of Vision-Language Model Finetuning for OOD Generalization

Yuhang Zang (Nanyang Technological University), Chen Huang (Apple Inc.)

ClassificationDomain AdaptationKnowledge DistillationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality

🎯 What it does: This study investigates the overfitting defects of vision-language models in out-of-distribution (OOD) generalization and proposes the OGEN method, which synthesizes unknown class features through a class-conditional feature generator and implements adaptive self-distillation for regularization, enhancing the OOD generalization performance of the fine-tuned model.

Overthinking the Truth: Understanding how Language Models Process False Demonstrations

Danny Halawi (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)

ClassificationTransformerLarge Language ModelText

🎯 What it does: This study investigates the phenomenon of 'error imitation' in large language models during few-shot learning, revealing how models produce incorrect outputs under erroneous prompts.

OVOR: OnePrompt with Virtual Outlier Regularization for Rehearsal-Free Class-Incremental Learning

Wei-Cheng Huang (JPMorgan Chase), Hsiang Hsu (JPMorgan Chase)

ClassificationTransformerPrompt EngineeringImage

🎯 What it does: This paper proposes a prompt-based replay-free class incremental learning method, combining Virtual Outlier Regularization (VOR) to tighten the classifier's decision boundary, and introducing a single prompt scheme (OnePrompt) to simplify the prompt pool, ultimately forming the OVOR method.

OWL: A Large Language Model for IT Operations

Hongcheng Guo (Beihang University), Zhoujun Li (Cloudwise Research)

Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark

🎯 What it does: A large language model tailored for IT operations tasks, named OWL, has been developed, along with the OWL-Instruct instruction set and the OWL-Bench bilingual evaluation benchmark; the HMCE long-context extension method and the Mixture-of-Adapter parameter-efficient fine-tuning strategy have been proposed to enhance the model's performance in operational scenarios.

P$^2$OT: Progressive Partial Optimal Transport for Deep Imbalanced Clustering

Chuyu Zhang (ShanghaiTech University), Xuming He (ShanghaiTech University)

ClassificationRepresentation LearningTransformerImage

🎯 What it does: A deep imbalance clustering method based on progressive pseudo-labels is proposed, utilizing progressive partial optimal transport to generate high-quality pseudo-labels that are sensitive to class imbalance, thereby learning more robust representations and clustering.

P2Seg: Pointly-supervised Segmentation via Mutual Distillation

Zipeng Wang (University of Chinese Academy of Sciences), Zhenjun Han (University of Chinese Academy of Sciences)

SegmentationKnowledge DistillationConvolutional Neural NetworkImage

🎯 What it does: A point-level supervised instance segmentation framework P2Seg is proposed, which utilizes a mutual distillation module (MDM) for bidirectional knowledge transfer between semantic and instance information, generating high-quality pseudo labels and training the instance segmentation network.

PAC Prediction Sets Under Label Shift

Wenwen Si (University of Pennsylvania), Osbert Bastani (University of Pennsylvania)

ClassificationAnomaly DetectionImageTextTabularBiomedical DataComputed Tomography

🎯 What it does: A PAC-guaranteed prediction set construction algorithm is designed for the label shift scenario, utilizing confidence intervals and Gaussian elimination to propagate uncertainty.

PAC-FNO: Parallel-Structured All-Component Fourier Neural Operators for Recognizing Low-Quality Images

Jinsung Jeon (Yonsei University), Noseong Park (KAIST)

RecognitionImage TranslationImage

🎯 What it does: A parallel structure, all-frequency Fourier Neural Operator (PAC-FNO) is proposed to simultaneously address image recognition tasks with low resolution and various natural input variations while maintaining a single model.

PAE: Reinforcement Learning from External Knowledge for Efficient Exploration

Zhe Wu (QiYuan Lab), Yuanchun Shi (Tsinghua University)

TransformerReinforcement LearningAgentic AIText

🎯 What it does: The Planner-Actor-Evaluator (PAE) framework is proposed, allowing agents to learn guided by external natural language knowledge in sparse reward environments, thereby enhancing exploration efficiency.

PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization

Yidong Wang (Peking University), Yue Zhang (Westlake University)

OptimizationHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark

🎯 What it does: An automatic evaluation model named PandaLM is proposed for multi-dimensional evaluation of instruction-tuned large language models, and this evaluator is used for hyperparameter optimization.

PanoDiffusion: 360-degree Panorama Outpainting via Diffusion

Tianhao Wu (Nanyang Technological University), Tat-Jen Cham (Nanyang Technological University)

RestorationGenerationData SynthesisDepth EstimationSuper ResolutionDiffusion modelGenerative Adversarial NetworkImageMultimodality

🎯 What it does: This paper studies an outpainting method for 360° indoor RGB-D panoramas using a latent diffusion model (LDM), capable of generating missing areas without depth input while ensuring surrounding consistency.

Parallelizing non-linear sequential models over the sequence length

Yi Heng Lim (Machine Discovery), Muhammad Firmansyah Kasim (Machine Discovery)

OptimizationComputational EfficiencyRecurrent Neural NetworkTime SeriesSequentialPhysics RelatedOrdinary Differential Equation

🎯 What it does: A new algorithm is proposed that enables parallel evaluation and training of nonlinear sequence models (such as RNNs and neural ordinary differential equations) without changing the architecture of the sequence model, significantly accelerating the training speed.

Parameter-Efficient Multi-Task Model Fusion with Partial Linearization

Anke Tang (Wuhan University), Dacheng Tao (Nanyang Technological University)

ClassificationRecognitionTransformerSupervised Fine-TuningImageText

🎯 What it does: By locally linearizing the LoRA adapter (i.e., performing a first-order Taylor expansion only on the trainable modules), L-LoRA is proposed, and based on this, a unified multi-task model is constructed by integrating multiple task models.

Parameter-Efficient Orthogonal Finetuning via Butterfly Factorization

Weiyang Liu (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)

OptimizationComputational EfficiencyTransformerSupervised Fine-TuningDiffusion modelText

🎯 What it does: Proposes Orthogonal Butterfly (BOFT) - a parameter-efficient fine-tuning method that utilizes a butterfly structure for the factorization of orthogonal matrices, significantly reducing the number of trainable parameters while preserving pre-trained knowledge.

Parametric Augmentation for Time Series Contrastive Learning

Xu Zheng (Florida International University), Dongsheng Luo (Florida International University)

ClassificationOptimizationRepresentation LearningRecurrent Neural NetworkGenerative Adversarial NetworkContrastive LearningTime Series

🎯 What it does: This paper proposes an adaptive parameterized data augmentation framework for time series contrastive learning, called AutoTCL, which achieves information retention and view diversity through factorization and learns the optimal augmentation strategy during the pre-training phase.

Pareto Deep Long-Tailed Recognition: A Conflict-Averse Solution

Zhipeng Zhou (University of Science and Technology of China), Wei Gong (University of Science and Technology of China)

RecognitionOptimizationRepresentation LearningContrastive LearningImage

🎯 What it does: A dynamic rebalancing method based on Pareto multi-objective optimization (PLOT) is proposed to address gradient conflicts and representation learning degradation in long-tail recognition.

PARL: A Unified Framework for Policy Alignment in Reinforcement Learning from Human Feedback

Souradip Chakraborty (University of Maryland), Furong Huang (Oregon State University)

Reinforcement Learning from Human FeedbackReinforcement LearningSequential

🎯 What it does: A unified two-layer optimization framework called PARL is proposed to address the policy alignment problem in reinforcement learning, particularly in scenarios utilizing human feedback.

Parsing neural dynamics with infinite recurrent switching linear dynamical systems

Victor Geadah (Princeton University), Jonathan W. Pillow (Princeton University)

Recurrent Neural NetworkTime SeriesSequential

🎯 What it does: A model of infinite recursive switching linear dynamical systems (irSLDS) is constructed to analyze neural dynamics with non-stationarity and variable state numbers.

Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models

Gabriele Corso (Massachusetts Institute of Technology), Tommi S. Jaakkola

GenerationData SynthesisDrug DiscoveryDiffusion modelImageTextMultimodalityStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This paper proposes a Particle Guidance framework that utilizes time-evolving joint potential energy to achieve non-independent and more diverse sampling in diffusion models.

Partitioning Message Passing for Graph Fraud Detection

Wei Zhuo (Shenzhen Campus of Sun Yat-sen University), Jia Chen (GrabTaxi Holdings Pte. Ltd.)

Anomaly DetectionGraph Neural NetworkGraphFinance Related

🎯 What it does: This paper proposes the Partitioning Message Passing (PMP) framework, which utilizes neighbor label information to distinguish between similar and dissimilar neighbors. During the information aggregation phase, it employs different aggregation weights that are adaptively generated by the central node to address the issues of label imbalance and similarity/dissimilarity in graph fraud detection.

Patched Denoising Diffusion Models For High-Resolution Image Synthesis

Zheng Ding (University of California San Diego), Zhuowen Tu (Stanford University)

GenerationData SynthesisDiffusion modelImage

🎯 What it does: A block-based denoising diffusion model, Patch-DM, has been developed, capable of directly generating high-resolution images (such as 1024×1024, 2048×1024) and avoiding block boundary artifacts through a feature collage strategy.

Path Choice Matters for Clear Attributions in Path Methods

Borui Zhang (Tsinghua University), Jiwen Lu (Tsinghua University)

Explainability and InterpretabilityConvolutional Neural NetworkImage

🎯 What it does: A path selection-based explanation method called SAMP is proposed to address the issue of unclear explanations caused by path uncertainty in traditional path methods.

Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting

Peng Chen (East China Normal University), Chenjuan Guo (East China Normal University)

TransformerTime Series

🎯 What it does: Proposes Pathformer, a multi-scale Transformer with adaptive paths for time series forecasting.

PB-LLM: Partially Binarized Large Language Models

Zhihang Yuan (Houmo AI), Zhen Dong (University of California Berkeley)

TransformerLarge Language ModelText

🎯 What it does: This paper proposes a Partial Binarized Large Language Model (PB-LLM) framework that can quantize the weights of LLMs to extremely low bits (even 1 bit) while maintaining inference capability.

PBADet: A One-Stage Anchor-Free Approach for Part-Body Association

Zhongpai Gao (United Imaging Intelligence), Ziyan Wu (Shanghai Jiao Tong University)

Object DetectionPose EstimationImage

🎯 What it does: PBADet is proposed, a one-stage anchor-free body part-to-body association detection framework.

Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models

Hritik Bansal (University of California Los Angeles), Aditya Grover (University of California Los Angeles)

Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText

🎯 What it does: This paper studies the impact of sparse feedback protocols (scoring and ranking) on the alignment and evaluation of large language models (LLMs), and systematically assesses the differences between the two protocols in terms of preference consistency and evaluation consistency.

PeFLL: Personalized Federated Learning by Learning to Learn

Jonathan Scott (Institute of Science and Technology Austria), Christoph H Lampert

Federated LearningMeta LearningImage

🎯 What it does: We propose PeFLL, a personalized federated learning algorithm based on meta-learning, which can generate personalized models for any client through a single forward inference on the server side;

PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts

Bang An (University of Maryland), Furong Huang (Bosch Center for Artificial Intelligence)

ClassificationTransformerPrompt EngineeringContrastive LearningImage

🎯 What it does: The PerceptionCLIP method is proposed, which first uses CLIP to infer the contextual attributes of images (such as background, orientation, etc.), and then uses these attributes as conditions for zero-shot classification of images.

Perceptual Group Tokenizer: Building Perception with Iterative Grouping

Zhiwei Deng (Google Research), Yang Li (Google Research)

SegmentationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerContrastive LearningImage

🎯 What it does: A Perceptual Group Tokenizer (PGT) based entirely on perceptual grouping is proposed and trained within an unsupervised learning framework.

Perceptual Scales Predicted by Fisher Information Metrics

Jonathan Vacher (University Paris Cité), Pascal Mamassian (École Normale Supérieure)

Image

🎯 What it does: This paper measures and predicts perceptual scales such as spatial frequency, direction, and bandwidth through differential scaling experiments, and explores the relationship between perceptual scales and image power spectra by combining it with Fisher information theory.

Performance Gaps in Multi-view Clustering under the Nested Matrix-Tensor Model

Hugo Lebeau (Université Grenoble Alpes), José Henrique De Morais Goulart

Multimodality

🎯 What it does: Under the nested matrix-tensor model, the performance of two spectral estimation methods, tensor and unfolding, is analyzed, and the error gap between the two in multi-view clustering is quantitatively evaluated;

Periodicity Decoupling Framework for Long-term Series Forecasting

Tao Dai (Shenzhen University), Shu-Tao Xia (Tsinghua University)

TransformerTime Series

🎯 What it does: A periodic decoupling framework (PDF) is proposed, which splits the original 1D time series into short-term and long-term 2D variation sequences, and then performs parallel modeling to achieve long-period forecasting.

Personalize Segment Anything Model with One Shot

Renrui Zhang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)

SegmentationImageVideo

🎯 What it does: By providing only a reference image and its mask, the Segment Anything Model (SAM) is personalized to automatically segment user-specified visual concepts.

Pessimistic Nonlinear Least-Squares Value Iteration for Offline Reinforcement Learning

Qiwei Di (University of California), Quanquan Gu (University of California)

Reinforcement Learning

🎯 What it does: The paper proposes an offline reinforcement learning algorithm named PNLSVI, which utilizes nonlinear least squares value iteration to address the problem of nonlinear function approximation.

PF-LRM: Pose-Free Large Reconstruction Model for Joint Pose and Shape Prediction

Peng Wang (Adobe Research), Kai Zhang (Adobe Research)

Object DetectionPose EstimationTransformerNeural Radiance FieldImage

🎯 What it does: A system called PF-LRM is proposed, which can simultaneously predict camera pose and 3D object shape (NeRF) from a very small number of uncalibrated images.

Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement

Linlu Qiu (Massachusetts Institute of Technology), Xiang Ren (University of Southern California)

TransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: Through the iterative hypothesis refinement (generation, selection, refinement rules) method, the system evaluates the capabilities of language models in inductive reasoning tasks;

PhyloGFN: Phylogenetic inference with generative flow networks

Ming Yang Zhou, Yoshua Bengio (University of Montreal)

TransformerFlow-based ModelSequential

🎯 What it does: We propose PhyloGFN, which utilizes Generative Flow Networks (GFlowNet) to simultaneously infer tree topology and branch lengths in the full tree space, achieving both Bayesian and minimum parsimony phylogenetic inference.

Physics-Regulated Deep Reinforcement Learning: Invariant Embeddings

Hongpeng Cao (Technical University of Munich), Marco Caccamo (Washington State University)

Safty and PrivacyRobotic IntelligenceReinforcement LearningSequentialPhysics Related

🎯 What it does: A physics-constrained deep reinforcement learning framework, Phy-DRL, is proposed for safety-critical autonomous systems.

Piecewise Linear Parametrization of Policies: Towards Interpretable Deep Reinforcement Learning

Maxime Wabartha (McGill University), Joelle Pineau (Meta)

Explainability and InterpretabilityRobotic IntelligenceReinforcement LearningSequential

🎯 What it does: Proposes HyperCombinator (HC) - a piecewise linear reinforcement learning strategy with controllable sub-strategies, combining interpretability and performance;

PILOT: An $\mathcal{O}(1/K)$-Convergent Approach for Policy Evaluation with Nonlinear Function Approximation

Zhuqing Liu (Ohio State University), Songtao Lu (IBM Research)

OptimizationReinforcement LearningSequential

🎯 What it does: A new path-integrated primal-dual stochastic gradient algorithm (PILOT) is proposed for policy evaluation problems with nonlinear function approximation, aimed at improving convergence speed and sample complexity.

PINNACLE: PINN Adaptive ColLocation and Experimental points selection

Gregory Kang Ruey Lau (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)

Time SeriesPhysics Related

🎯 What it does: The PINNACLE algorithm is proposed, which jointly optimizes the selection of all types of training points (experimental points, PDE collocation points, IC/BC collocation points) in the Physics-Informed Neural Network (PINN) and automatically adjusts the ratio of collocation points based on training progress.

PINNsFormer: A Transformer-Based Framework For Physics-Informed Neural Networks

Zhiyuan Zhao (Georgia Institute of Technology), B. Aditya Prakash (Georgia Institute of Technology)

TransformerTime SeriesSequentialPhysics Related

🎯 What it does: This paper presents PINNsFormer, a Transformer-based framework designed to capture temporal dependencies and approximate PDE analytical solutions within Physics-Informed Neural Networks.

PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis

Junsong Chen (Dalian University of Technology), Zhenguo Li (Huawei Noah's Ark Lab)

GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText

🎯 What it does: PIXARTα is proposed, a text-to-image diffusion model based on Transformer, capable of achieving high-quality image synthesis comparable to existing SOTA at low cost and low energy consumption, supporting high-resolution generation (1024×1024).

Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks

Murtaza Dalal (Carnegie Mellon University), Ruslan Salakhutdinov (Mistral AI)

Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodality

🎯 What it does: Using a modular framework PSL that integrates large language model planning, motion planning, and reinforcement learning to learn online from visual inputs and solve long-horizon robotic tasks.

PlaSma: Procedural Knowledge Models for Language-based Planning and Re-Planning

Faeze Brahman (Allen Institute for Artificial Intelligence), Yejin Choi (Allen Institute for Artificial Intelligence)

OptimizationExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelText

🎯 What it does: This paper proposes the PLASMA framework, which utilizes small language models to achieve interpretable and high-quality programmatic planning and replanning through symbolic program knowledge distillation and verification-guided decoding during inference.

Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents

Yang Deng (National University of Singapore), Tat-Seng Chua (National University of Singapore)

Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText

🎯 What it does: A pluggable dialogue strategy planning plugin (PPDPP) is proposed, enabling large language models to achieve better goal attainment in active dialogue tasks through learning strategies.

Plug-and-Play Posterior Sampling under Mismatched Measurement and Prior Models

Marien Renaud (Washington University in St. Louis), Ulugbek Kamilov (Washington University in St. Louis)

RestorationImageStochastic Differential Equation

🎯 What it does: This paper studies the posterior sampling error of the Plug-and-Play Unadjusted Langevin Algorithm (PnP-ULA) when there is a mismatch between the measurement model and the prior (denoiser). It introduces the Posterior L₂ pseudo-metric and provides an upper bound for the error, followed by validation of the theory through experiments on Gaussian mixture models, grayscale deblurring, and color deblurring.

Plug-and-Play: An Efficient Post-training Pruning Method for Large Language Models

Yingtao Zhang (Tsinghua University), Carlo Vittorio Cannistraci (Tsinghua University)

CompressionComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: This paper proposes a post-training pruning framework that can directly perform one-time pruning on large language models, balancing model compression and hardware acceleration.

Plugin estimators for selective classification with out-of-distribution detection

Harikrishna Narasimhan (Google Research), Sanjiv Kumar (Google Research)

ClassificationAnomaly DetectionImage

🎯 What it does: A statistical formula for the problem of selective classification and out-of-distribution detection (SCOD) is proposed, along with the Bayes optimal rejection rule; based on this optimal rule, a plug-in estimator is designed that can achieve efficient rejection decisions in two settings: using only ID samples or using both ID and 'wild' mixed samples.

PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code

Xuan Ju (International Digital Economy Academy), Qiang Xu (Chinese University of Hong Kong)

GenerationOptimizationPrompt EngineeringDiffusion modelImageBenchmark

🎯 What it does: PnP Inversion is proposed, which decouples the source and target branches of the diffusion model, allowing for improved image editing quality with just three lines of code.

Point2SSM: Learning Morphological Variations of Anatomies from Point Clouds

Jadie Adams (University of Utah), Shireen Elhabian

GenerationData SynthesisOptimizationGraph Neural NetworkPoint Cloud

🎯 What it does: A novel unsupervised deep learning framework called Point2SSM is proposed, which can directly generate corresponding statistical shape models from raw point clouds.

Poisoned Forgery Face: Towards Backdoor Attacks on Face Forgery Detection

Jiawei Liang (Sun Yat-Sen University), Xiaochun Cao (Sun Yat-Sen University)

RecognitionAdversarial AttackConvolutional Neural NetworkImageVideo

🎯 What it does: This paper proposes a clean-label backdoor attack framework for face forgery detection models called Poisoned Forgery Face, which can implant a backdoor during the training phase and induce the model to incorrectly identify forged faces as real faces through a trigger.

Policy Rehearsing: Training Generalizable Policies for Reinforcement Learning

Chengxing Jia (Nanjing University), Yang Yu (Nanjing University)

Reinforcement LearningSequential

🎯 What it does: Proposed Policy Rehearsing through Dynamics Model Generation (ReDM) method, which trains adaptive policies in unknown environments using non-interactive or minimal interactive data.

Poly-View Contrastive Learning

Amitis Shidani (University of Oxford), Dan Busbridge (Apple)

Representation LearningConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: This paper proposes a Poly-View Contrastive Learning framework, investigating the representation learning objectives when using more than two views (multi-view tasks) on a single sample, and derives a new loss function through information maximization and sufficient statistics theory.

PolyGCL: GRAPH CONTRASTIVE LEARNING via Learnable Spectral Polynomial Filters

Jingyu Chen (Renmin University of China), Zhewei Wei (Renmin University of China)

ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph

🎯 What it does: The POLYGCL framework is proposed, which utilizes learnable spectral polynomial filters to generate low-pass and high-pass views, achieving unsupervised graph representation learning through contrastive learning joint training.

Polynomial Width is Sufficient for Set Representation with High-dimensional Features

Peihao Wang (University of Texas at Austin), Pan Li (Georgia Tech)

🎯 What it does: The study investigates the minimum bottleneck dimension L required in the DeepSets architecture to represent any continuous set function when the feature dimension D of the set elements is greater than 1; it is proven that for two common embedding layers (linear + power mapping LP and linear + exponential activation LE), a polynomial level of L is sufficient.

Polynormer: Polynomial-Expressive Graph Transformer in Linear Time

Chenhui Deng (Cornell University), Zhiru Zhang (Cornell University)

ClassificationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerGraph

🎯 What it does: A linear-time graph Transformer model named Polynormer is proposed, which can learn high-order polynomial features in node classification tasks while maintaining permutation invariance.

PolyVoice: Language Models for Speech to Speech Translation

Qian qian Dong, Yuxuan Wang (ByteDance)

TransformerLarge Language ModelPrompt EngineeringChain-of-ThoughtAudio

🎯 What it does: Proposes the PolyVoice framework, which uses three decoder-only language models to achieve speech-to-speech translation and supports non-written languages.

Pooling Image Datasets with Multiple Covariate Shift and Imbalance

Sotirios Panagiotis Chytas (University of Wisconsin Madison), Vikas Singh (University of Wisconsin Madison)

ClassificationSegmentationDomain AdaptationAuto EncoderImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease

🎯 What it does: This paper studies how to constrain invariance/covariance in the latent space through a category theory framework when pooling multi-source medical imaging data, in order to address covariate shift and imbalance issues.

PORF: POSE RESIDUAL FIELD FOR ACCURATE NEURAL SURFACE RECONSTRUCTION

Jia-Wang Bian (University of Oxford), Philip Torr (University of Oxford)

Pose EstimationDepth EstimationNeural Radiance FieldPoint Cloud

🎯 What it does: This paper proposes a new Pose Residual Field (PoRF) and an essential geometry-based loss to jointly optimize camera poses and neural surface reconstruction.

Pose Modulated Avatars from Video

Chunjin Song (University of British Columbia), Helge Rhodin (Bielefeld University)

GenerationPose EstimationGraph Neural NetworkNeural Radiance FieldVideo

🎯 What it does: This paper proposes a pose-driven frequency modulation network to reconstruct high-detail dynamic human heads from videos.

PoSE: Efficient Context Window Extension of LLMs via Positional Skip-wise Training

Dawei Zhu (Peking University), Sujian Li (Peking University)

RetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: By implementing a jump offset on position indices during training, the PoSE method achieves the goal of extending the LLM context window to hundreds of thousands or even infinite lengths while maintaining the original context window's training length.

Post-hoc bias scoring is optimal for fair classification

Wenlong Chen (Imperial College London), Yang Liu (ByteDance Research)

ClassificationOptimizationHyperparameter SearchConvolutional Neural NetworkSupervised Fine-TuningTabular

🎯 What it does: This paper proposes a post-processing method called MBS based on instance-level bias scores to achieve multiple group fairness constraints (DP, EOp, EO, and their combinations) without retraining the model, and provides analytical improvement rules for the Bayesian optimal classifier.

Posterior Sampling Based on Gradient Flows of the MMD with Negative Distance Kernel

Paul Hagemann (Technische Universität Berlin), Gabriele Steidl (Humboldt-Universität zu Berlin)

RestorationGenerationConvolutional Neural NetworkFlow-based ModelGenerative Adversarial NetworkImageComputed Tomography

🎯 What it does: This paper proposes a conditional flow method based on the Maximum Mean Discrepancy (MMD) negative distance kernel for sampling the posterior distribution of Bayesian inverse problems, and constructs a conditional generative model based on this.

Pre-Training and Fine-Tuning Generative Flow Networks

Ling Pan (Hong Kong University of Science and Technology), Yoshua Bengio (Mila - Quebec AI Institute)

GenerationReinforcement Learning from Human FeedbackFlow-based ModelContrastive LearningSequentialBiomedical Data

🎯 What it does: Proposes an unsupervised pre-training method for Generative Flow Networks (GFlowNet), using outcome-conditioned GFlowNet (OC-GFN) to learn to reach any target state without task rewards, and achieves rapid fine-tuning of downstream task reward functions through an amortized predictor;

Pre-Training Goal-based Models for Sample-Efficient Reinforcement Learning

Haoqi Yuan (Peking University), Zongqing Lu (Peking University)

Robotic IntelligenceTransformerReinforcement LearningVideo

🎯 What it does: Pretrain a goal-oriented model on large-scale unstructured data, significantly improving the sample efficiency of reinforcement learning through temporal abstraction provided by this model.

Pre-training LiDAR-based 3D Object Detectors through Colorization

Tai-Yu Pan (Ohio State University), Wei-Lun Chao (Ohio State University)

Object DetectionAutonomous DrivingContrastive LearningPoint Cloud

🎯 What it does: A Grounded Point Colorization (GPC) pre-training framework is proposed, which utilizes the corresponding colored images of point clouds for colorization learning of LiDAR 3D detectors, enhancing their perception of semantic structures.

Pre-training Sequence, Structure, and Surface Features for Comprehensive Protein Representation Learning

Youhan Lee (Kakao Brain), Jaehoon Kim (Kakao Brain)

Representation LearningProtein Structure PredictionTransformerContrastive LearningMultimodalityPoint CloudGraph

🎯 What it does: Combining protein sequences, 3D structures, and molecular surface information, a multimodal pre-training is conducted for proteins and applied to downstream functional prediction tasks.

Pre-training with Random Orthogonal Projection Image Modeling

Maryam Haghighat (Data61 CSIRO), Piotr Koniusz (Data61 CSIRO)

Computational EfficiencyRepresentation LearningTransformerContrastive LearningImage

🎯 What it does: A Random Orthogonal Projection Image Modeling (ROPIM) framework is proposed for self-supervised pre-training of Vision Transformers, replacing traditional binary masking methods.

Pre-training with Synthetic Data Helps Offline Reinforcement Learning

Zecheng Wang (New York University), Keith W. Ross (New York University)

TransformerReinforcement LearningTabular

🎯 What it does: This study investigates the pre-training of offline reinforcement learning models (Decision Transformer and CQL) using simple synthetic data, demonstrating that this method can enhance performance, even surpassing pre-training with large language corpora.

Predicting Emergent Abilities with Infinite Resolution Evaluation

Shengding Hu (Tsinghua University), Maosong Sun (Tsinghua University)

AI Code AssistantTransformerLarge Language ModelText

🎯 What it does: The PASSUNTIL evaluation strategy is proposed to continuously assess task performance during the decoding phase through a large number of random samples, resulting in a theoretically infinite resolution task performance estimate (PASSUNTIL score PU), which can quantify and predict the task performance scaling of large language models and quantitatively study the phenomenon of 'accelerated emergence.'

Prediction Error-based Classification for Class-Incremental Learning

Michał Zając (KU Leuven), Gido M van de Ven

ClassificationKnowledge DistillationImageStochastic Differential Equation

🎯 What it does: A new class-incremental learning method called Prediction Error Classification (PEC) is proposed, which trains a student network for each class to fit the output of a random teacher network, using the student's prediction error against the teacher as the class score.

Prediction without Preclusion: Recourse Verification with Reachable Sets

Avni Kothari (University of California San Francisco), Berk Ustun (University of California San Diego)

OptimizationExplainability and InterpretabilityTabularFinance Related

🎯 What it does: This paper proposes a 'recourse verification' method that audits machine learning models by constructing reachable sets to detect whether model predictions can be altered through actionable constraints that are achievable by the subjects, thereby identifying whether the model exhibits admission discrimination due to fixed predictions.

Predictive auxiliary objectives in deep RL mimic learning in the brain

Ching Fang, Kim Stachenfeld

Representation LearningReinforcement LearningContrastive LearningSequential

🎯 What it does: This paper studies the impact of incorporating predictive auxiliary objectives (contrastive predictive coding) on network representation learning and learning efficiency in deep reinforcement learning, and maps the various modules of the RL network to the hippocampus, basal ganglia, and visual cortex, attempting to explain the neural representation changes observed in various animal experiments.

Predictive, scalable and interpretable knowledge tracing on structured domains

Hanqi Zhou (University of Tübingen), Álvaro Tejero-Cantero (University of Tübingen)

Recommendation SystemExplainability and InterpretabilityTabularStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: An interpretable, scalable, and predictable knowledge tracing model PSI-KT is proposed, which utilizes a hierarchical generative state space model to simultaneously estimate learners' cognitive features and knowledge prerequisite structures.

PRES: Toward Scalable Memory-Based Dynamic Graph Neural Networks

Junwei Su (University of Hong Kong), Chuan Wu (University of Hong Kong)

Graph Neural NetworkGraphTime Series

🎯 What it does: This study addresses and solves the performance degradation issue caused by temporal discontinuity in Memory-based Dynamic Graph Neural Networks (MDGNN) as the temporal batch size increases, proposing a training framework named PRES (Predict-to-Smooth);

PRIME: Prioritizing Interpretability in Failure Mode Extraction

Keivan Rezaei (University of Maryland), Soheil Feizi (University of Maryland)

ClassificationExplainability and InterpretabilityVision Language ModelImage

🎯 What it does: A failure mode detection method based on human-interpretable labels, called PRIME, is proposed, which can identify and describe hard-to-classify subgroups in a trained image classification model using readable text.

Principled Architecture-aware Scaling of Hyperparameters

Wuyang Chen (University of Texas), Boris Hanin (Princeton University)

Hyperparameter SearchNeural Architecture SearchConvolutional Neural NetworkImage

🎯 What it does: This study investigates the impact of neural network architecture (in DAG form) on initialization and learning rate, proposing topology-aware initialization and learning rate scaling formulas, and experimentally validating them on MLP, CNN, and NAS-Bench-201.

Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting

Enyi Jiang (University of Illinois Urbana-Champaign), Sanmi Koyejo (Stanford University)

Domain AdaptationFederated LearningImage

🎯 What it does: A novel aggregation rule for domain adaptation between source clients and target clients in federated learning is proposed, and performance is further enhanced through adaptive weighting.

Prioritized Soft Q-Decomposition for Lexicographic Reinforcement Learning

Finn Rietz (Orebro University), Johannes A. Stork (Orebro University)

Reinforcement Learning

🎯 What it does: This paper proposes an algorithm called PSQD that can solve Lexicographic Multi-Objective Reinforcement Learning (MORL) in continuous state-action spaces, enabling zero-shot reuse and online/offline adaptation to subtask solutions.

Privacy Amplification for Matrix Mechanisms

Christopher A. Choquette-Choo (Google DeepMind), Abhradeep Guha Thakurta

Safty and PrivacyMixture of ExpertsImage

🎯 What it does: The MMCC algorithm is proposed for calculating the privacy amplification guarantees of arbitrary matrix mechanisms (including DP-FTRL) and provides approximately optimal ε-δ results; it also proves the amplification effect of the binary tree mechanism under shuffling and experimentally validates its privacy-efficiency enhancement in machine learning tasks.

Privacy-Preserving In-Context Learning for Large Language Models

Tong Wu (Princeton University), Prateek Mittal (Princeton University)

GenerationSafty and PrivacyLarge Language ModelText

🎯 What it does: A differential privacy-based in-context learning (DP-ICL) framework is proposed and implemented, utilizing chunked example sets for parallel inference and noise aggregation to achieve privacy protection in text classification and language generation tasks.

Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation

Xinyu Tang (Princeton University), Robert Sim (Microsoft Research)

GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper proposes an algorithm for synthesizing a small number of examples (few-shot) from private datasets while maintaining differential privacy (DP) guarantees, and using them as demonstrations for in-context learning (ICL) under large language models (LLM);

Private Zeroth-Order Nonsmooth Nonconvex Optimization

Qinzi Zhang (Boston University), Ashok Cutkosky (Boston University)

OptimizationSafty and Privacy

🎯 What it does: A zero-order differentially private gradient optimization algorithm for non-convex and non-smooth objectives is proposed, which can find (δ, ϵ) Goldstein stationary points while maintaining RDP privacy.

Privately Aligning Language Models with Reinforcement Learning

Fan Wu (University of Illinois Urbana-Champaign), Robert Sim (Microsoft Research)

Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: This study investigates how to incorporate differential privacy into the reinforcement learning (PPO) framework for aligning large language models, proposing a three-step DP alignment process and implementing a DP version of PPO.

Privileged Sensing Scaffolds Reinforcement Learning

Edward S. Hu (University of Pennsylvania), Dinesh Jayaraman (University of Pennsylvania)

Robotic IntelligenceReinforcement LearningWorld Model

🎯 What it does: This paper proposes a 'Perceptual Scaffolding' framework that utilizes privileged sensor information obtainable during training to accelerate and enhance the performance of reinforcement learning under target sensors (only limited sensors) for robots. It also designs a suite of ten diverse perceptual scaffolding tasks (S3) to evaluate this method.

Probabilistic Adaptation of Black-Box Text-to-Video Models

Sherry Yang (Google DeepMind), Pieter Abbeel (University of California Berkeley)

GenerationDomain AdaptationDiffusion modelScore-based ModelVideoText

🎯 What it does: This paper proposes Video Adapter, which utilizes the scores from a black-box text-to-video diffusion model as a probabilistic prior, combined with a small domain-specific model for weightless adaptation, to generate high-quality domain-specific videos.

Probabilistic Self-supervised Representation Learning via Scoring Rules Minimization

Amirhossein Vahidi (Wellcome Sanger Institute), Mina Rezaei (Munich Center for Machine Learning)

Knowledge DistillationRepresentation LearningTransformerContrastive LearningImage

🎯 What it does: The ProSMin method is proposed, which minimizes the dimensional collapse of representations and enhances representation quality during the pre-training phase by using a loss function based on appropriate scoring rules through probability knowledge distillation between the online network and the target network.

Probabilistically Rewired Message-Passing Neural Networks

Chendi Qian (RWTH Aachen University), Christopher Morris (RWTH Aachen University)

Graph Neural NetworkGraph

🎯 What it does: A probabilistic reconnection graph neural network framework PR-MPNN is proposed, which utilizes differentiable k-subset sampling to learn to add and remove edges in the graph to enhance expressive power;

Procedural Fairness Through Decoupling Objectionable Data Generating Components

Zeyu Tang (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)

Tabular

🎯 What it does: A framework for program fairness based on causal graphs is proposed, which decouples objectionable causal paths from neutral paths during the data generation process using 'reference points' and 'value instantiation rules', predicting only with neutral paths to avoid hidden impacts of procedural unfairness.