ICLR 2023 Papers — Page 11
International Conference on Learning Representations · 1573 papers
Not All Tasks Are Born Equal: Understanding Zero-Shot Generalization
Jing Zhou (Tsinghua University), Zhilin Yang
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper reveals that the zero-shot generalization performance of T0 multi-task prompt-based pre-training is mainly determined by a small number of key tasks, especially QA-type tasks, through experiments. It also proposes a task resampling method based on the generalization results between task pairs to enhance model performance.
Novel View Synthesis with Diffusion Models
Daniel Watson (Google Research), Mohammad Norouzi (Google Research)
GenerationData SynthesisDiffusion modelNeural Radiance FieldImage
🎯 What it does: We propose 3DiM—a 3D novel view synthesis method based on diffusion models, capable of generating multiple geometrically consistent and detail-rich views from just one or a few input views.
NTFields: Neural Time Fields for Physics-Informed Robot Motion Planning
Ruiqi Ni (Purdue University), Ahmed H Qureshi
OptimizationRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningPoint CloudPhysics Related
🎯 What it does: This paper proposes a physics-driven Neural Time Field (NTFields) that completes high-dimensional robot motion planning by directly solving the Eikonal equation, without the need for expert trajectory data.
NTK-SAP: Improving neural network pruning by aligning training dynamics
Yite Wang (University of Illinois Urbana-Champaign), Ruoyu Sun (Chinese University of Hong Kong)
CompressionOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a pre-pruning method called NTK-SAP, which utilizes the spectral information of the Neural Tangent Kernel (NTK) to evaluate and prune connections that have the least impact on the NTK spectrum, aiming to maintain the training dynamics of the pruned network similar to that of the full network.
ODAM: Gradient-based Instance-Specific Visual Explanations for Object Detection
Chenyang ZHAO, Antoni B. Chan (City University of Hong Kong)
Object DetectionExplainability and InterpretabilityImage
🎯 What it does: This paper proposes ODAM, a gradient-based instance-level visual explanation method for interpreting the decisions of object detectors regarding each predicted attribute (class score, bounding box coordinates). It also designs the Odam-Train training strategy to enhance explanation consistency and discriminability, and further implements Odam-NMS based on ODAM's explanations to better eliminate duplicate detections in crowded scenes.
Offline Congestion Games: How Feedback Type Affects Data Coverage Requirement
Haozhe Jiang (Tsinghua University), Simon Shaolei Du
OptimizationReinforcement Learning
🎯 What it does: The study investigates how to effectively recover approximate Nash equilibria in offline congestion games under different types of feedback (facility-level, agent-level, game-level).
Offline Q-learning on Diverse Multi-Task Data Both Scales And Generalizes
Aviral Kumar (Google Research), Sergey Levine (Google Research)
Convolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: Trained a large-scale offline Q-learning model on multi-task Atari games, demonstrating scalability and transferability through improvements in network architecture, loss function, and feature normalization.
Offline Reinforcement Learning via High-Fidelity Generative Behavior Modeling
Huayu Chen (Tsinghua University), Jun Zhu (Tsinghua University)
Reinforcement LearningDiffusion modelTabular
🎯 What it does: An offline reinforcement learning method is proposed, which uses a diffusion generative model to learn high-fidelity behavior policies, and then generates executable policies through a combination of importance sampling and action evaluation models.
Offline Reinforcement Learning with Differentiable Function Approximation is Provably Efficient
Ming Yin (University of California), Yu-Xiang Wang (University of California)
Reinforcement Learning
🎯 What it does: This study investigates differentiable function approximation methods in offline reinforcement learning and provides theoretical sample complexity guarantees.
Offline RL for Natural Language Generation with Implicit Language Q Learning
Charlie Victor Snell (University of California Berkeley), Sergey Levine (University of California Berkeley)
GenerationTransformerReinforcement LearningText
🎯 What it does: This paper proposes an offline reinforcement learning method for language generation called ILQL, which utilizes Q-learning and implicit data support constraints to self-supervise fine-tune existing large-scale text data, optimizing any user-specified reward function.
Offline RL with No OOD Actions: In-Sample Learning via Implicit Value Regularization
Haoran Xu (Institute for AI Industry Research), Xianyuan Zhan (Institute for AI Industry Research)
Reinforcement LearningTabular
🎯 What it does: An Implicit Value Regularization (IVR) framework is proposed, and based on this framework, two sample-based offline reinforcement learning algorithms, SQL (Sparse Q-Learning) and EQL (Exponential Q-Learning), are designed.
Ollivier-Ricci Curvature for Hypergraphs: A Unified Framework
Corinna Coupette (Max Planck Institute for Informatics), Bastian Rieck (Helmholtz Munich)
Graph
🎯 What it does: A unified Ollivier-Ricci curvature framework (ORCHID) is proposed for the geometric analysis and learning tasks of hypergraphs.
Omnigrok: Grokking Beyond Algorithmic Data
Ziming Liu (Massachusetts Institute of Technology), Max Tegmark (Massachusetts Institute of Technology)
Representation LearningImageText
🎯 What it does: This paper explains and induces the grokking phenomenon on algorithmic and non-algorithmic datasets by analyzing the weight norm of neural networks and the loss curves.
On Accelerated Perceptrons and Beyond
Guanghui Wang (Georgia Tech), Jacob Abernethy (Google Research)
Optimization
🎯 What it does: This paper proposes a unified framework that views all known accelerated perceptrons as an optimistic online learning process in a two-player zero-sum game, and improves the convergence speed of several related problems based on this framework.
On Achieving Optimal Adversarial Test Error
Justin D. Li (University of Illinois), Matus Telgarsky (University of Illinois)
OptimizationAdversarial Attack
🎯 What it does: This paper theoretically proves that under idealized conditions, adversarial training of shallow ReLU networks (combined with early stopping and ideal adversaries) can achieve optimal adversarial test error;
On amortizing convex conjugates for optimal transport
Brandon Amos (Meta AI)
GenerationOptimizationSupervised Fine-TuningImageBenchmark
🎯 What it does: This paper proposes a method to amortize the convex conjugate operation and use fine-tuning to efficiently solve the dual potential of Wasserstein-2 optimal transport in continuous space.
On Compositional Uncertainty Quantification for Seq2seq Graph Parsing
Zi Lin (University of California San Diego), Jingbo Shang (University of California San Diego)
TransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: This paper proposes the Graph Autoregressive Process (GAP) framework and the Compositional Expected Calibration Error (CECE) metric, which can map the sequence probabilities of seq2seq models to the conditional probabilities of graph nodes/edges, thereby quantifying and evaluating the compositional uncertainty in graph parsing tasks.
On Explaining Neural Network Robustness with Activation Path
Ziping Jiang (Lancaster University)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: By decomposing the activation patterns of neural networks, the network is divided into fixed paths and floating paths. The impact of floating paths on robustness is analyzed, and a mechanism to suppress floating paths (SC-RFP) is introduced in randomized smoothing, thereby enhancing the verifiable robustness of the model.
On Pre-training Language Model for Antibody
Danqing Wang (ByteDance Research), Hao Zhou (Institute for AI Industry Research)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningBiomedical DataBenchmark
🎯 What it does: This paper studies the performance of protein and antibody-specific pre-trained language models on different specificity antibody tasks, proposing the first benchmark ATUE covering four types of antibody tasks, and designing and training the EATLM model based on antibody evolutionary mechanisms.
On Representing Linear Programs by Graph Neural Networks
Ziang Chen (Duke University), Wotao Yin (Alibaba US)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proves that Graph Neural Networks (GNNs) can express and approximate the feasibility, optimal objective value, and optimal solution of Linear Programming (LP), establishing a theoretical foundation for using GNNs in learning optimization.
On Representing Mixed-Integer Linear Programs by Graph Neural Networks
Ziang Chen (Duke University), Wotao Yin (Alibaba US)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the expressive power of Graph Neural Networks (GNN) in representing Mixed Integer Linear Programming (MILP) problems. It proves that, in general, GNNs cannot distinguish between certain feasible and infeasible MILP instances (foldable MILP) and proposes to restore the expressiveness of GNNs by restricting instances to unfoldable MILP or by adding random features.
On the complexity of nonsmooth automatic differentiation
Jerome Bolte (Toulouse School of Economics), Béatrice Pesquet-Popescu (Thales LAS France)
OptimizationComputational Efficiency
🎯 What it does: This paper studies the computational complexity of nonsmooth automatic differentiation, proposes a dimensionless overhead theorem based on conservative gradients, and analyzes the complexity differences between forward mode and backward mode.
On the Convergence of AdaGrad(Norm) on $\mathbb{R}^d$: Beyond Convexity, Non-Asymptotic Rate and Acceleration
Zijian Liu (New York University), Huy Nguyen
Optimization
🎯 What it does: This paper conducts an in-depth analysis of the convergence of AdaGrad and its variants under unconstrained smooth convex functions and more general γ-quasar convex functions. It provides explicit non-asymptotic convergence rates for both deterministic and stochastic cases, and proposes two new variants that ensure the convergence of the last iteration; additionally, it designs two accelerated adaptive algorithms that achieve a convergence rate of O(1/T²).
On the Data-Efficiency with Contrastive Image Transformation in Reinforcement Learning
Sicong Liu (Institute of Automation, Chinese Academy of Sciences), Jian Cheng (Nanjing University of Science and Technology)
Reinforcement LearningContrastive LearningImage
🎯 What it does: A learnable Contrast-Invariant Image Transformation (CoIT) has been designed and implemented to enhance the data efficiency of visual reinforcement learning.
On the duality between contrastive and non-contrastive self-supervised learning
Quentin Garrido (Meta AI), Yann LeCun (New York University)
OptimizationRepresentation LearningContrastive LearningImage
🎯 What it does: A unified theoretical analysis of sample contrastive and dimension contrastive methods in contrastive learning is conducted, and their similarities and interchangeability in practical training are verified through experiments.
On the Effectiveness of Out-of-Distribution Data in Self-Supervised Long-Tail Learning.
Jianhong Bai (Zhejiang University), Haoji Hu (Zhejiang University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A framework called COLT has been developed to utilize external unlabeled OOD data for self-supervised long-tail learning, addressing the issue of traditional SSL's excessive focus on the main class under long-tail distributions.
On the Feasibility of Cross-Task Transfer with Model-Based Reinforcement Learning
Yifan Xu (University of California San Diego), Zhuowen Tu (University of California San Diego)
Knowledge DistillationReinforcement LearningWorld ModelVideoBenchmark
🎯 What it does: A model-based cross-task transfer framework called XTRA is proposed, which first pre-trains a world model on multi-task offline data and then fine-tunes it in the online interaction of the target task by parallelly using offline task data to improve sample efficiency.
On the Importance and Applicability of Pre-Training for Federated Learning
Hong-You Chen (Ohio State University), Wei-Lun Chao (Ohio State University)
Federated LearningContrastive LearningImageBiomedical Data
🎯 What it does: This paper systematically studies the feasibility and effectiveness of introducing pre-training in federated learning, evaluating real data pre-training, synthetic fractal image pre-training, and self-supervised two-stage pre-training, and conducts comparative experiments on their performance under different federated settings.
On The Inadequacy of Optimizing Alignment and Uniformity in Contrastive Learning of Sentence Representations
Zhijie Nie (Beihang University), Yongyi Mao (University of Ottawa)
OptimizationRepresentation LearningTransformerContrastive LearningText
🎯 What it does: This study investigates the shortcomings of the alignment and decoupling forms of contrastive learning loss in sentence representation learning, and proposes a gradient dissipation mechanism to explain the performance gap.
On the Performance of Temporal Difference Learning With Neural Networks
HAOXING TIAN, Alex Olshevsky (Boston University)
Reinforcement LearningSequential
🎯 What it does: This paper studies TD learning of neural networks under fixed radius projection and provides an upper bound on convergence error, proving that the final error in a network with width m is O(ε + 1/√m);
On the Perils of Cascading Robust Classifiers
Ravi Mangal (Carnegie Mellon University), Matt Fredrikson (Carnegie Mellon University)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Discussed the limitations of cascade robust classifiers, proving their certification is unreliable and proposing attack methods.
On The Relative Error of Random Fourier Features for Preserving Kernel Distance
Kuan Cheng (Peking University), Zhide Wei (Peking University)
Tabular
🎯 What it does: This study investigates the performance of Random Fourier Features (RFF) in maintaining the relative error of kernel distances and proposes a data-independent dimensionality reduction method for the Laplacian kernel.
On the Robustness of Safe Reinforcement Learning under Observational Perturbations
Zuxin Liu (Carnegie Mellon University), Ding Zhao (Carnegie Mellon University)
Safty and PrivacyRobotic IntelligenceReinforcement LearningSequentialBenchmark
🎯 What it does: This paper studies the robustness issue of safe reinforcement learning in the face of observation disturbances, proposing an evaluation framework and demonstrating that optimal policies are prone to failure under observation attacks.
On the Saturation Effect of Kernel Ridge Regression
Yicheng Li (Tsinghua University), Qian Lin (Beijing Academy of Artificial Intelligence)
🎯 What it does: It proves the saturation effect of kernel ridge regression under high smoothness, that is, regardless of the tuning parameters, the error convergence rate is limited to n^{-2/(2+β)}.
On the Sensitivity of Reward Inference to Misspecified Human Models
Joey Hong (University of California Berkeley), Anca Dragan (University of California Berkeley)
Reinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: This study investigates the sensitivity of reward inference to human model errors in reward learning, providing counterexamples where small errors lead to large mistakes, as well as positive results with linear upper bounds under the assumption of log-concavity, validated through experiments.
On the Soft-Subnetwork for Few-Shot Class Incremental Learning
Haeyong Kang (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes SoftNet, a few-shot class incremental learning method implemented through soft sub-networks (major + minor sub-networks), aimed at alleviating catastrophic forgetting and overfitting simultaneously.
On The Specialization of Neural Modules
Devon Jarvis (University of the Witwatersrand), Andrew M Saxe (University College London)
Convolutional Neural NetworkImage
🎯 What it does: This paper studies the specialization capability of neural modules in datasets and verifies through theoretical analysis and experiments whether modular networks can achieve systematic generalization.
On the Trade-Off between Actionable Explanations and the Right to be Forgotten
Martin Pawelczyk (University of Tübingen), Gjergji Kasneci (University of Tübingen)
Explainability and InterpretabilityTabular
🎯 What it does: This paper explores the trade-off between actionable explanations in machine learning models and the 'right to be forgotten', and proposes an algorithm to find the most critical training samples to maximize explanation failure.
On the Usefulness of Embeddings, Clusters and Strings for Text Generation Evaluation
Tiago Pimentel (University of Cambridge), Ryan Cotterell (ETH Zürich)
GenerationTransformerLarge Language ModelText
🎯 What it does: This paper explores why the MAUVE metric is highly correlated with human judgments through theoretical and experimental analysis, ultimately finding that its core advantage lies in using clustering distributions based on pre-trained model embeddings to approximate string distributions, rather than the newly proposed AUC Divergence itself.
On the Word Boundaries of Emergent Languages Based on Harris's Articulation Scheme
Ryo Ueda (University of Tokyo), Yusuke Miyao (University of Tokyo)
Recurrent Neural NetworkSequential
🎯 What it does: Investigate whether Harris's pronunciation scheme (HAS) holds in simulated signal games, and evaluate whether the word boundaries of generated language have semantic significance through unsupervised entropy-based word boundary detection.
One Transformer Can Understand Both 2D & 3D Molecular Data
Shengjie Luo (Peking University), Di He (Peking University)
Drug DiscoveryTransformerMultimodalityGraph
🎯 What it does: Design a Transformer-M model that can simultaneously handle molecular 2D graphics and 3D geometric data, achieving cross-modal learning with a single model.
One-Pixel Shortcut: On the Learning Preference of Deep Neural Networks
Shutong Wu (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Designed and validated the One-Pixel Shortcuts (OPS) technique to generate unlearnable samples, misleading deep networks during training and resulting in poor performance on clean test sets;
Online Bias Correction for Task-Free Continual Learning
Aristotelis Chrysakis (KU Leuven), Marie-Francine Moens (KU Leuven)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: The research task addresses the prediction bias in task-agnostic continual learning, explains the 'freshness bias' caused by experience replay, and proposes an online bias correction method (OBC) that can continuously correct bias during the learning process.
Online Boundary-Free Continual Learning by Scheduled Data Prior
Hyunseo Koh (Gwangju Institute of Science and Technology), Jonghyun Choi (Yonsei University)
Knowledge DistillationImage
🎯 What it does: Proposes a boundary-free online continual learning framework to address the problem of data stream learning without task boundaries.
Online Low Rank Matrix Completion
Soumyabrata Pal (Google Research), Prateek Jain (Google Research)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the online low-rank matrix completion problem, aiming to recommend items for M users and N items over T rounds, and to obtain noise rewards based on a low-rank user-item preference matrix.
Open-Vocabulary Object Detection upon Frozen Vision and Language Models
Weicheng Kuo (Google Research), Anelia Angelova (Google Research)
Object DetectionConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A vocabulary-free object detection method based on frozen vision-language models (such as CLIP) called F-VLM is proposed, which only requires training the detection head to complete the detection task.
Optimal Activation Functions for the Random Features Regression Model
Jianxin Wang (Rice University), José Bento (Boston College)
🎯 What it does: This paper derives the optimal activation function in the Random Feature Regression (RFR) model, provides its closed-form analytical solution, and analyzes its impact on test error, sensitivity, and the double descent curve.
Optimal Conservative Offline RL with General Function Approximation via Augmented Lagrangian
Paria Rashidinejad (University of California), Jiantao Jiao (University of California)
OptimizationReinforcement LearningTabular
🎯 What it does: In offline reinforcement learning, an algorithm is proposed that combines the augmented Lagrangian method with marginal importance sampling (MIS), achieving statistically optimal and practically feasible conservative policy learning under general function approximation and single-policy concentration assumptions.
Optimal Transport for Offline Imitation Learning
Yicheng Luo (University College London), Marc Peter Deisenroth (University College London)
Robotic IntelligenceReinforcement LearningSequentialBenchmark
🎯 What it does: Align expert demonstrations with reward-free offline trajectories using optimal transport, automatically generating rewards for each step, allowing offline RL to directly learn good policies.
Optimistic Exploration with Learned Features Provably Solves Markov Decision Processes with Neural Dynamics
Sirui Zheng (Northwestern University), Zhaoran Wang (Northwestern University)
OptimizationReinforcement Learning
🎯 What it does: A novel algorithm ELNF suitable for MDPs with neural network features is proposed, which can explore in structured environments with theoretical guarantees and provides an upper bound on sample complexity.
Optimizing Bi-Encoder for Named Entity Recognition via Contrastive Learning
Sheng Zhang (Microsoft Research), Hoifung Poon (Microsoft Research)
RecognitionOptimizationTransformerContrastive LearningTextBiomedical Data
🎯 What it does: A bi-encoder based named entity recognition framework is proposed, which maps entity spans and entity types to the same vector space through contrastive learning, and introduces a dynamic threshold loss to distinguish between entities and non-entities.
Optimizing Spca-based Continual Learning: A Theoretical Approach
Chunchun Yang (University of Science and Technology of China), Zengfu Wang (University of Science and Technology of China)
ClassificationOptimizationSupervised Fine-TuningImage
🎯 What it does: This paper proposes a continuous learning algorithm OSCL based on supervised principal component analysis (SPCA) and provides an analytical classification error of the algorithm within a high-dimensional statistical theoretical framework. Furthermore, it theoretically avoids catastrophic forgetting through label optimization and achieves flexible adjustment of task weights.
OPTQ: Accurate Quantization for Generative Pre-trained Transformers
Elias Frantar (IST Austria), Dan Alistarh (IST Austria)
GenerationCompressionTransformerLarge Language ModelText
🎯 What it does: A one-time post-training quantization method called OPTQ is proposed, which can compress large Transformer models (up to 175B parameters) to 3/4 bits while maintaining almost no accuracy loss.
Order Matters: Agent-by-agent Policy Optimization
Xihuai Wang (Shanghai Jiao Tong University), Weinan Zhang (Shanghai Jiao Tong University)
OptimizationReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes an Agent-by-Agent policy optimization algorithm A2PO, which achieves trust region optimization for multiple agents under single round sampling while maintaining monotonic improvement guarantees for each agent and the overall policy.
Ordered GNN: Ordering Message Passing to Deal with Heterophily and Over-smoothing
Yunchong Song (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
Graph Neural NetworkGraph
🎯 What it does: Proposes Ordered GNN, which arranges the neurons of node representations according to the hierarchical structure of a root tree, forming an ordered message passing mechanism;
OTOv2: Automatic, Generic, User-Friendly
Tianyi Chen (Microsoft), Ilya Zharkov (Microsoft)
CompressionOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Proposes the Only-Train-Once v2 framework, which trains and compresses any deep neural network in a single pass without the need for pre-training or fine-tuning, automatically partitions Zero-Invariant Groups (ZIG) and constructs a compressed model.
Out-of-Distribution Detection and Selective Generation for Conditional Language Models
Jie Ren (Google Research), Peter J Liu (Google Research)
GenerationAnomaly DetectionTransformerLarge Language ModelText
🎯 What it does: A lightweight OOD detection method based on the internal embeddings of conditional language models is proposed, achieving selective generation in summarization and translation tasks, enhancing the model's robustness to inputs that do not conform to the training distribution.
Out-of-Distribution Detection based on In-Distribution Data Patterns Memorization with Modern Hopfield Energy
Jinsong Zhang (Nankai University), Dongmei Zhang
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A method for OOD detection based on the energy of modern Hopfield networks, called HE, is proposed, and a simplified version without hyperparameters, SHE, is derived based on it, using representative features from the training set ID as patterns for store-then-compare discrimination.
Out-of-distribution Detection with Implicit Outlier Transformation
Qizhou Wang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a model perturbation-based implicit data transformation and distribution-independent outlier exposure (DOE) method to enhance the anomaly detection capability of deep models on samples from unknown distributions.
Out-of-distribution Representation Learning for Time Series Classification
Wang Lu (Institute of Computing Technology, Chinese Academy of Sciences), Xing Xie (Fudan University)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerTime SeriesAudio
🎯 What it does: A domain-agnostic time series OOD representation learning method called DIVERSIFY is proposed, which utilizes adversarial self-supervised pseudo-domain label iterative mining to normalize the latent sub-distributions in the data.
Outcome-directed Reinforcement Learning by Uncertainty \& Temporal Distance-Aware Curriculum Goal Generation
Daesol Cho (Seoul National University), H. Jin Kim (Seoul National University)
Robotic IntelligenceMeta LearningReinforcement LearningSequential
🎯 What it does: A curriculum learning method based on uncertainty and temporal distance perception is proposed, allowing the agent to automatically progress towards the desired outcome without environmental rewards and only given target examples.
Over-parameterized Model Optimization with Polyak-{\L}ojasiewicz Condition
Yixuan Chen (Fudan University), Li Shang (Fudan University)
OptimizationTransformerMixture of ExpertsImageText
🎯 What it does: A condition number regularization method based on the Polyak-Łojasiewicz (PL) condition is proposed, which dynamically masks poorly performing sub-networks during training through a learnable structured pruning (gating network), thereby improving the convergence speed and generalization ability of over-parameterized models.
Over-Training with Mixup May Hurt Generalization
Zixuan Liu (University of Ottawa), Yongyi Mao (University of Ottawa)
Data-Centric LearningImage
🎯 What it does: This study investigates the U-shaped generalization curve that appears when Mixup is over-trained, providing theoretical explanations and experimental validation.
PAC Reinforcement Learning for Predictive State Representations
Wenhao Zhan, Jason D. Lee
Reinforcement Learning
🎯 What it does: A strategy learning algorithm CRANE based on maximum likelihood estimation is proposed, which can learn an approximately optimal policy in partially observable predictive state representation (PSR) environments with polynomial sample complexity.
PAC-NeRF: Physics Augmented Continuum Neural Radiance Fields for Geometry-Agnostic System Identification
Xuan Li (University of California Los Angeles), Chuang Gan (University of Massachusetts Amherst)
Neural Radiance FieldVideoPhysics Related
🎯 What it does: Simultaneously estimate the geometric and physical parameters of unknown objects through multi-view video.
Packed Ensembles for efficient uncertainty estimation
Olivier Laurent (University of Paris-Saclay), Gianni Franchi (Institut Polytechnique de Paris)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: The Packed-Ensembles (PE) framework is proposed, which parallelizes multiple sub-networks within a single network using grouped convolutions and masks, achieving efficient ensemble and improving uncertainty estimation.
PaLI: A Jointly-Scaled Multilingual Language-Image Model
Xi Chen (Google Research), Radu Soricut (Google Research)
GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Designed and trained PaLI, a unified multilingual visual-text generation model capable of simultaneously handling image, text, and image+text tasks.
PandA: Unsupervised Learning of Parts and Appearances in the Feature Maps of GANs
James Oldfield (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A novel unsupervised, architecture-agnostic semi-nonnegative tensor decomposition method is proposed, which jointly learns interpretable components and appearances in GAN feature maps, enabling pixel-level local editing and concept localization.
Panning for Gold in Federated Learning: Targeted Text Extraction under Arbitrarily Large-Scale Aggregation
Hong-Min Chu (University of Maryland), Tom Goldstein (University of Maryland)
Federated LearningAdversarial AttackTransformerText
🎯 What it does: This paper proposes a targeted information extraction attack (panning) based on Transformer models in federated learning, which maliciously modifies parameters to label and filter sequences containing keywords and recover complete sentences from aggregated gradients.
Parallel Deep Neural Networks Have Zero Duality Gap
Yifei Wang (Stanford University), Mert Pilanci (Stanford University)
Optimization
🎯 What it does: This paper studies the convex duality of deep neural networks, proving that standard deep networks have a dual gap, while parallel networks can achieve zero dual gap under appropriate regularization, leading to a globally optimal convex equivalent problem.
Parameter-Efficient Fine-Tuning Design Spaces
Jiaao Chen (Georgia Institute of Technology), Diyi Yang (Amazon Web Services)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Systematically explore the design space of parameter-efficient fine-tuning, identifying and validating a series of design patterns, and applying them to various pre-trained models and tasks.
Parametrizing Product Shape Manifolds by Composite Networks
Josua Sassen (University of Bonn), Benedikt Wirth (University of Münster)
Graph Neural NetworkMesh
🎯 What it does: Learning and approximating the Riemannian exponential map of shape space parameterization, proposing to decompose the shape space into low-dimensional factors and approximate it using composite networks.
Pareto Invariant Risk Minimization: Towards Mitigating the Optimization Dilemma in Out-of-Distribution Generalization
Yongqiang Chen (Chinese University of Hong Kong), James Cheng (Hong Kong Baptist University)
Domain AdaptationOptimizationImageBenchmark
🎯 What it does: In this work, the authors propose a new optimization framework called PAIR (Pareto Invariant Risk Minimization), aimed at addressing the optimization dilemma in the OOD (Out-of-Distribution) generalization process. By adopting a multi-objective optimization (MOO) perspective, it jointly considers ERM and OOD objectives, employing an adaptive weighting mechanism to dynamically balance both during the optimization process, ultimately achieving the recovery and enhancement of the original IRM objective. Additionally, the authors designed a corresponding model selector PAIR-s to better capture OOD performance during the validation phase.
Part-Based Models Improve Adversarial Robustness
Chawin Sitawarin (University of California), David Wagner (University of California)
ClassificationSegmentationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: By combining human prior knowledge with end-to-end learning, a part-based model is introduced to enhance the robustness of image classification.
Partial Label Unsupervised Domain Adaptation with Class-Prototype Alignment
Yan Yan (Carleton University), Yuhong Guo (CIFAR AI Chair, Amii)
Domain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the problem of Partial Label Unsupervised Domain Adaptation (PLUDA) in the source domain and introduces the PAPLUDA method based on class prototype alignment.
Partially Observable RL with B-Stability: Unified Structural Condition and Sharp Sample-Efficient Algorithms
Fan Chen (Peking University), Song Mei (University of California Berkeley)
Reinforcement Learning
🎯 What it does: A new structural condition called B-stability is proposed, which unifies and generalizes many existing sample-efficient subclasses of POMDP/PSR, and three model-based learning algorithms (OMLE, Explorative E2D, MOPS) are provided to achieve near-optimal policy learning for B-stable PSR.
Particle-based Variational Inference with Preconditioned Functional Gradient Flow
Hanze Dong (Hong Kong University of Science and Technology), Tong Zhang (Hong Kong University of Science and Technology)
Flow-based ModelTabular
🎯 What it does: A particle variational inference framework based on Preconditioned Functional Gradient Flow (PFG) is proposed, which can directly approximate the preconditioned Wasserstein gradient.
PASHA: Efficient HPO and NAS with Progressive Resource Allocation
Ondrej Bohdal (University of Edinburgh), Giovanni Zappella (AWS)
OptimizationComputational EfficiencyHyperparameter SearchNeural Architecture SearchImage
🎯 What it does: PASHA significantly reduces training costs by dynamically adjusting the maximum resource limits in multi-fidelity search to stop unpromising configuration evaluations earlier;
Patch-Level Contrasting without Patch Correspondence for Accurate and Dense Contrastive Representation Learning
Shaofeng Zhang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Object DetectionSegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes ADCLR, a framework for contrastive learning using query crops in visual Transformers, combining global and local contrastive losses, and employing unidirectional cross-attention to avoid collapse.
PatchDCT: Patch Refinement for High Quality Instance Segmentation
Qinrou Wen (Zhejiang University), Kewei Liang (Shanghai Jiao Tong University)
Object DetectionSegmentationSupervised Fine-TuningImage
🎯 What it does: Proposes PatchDCT, a multi-stage patch-based segmentation refinement framework based on compressed vectors for high-quality instance segmentation;
PD-MORL: Preference-Driven Multi-Objective Reinforcement Learning Algorithm
Toygun Basaklar (University of Wisconsin Madison), Umit Ogras
OptimizationReinforcement LearningSequential
🎯 What it does: A preference-driven multi-objective reinforcement learning algorithm (PD-MORL) is proposed, which can cover the entire preference space with a unified network, avoiding the need to train separate policies for each preference.
PEER: A Collaborative Language Model
Timo Schick (Meta AI Research), Sebastian Riedel (University College London)
GenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A collaborative writing language model named PEER has been designed and trained, capable of generating plans, executing edits, providing explanations, and repeating in iterative steps, thus achieving various editing tasks such as adding, deleting, modifying text, and citing.
Perfectly Secure Steganography Using Minimum Entropy Coupling
Christian Schroeder de Witt (University of Oxford), Martin Strohmeier (Armasuisse Science and Technology)
OptimizationSafty and PrivacyTransformerLarge Language ModelImageTextAudio
🎯 What it does: This paper proposes an information-theoretic steganography method based on Minimum Entropy Coupling (MEC), which can achieve perfectly secure steganography on any cover text distribution while maintaining high coding efficiency.
PerFedMask: Personalized Federated Learning with Optimized Masking Vectors
Mehdi Setayesh (University of British Columbia), Vincent W.S. Wong (University of British Columbia)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: A personalized federated learning framework called PerFedMask is proposed, which utilizes optimized masking vectors to freeze parts of the network based on device computing capabilities and fine-tunes after training.
Performance Bounds for Model and Policy Transfer in Hidden-parameter MDPs
Haotian Fu (Brown University), George Konidaris (Brown University)
OptimizationReinforcement Learning
🎯 What it does: In the Hidden-Parameter MDP (HiP-MDP) framework, this paper conducts a theoretical analysis of the robustness of two types of algorithms: model transfer and policy transfer. It proves that the value function can be approximated under the condition of Lipschitz continuity, and subsequently provides upper bounds on the regret for both transfer methods, demonstrating that this upper bound is tight under linear deterministic limits. The theoretical predictions are then validated through two continuous control experiments (ball-goal and ball-wind).
Personalized Federated Learning with Feature Alignment and Classifier Collaboration
Jian Xu (Tsinghua University), Shao-Lun Huang (Tsinghua University)
Federated LearningImage
🎯 What it does: This study investigates the implementation of personalized models in federated learning through global feature alignment and classifier collaboration.
Personalized Reward Learning with Interaction-Grounded Learning (IGL)
Jessica Maghakian (Stony Brook University), Cheng Tan (Microsoft Research)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: A personalized reward learning framework based on Interaction Grounded Learning (IGL) called IGL-P is proposed in the recommendation system, which can learn each user's implicit satisfaction through user interaction feedback without explicit reward signals, thereby directly optimizing the user's real experience.
Pessimism in the Face of Confounders: Provably Efficient Offline Reinforcement Learning in Partially Observable Markov Decision Processes
Miao Lu (University of Science and Technology of China), Zhuoran Yang (Yale University)
Reinforcement Learning
🎯 What it does: The P3O algorithm is proposed to address the offline reinforcement learning problem when observational data is confounded due to hidden states;
PGrad: Learning Principal Gradients For Domain Generalization
Zhe Wang (University of Virginia), Yanjun Qi (University of Virginia)
Domain AdaptationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Principal Gradient (PGrad) update strategy based on training trajectories. By sequentially iterating over different training domains to obtain parameter trajectories, and then performing principal component analysis (SVD) on these trajectories, robust gradients are aggregated along the principal directions to update the model, thereby enhancing generalization to unseen domains.
Phase transition for detecting a small community in a large network
Jiashun Jin (Carnegie Mellon University), Anru Zhang
Graph
🎯 What it does: This paper studies the global detection problem of small communities in extremely unbalanced networks, proposes and analyzes the SgnQ statistic, provides its limiting distribution, power threshold, information lower bound, and computational lower bound, and compares it with traditional χ² tests and scan statistics.
Phase2vec: dynamical systems embedding with a physics-informed convolutional network
Matt Ricci, Mor Nitzan (Hebrew University)
Convolutional Neural NetworkTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: An unsupervised low-dimensional embedding method called phase2vec has been developed, which can learn physically meaningful dynamic system representations from vector field data and can be used to reconstruct system equations.
Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions
Ruben Villegas (Google Brain), Dumitru Erhan (Google Brain)
GenerationData SynthesisTransformerLarge Language ModelGenerative Adversarial NetworkVideoTextMultimodality
🎯 What it does: The Phenaki model is proposed, which can generate high-quality long-duration videos based on variable-length, temporally variable text stories.
PiFold: Toward effective and efficient protein inverse folding
Zhangyang Gao (Westlake University), Stan Z. Li (Westlake University)
Protein Structure PredictionGraph Neural NetworkGraph
🎯 What it does: A protein structure inverse folding model named PiFold is proposed, which can generate the corresponding amino acid sequence all at once given a 3D structure.
Pink Noise Is All You Need: Colored Noise Exploration in Deep Reinforcement Learning
Onno Eberhard (Max Planck Institute for Intelligent Systems), Georg Martius (Max Planck Institute for Intelligent Systems)
Reinforcement LearningSequential
🎯 What it does: Conduct a systematic experimental evaluation of action noise in continuous action reinforcement learning, comparing white noise, OU noise, and colored noise with different color parameters, and propose pink noise (β=1) as a universal default noise.
PINTO: Faithful Language Reasoning Using Prompt-Generated Rationales
PeiFeng Wang, Xiang Ren (University of Southern California)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A two-stage reasoning pipeline is designed and implemented, where a medium-sized frozen model first generates free-text reasoning processes (rationale) using prompts, and then a small fine-tuned model predicts answers based on the generated reasoning process.
Pitfalls of Gaussians as a noise distribution in NCE
Holden Lee (Johns Hopkins University), Andrej Risteski (Carnegie Mellon University)
OptimizationContrastive Learning
🎯 What it does: This paper discusses the shortcomings of using Gaussian distribution as the noise distribution in Noise Contrastive Estimation (NCE), pointing out that this choice may lead to an exponential deterioration of the condition number of the Hessian matrix in high-dimensional spaces, thereby affecting the statistical and algorithmic efficiency of NCE.
Planckian Jitter: countering the color-crippling effects of color jitter on self-supervised training
Simone Zini (University of Milano - Bicocca), Joost van de weijer
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A Planckian Jitter color augmentation method based on a physical blackbody radiation model is proposed for self-supervised learning.
Planning Goals for Exploration
Edward S. Hu (University of Pennsylvania), Dinesh Jayaraman (University of Pennsylvania)
OptimizationRobotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: This paper proposes a planning exploration target method based on a world model called PEG, which directly optimizes during the training phase to enable GCRL agents to generate high exploration value trajectories in the Go-Explore mode, thereby accelerating multi-task learning in unlabelled task environments.
Planning with Large Language Models for Code Generation
Shun Zhang (MIT IBM Watson AI Lab), Chuang Gan (UMass Amherst)
GenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes Planning-Guided Transformer Decoding (PG-TD), a code generation decoding framework that combines large language model Transformers with MCTS-based planning algorithms.
Planning with Sequence Models through Iterative Energy Minimization
Hongyi Chen (Georgia Institute of Technology), Patricio A. Vela (Massachusetts Institute of Technology)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningSequential
🎯 What it does: A planning framework named LEAP is proposed, which utilizes sequence models for iterative energy minimization to generate and refine action trajectories, overcoming the limitation of traditional autoregressive generation that can only predict the next step at a time;