arXivSub Start free trial

NeurIPS 2023 Papers — Page 15

Conference on Neural Information Processing Systems · 3218 papers

Hypernetwork-based Meta-Learning for Low-Rank Physics-Informed Neural Networks

Woojin Cho (Yonsei University), Noseong Park (Yonsei University)

OptimizationMeta LearningTime SeriesPhysics Related

🎯 What it does: A two-stage meta-learning framework combining low-rank physics-informed neural networks (LR-PINN) with hyper-networks is proposed for quickly approximating solutions in multi-query parameterized PDE scenarios.

Hypervolume Maximization: A Geometric View of Pareto Set Learning

Xiaoyuan Zhang (City University of Hong Kong), Qingfu Zhang (Hong Kong Baptist University)

Optimization

🎯 What it does: By constructing a neural network model with polar angles as input and transforming Pareto set learning into a geometric problem of maximizing hypervolume, the approach achieves approximation and prediction of the entire Pareto set.

Hypothesis Selection with Memory Constraints

Maryam Aliakbarpour (Rice University), Adam Smith (Boston University)

🎯 What it does: This paper proposes an algorithm to solve the Hypothesis Selection problem under memory-constrained conditions, primarily by constructing a Random Ladder Tournament and a Filter strategy, achieving a near-optimal trade-off between sample size and memory usage while ensuring that the total variation distance between the output distribution and the unknown distribution is close to the optimal value.

HyTrel: Hypergraph-enhanced Tabular Data Representation Learning

Pei Chen (Texas A&M University), George Karypis (Amazon Web Services)

Representation LearningTransformerContrastive LearningTabular

🎯 What it does: A hypergraph-based table language model called HYTREL is proposed, which utilizes the hypergraph structure to unify the representation of table cell units, rows, columns, and the entire table, and conducts table representation learning based on this.

IBA: Towards Irreversible Backdoor Attacks in Federated Learning

Dung Thuy Nguyen (Vanderbilt University), KOK SENG WONG

Federated LearningAdversarial AttackConvolutional Neural NetworkAuto EncoderImage

🎯 What it does: A two-stage irreversible backdoor attack (IBA) is proposed under the federated learning framework, which first learns a visually covert trigger generator and then gradually injects the backdoor into the global model without affecting normal classification performance.

ID and OOD Performance Are Sometimes Inversely Correlated on Real-world Datasets

Damien Teney (Idiap Research Institute), Ehsan Abbasnejad (University of Adelaide)

ClassificationDomain AdaptationSupervised Fine-TuningImageTextBiomedical DataMagnetic Resonance Imaging

🎯 What it does: The study found an inverse correlation between ID and OOD performance across various real datasets, challenging the traditional positive correlation assumption.

IDEA: An Invariant Perspective for Efficient Domain Adaptive Image Retrieval

Haixin Wang (Peking University), Xiao Luo (University of California)

RetrievalDomain AdaptationContrastive LearningImage

🎯 What it does: This paper addresses the problem of unsupervised domain adaptation retrieval and proposes the IDEA model, which decomposes each image into causal features and non-causal features, and uses causal features to generate domain-invariant binary hash codes.

Idempotent Learned Image Compression with Right-Inverse

Yanghao Li (Tsinghua University), Ya-Qin Zhang (Tsinghua University)

CompressionConvolutional Neural NetworkImage

🎯 What it does: A framework for idempotent learning image compression using right invertible transformations is proposed.

Identifiability Guarantees for Causal Disentanglement from Soft Interventions

Jiaqi Zhang (Massachusetts Institute of Technology), Caroline Uhler (Massachusetts Institute of Technology)

Auto EncoderBiomedical Data

🎯 What it does: Study the causal separability identifiability of unobserved causal variables under soft intervention data, and propose a variational autoencoder-based algorithm to achieve this separation.

Identifiable Contrastive Learning with Automatic Feature Importance Discovery

Qi Zhang (Peking University), Yisen Wang (Peking University)

Explainability and InterpretabilityRepresentation LearningContrastive LearningImage

🎯 What it does: This paper proposes triCL, a third-order contrastive learning method that incorporates a learnable diagonal importance matrix S into the contrastive loss, allowing features to be uniquely identified and ranked by importance.

Identification of Nonlinear Latent Hierarchical Models

Lingjing Kong (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)

TabularSequential

🎯 What it does: This paper proposes a theory of identifiability and a complete identification algorithm for nonlinear latent hierarchical causal models, which can simultaneously recover latent variables and causal structures (up to reversible transformations) based solely on observational data, and achieves this under relaxed structural and functional conditions;

IDRNet: Intervention-Driven Relation Network for Semantic Segmentation

Zhenchao Jin (University of Hong Kong), Lequan Yu (Peking University)

Object DetectionSegmentationConvolutional Neural NetworkImage

🎯 What it does: This paper proposes an Intervention Driven Relationship Network (IDRNet) based on Deletion Diagnostics, which first aggregates pixel-level features into semantic-level representations, and then updates the semantic-level relationship matrix using deletion diagnostics to enhance the mutual reinforcement of semantic-level representations. Finally, the enhanced representations are fed back to the pixel features to improve the pixel-level prediction accuracy of semantic segmentation.

IEBins: Iterative Elastic Bins for Monocular Depth Estimation

Shuwei Shao (Beihang University), Zhengguo Li (Institute for Infocomm Research)

Depth EstimationAutonomous DrivingRecurrent Neural NetworkTransformerImage

🎯 What it does: An Iterative Elastic Bin (IEBins) framework is proposed, which implements monocular depth estimation through multi-stage small bin iterative refinement search combined with a Swin Transformer encoder and a GRU iterative optimizer.

Ignorance is Bliss: Robust Control via Information Gating

Manan Tomar (University of Alberta), Philip Bachman (Microsoft Research Montreal)

Robotic IntelligenceReinforcement LearningContrastive LearningImage

🎯 What it does: Proposes the Information Gating (InfoGating) framework, which learns to retain only the minimal necessary information in control tasks by masking inputs or intermediate features through a differentiable signal-to-noise ratio parameterization, and jointly trains it with various downstream objectives (multi-step inverse dynamics, Q-learning, behavior cloning).

Im-Promptu: In-Context Composition from Image Prompts

Bhishma Dedhia (Princeton University), Niraj Jha

GenerationData SynthesisTransformerLarge Language ModelAuto EncoderImageBenchmark

🎯 What it does: The research achieves image context learning through visual analogical reasoning, proposing the Im-Promptu framework and three visual benchmarks;

Image Captioners Are Scalable Vision Learners Too

Michael Tschannen (Google DeepMind), Lucas Beyer (Google DeepMind)

GenerationRetrievalTransformerVision Language ModelContrastive LearningImageText

🎯 What it does: Evaluate the performance and capabilities of Contrastive Pre-training on Image-Text Pairs (CLIP) versus pure image captioning pre-training on visual encoders.

ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation

Yasheng SUN, Hideki Koike

Image TranslationRestorationGenerationTransformerDiffusion modelImageVideo

🎯 What it does: A visual example-based image editing framework called ImageBrush is proposed, which utilizes a pair of example images as visual instructions to edit the target image.

ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation

Jiazheng Xu (Tsinghua University), Yuxiao Dong (Tsinghua University)

GenerationReinforcement LearningDiffusion modelImageText

🎯 What it does: Constructed the ImageReward human preference scoring model and implemented the ReFL direct tuning strategy for diffusion models based on it.

Imagine That! Abstract-to-Intricate Text-to-Image Synthesis with Scene Graph Hallucination Diffusion

Shengqiong Wu (National University of Singapore), Tat-Seng Chua (National University of Singapore)

GenerationData SynthesisTransformerDiffusion modelImageText

🎯 What it does: A text-to-image (T2I) system based on scene graph (SG) hallucination is proposed, capable of generating complex and detailed images from brief abstract texts.

Imbalanced Mixed Linear Regression

Pini Zilber (Weizmann Institute of Science), Boaz Nadler (Weizmann Institute of Science)

Tabular

🎯 What it does: An iterative method called Mix-IRLS has been developed to solve the mixed linear regression problem, which can still maintain effective recovery in the case of imbalanced sample proportions.

Imitation Learning from Imperfection: Theoretical Justifications and Algorithms

Ziniu Li (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)

Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningImageVideo

🎯 What it does: A framework for offline imitation learning using supplementary non-expert data is proposed, along with theoretical analysis;

Imitation Learning from Vague Feedback

Xin-Qiang Cai (University of Tokyo), Masashi Sugiyama (RIKEN AIP)

Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningGenerative Adversarial NetworkSequential

🎯 What it does: Proposes the Vaguely Pairwise Imitation Learning (VPIL) framework, and presents the algorithm COMPILER/COMPILER-E, which reconstructs the expert occupancy distribution through risk rewriting and mixed ratio estimation under known or unknown expert ratio α, combined with adversarial IL methods like GAIL.

Implicit Bias of (Stochastic) Gradient Descent for Rank-1 Linear Neural Network

Bochen Lyu (DataCanvas), Zhanxing Zhu (Peking University)

OptimizationTabularStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This study investigates the implicit bias of rank-1 linear networks under Gradient Descent (GD) and Stochastic Gradient Descent (SGD), constructs potential functions to describe the convergence solutions of GD, and analyzes how network depth, initialization, and SGD noise collectively influence the final solution. It also compares this model with standard linear networks and diagonal linear networks, demonstrating that the rank-1 model aligns more closely with the behavior of standard linear networks.

Implicit Bias of Gradient Descent for Logistic Regression at the Edge of Stability

Jingfeng Wu (Johns Hopkins University), Jason D. Lee (Princeton University)

OptimizationTabular

🎯 What it does: This paper studies the convergence and implicit bias of constant step size gradient descent (GD) in logistic regression under the edge of stability (EoS) conditions, proving that despite local oscillations, GD can still minimize logistic loss over long time scales.

Implicit Bias of Gradient Descent for Two-layer ReLU and Leaky ReLU Networks on Nearly-orthogonal Data

Yiwen Kou (University of California), Quanquan Gu (University of California)

OptimizationImage

🎯 What it does: This study investigates the implicit bias of two-layer fully connected (Leaky) ReLU neural networks during gradient descent training, particularly for nearly orthogonal datasets.

Implicit Contrastive Representation Learning with Guided Stop-gradient

Byeongchan Lee (Gauss Labs), Sehyun Lee (KAIST)

Object DetectionSegmentationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: This paper proposes an implicit contrastive learning method using the asymmetric structure of Siamese networks without explicit contrastive loss, implemented through the 'guided stop-gradient (GSG)' method in SimSiam and BYOL.

Implicit Convolutional Kernels for Steerable CNNs

Maksim Zhdanov (University of Amsterdam), Gabriele Cesa (Qualcomm AI Research)

Convolutional Neural NetworkPoint Cloud

🎯 What it does: This paper proposes the use of implicit neural networks (G-equivariant multilayer perceptrons) to parameterize the convolution kernels of Steerable CNNs, which can easily achieve equivariance for any compact group and insert scalable features into the convolution kernels, greatly enhancing the model's expressiveness.

Implicit Differentiable Outlier Detection Enable Robust Deep Multimodal Analysis

Zhu Wang (University of Illinois at Chicago), Sathya N. Ravi (University of Illinois at Chicago)

Anomaly DetectionOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality

🎯 What it does: An implicit differentiable outlier detection layer is proposed, which combines external knowledge graphs to efficiently eliminate noisy concepts in multimodal models, enhancing training efficiency and robustness.

Implicit Manifold Gaussian Process Regression

Bernardo Fichera (École Polytechnique Fédérale de Lausanne), Aude Billard (École Polytechnique Fédérale de Lausanne)

TabularComputed Tomography

🎯 What it does: This paper proposes a technique (IMGP) that can adaptively infer implicit low-dimensional manifolds from both labeled and unlabeled data, and construct Gaussian process regression models on that manifold.

Implicit Regularization in Over-Parameterized Support Vector Machine

Yang Sui (Shanghai University of Finance and Economics), Yang Bai (Shanghai University of Finance and Economics)

OptimizationTabular

🎯 What it does: A high-dimensional support vector machine (SVM) algorithm without regularization is designed, achieving implicit regularization through over-parameterization and gradient descent.

Implicit Transfer Operator Learning: Multiple Time-Resolution Models for Molecular Dynamics

Mathias Schreiner (Chalmers University of Technology), Simon Olsson (Chalmers University of Technology)

Drug DiscoveryGraph Neural NetworkDiffusion modelSequential

🎯 What it does: Proposes the Implicit Transfer Operator (ITO) framework, which uses conditional denoising diffusion models to approximate molecular dynamics across multiple time scales;

Implicit variance regularization in non-contrastive SSL

Manu Srinath Halvagal (Friedrich Miescher Institute for Biomedical Research), Friedemann Zenke (Friedrich Miescher Institute for Biomedical Research)

OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: This study investigates the mechanism by which the predictor and stop-gradient prevent representation collapse through implicit variance regularization in non-contrastive self-supervised learning. It derives the feature space dynamics of Euclidean and cosine losses and proposes IsoLoss for isometric convergence based on this.

Implicit Variational Inference for High-Dimensional Posteriors

Anshuk Uppal (Technical University of Denmark), Jes Frellsen (Technical University of Denmark)

ClassificationGenerationDomain AdaptationGenerative Adversarial NetworkImageTabular

🎯 What it does: A linearized implicit variational inference (LIVI) method is proposed, which uses a neural sampler to approximate high-dimensional posteriors, avoiding the training of adversarial objectives;

Importance Weighted Actor-Critic for Optimal Conservative Offline Reinforcement Learning

Hanlin Zhu (University of California Berkeley), Jiantao Jiao (University of California Berkeley)

Reinforcement LearningTabular

🎯 What it does: This paper proposes A-Crab, a new offline reinforcement learning algorithm that utilizes an actor-critic structure combined with importance-weighted average Bellman error to achieve a relatively pessimistic evaluation of the policy.

Importance-aware Co-teaching for Offline Model-based Optimization

Ye Yuan (McGill University), Xue Liu (McGill University)

OptimizationMeta LearningReinforcement Learning from Human FeedbackNeural Architecture SearchTabular

🎯 What it does: A method called Importance-aware Co-Teaching (ICT) based on pseudo-label co-teaching and meta-learning weighting is proposed, which uses three proxy models for collaborative learning to enhance the optimization performance of offline models.

IMPRESS: Evaluating the Resilience of Imperceptible Perturbations Against Unauthorized Data Usage in Diffusion-Based Generative AI

Bochuan Cao (Pennsylvania State University), Jinghui Chen (Pennsylvania State University)

RestorationData SynthesisDiffusion modelImage

🎯 What it does: This paper studies the effectiveness of using invisible perturbations for data protection in diffusion models (such as Stable Diffusion) and proposes a platform called IMPRESS to evaluate and remove these perturbations.

Improved Algorithms for Stochastic Linear Bandits Using Tail Bounds for Martingale Mixtures

Hamish Flynn (Bosch Center for Artificial Intelligence), Jan Peters (Technische Universität Darmstadt)

OptimizationHyperparameter SearchTabular

🎯 What it does: A new adaptive Markov mixture tail bound is proposed for constructing confidence sequences for linear bandits, and based on this, two UCB algorithms (CMM-UCB and AMM-UCB) are implemented.

Improved Bayes Risk Can Yield Reduced Social Welfare Under Competition

Meena Jagadeesan (University of California Berkeley), Nika Haghtalab (University of California Berkeley)

ClassificationOptimizationConvolutional Neural NetworkImage

🎯 What it does: This paper discusses that improving data representation quality (Bayes risk) in a multi-model competitive environment does not necessarily enhance overall user prediction accuracy (social welfare) and may even lead to non-monotonic social losses.

Improved Bayesian Regret Bounds for Thompson Sampling in Reinforcement Learning

Ahmadreza Moradipari (Toyota Motor North America), Vaneet Aggarwal (Purdue University)

Reinforcement Learning

🎯 What it does: This paper addresses Bayesian reinforcement learning in time-inhomogeneous Markov decision processes, and for the first time provides upper bounds on Bayesian returns for Thompson Sampling (posterior sampling) under various settings, achieved through the construction of a discretized proxy environment and improved information ratio analysis.

Improved Best-of-Both-Worlds Guarantees for Multi-Armed Bandits: FTRL with General Regularizers and Multiple Optimal Arms

Tiancheng Jin (University of Southern California), Haipeng Luo (University of Southern California)

OptimizationReinforcement Learning

🎯 What it does: A general adaptive multi-armed bandit algorithm is proposed, capable of achieving optimal performance in both random and adversarial environments simultaneously;

Improved Communication Efficiency in Federated Natural Policy Gradient via ADMM-based Gradient Updates

Guangchen Lan (Purdue University), Vaneet Aggarwal (Purdue University)

Federated LearningComputational EfficiencyReinforcement LearningSequential

🎯 What it does: A new federated natural policy gradient algorithm, FedNPG-ADMM, is proposed, which estimates the global NPG direction through the Alternating Direction Method of Multipliers (ADMM) to reduce communication complexity.

Improved Convergence in High Probability of Clipped Gradient Methods with Heavy Tailed Noise

Ta Duy Nguyen (Boston University), Huy Nguyen

Optimization

🎯 What it does: This study investigates the high-probability convergence of stochastic gradient and mirror descent methods with gradient clipping in the presence of heavy-tailed noise.

Improved Frequency Estimation Algorithms with and without Predictions

Anders Aamand (Massachusetts Institute of Technology), Ali Vakilian (Toyota Technological Institute at Chicago)

Machine LearningTime Series

🎯 What it does: This paper proposes a new algorithm for frequency estimation on Zipfian streaming data, which can be used without prediction and can improve accuracy when gravity point predictions are available.

Improvements on Uncertainty Quantification for Node Classification via Distance Based Regularization

Russell Alan Hart (University of Texas at Dallas), Feng Chen (University of Texas at Dallas)

ClassificationAnomaly DetectionGraph Neural NetworkGraph

🎯 What it does: This paper addresses the uncertainty quantification in graph node classification and proposes and validates an improved method based on distance regularization, aiming to enhance the performance of OOD detection and misclassification detection.

Improving *day-ahead* Solar Irradiance Time Series Forecasting by Leveraging Spatio-Temporal Context

Oussama Boussif (Mila - Quebec AI Institute), Yoshua Bengio (Mila - Quebec AI Institute)

TransformerOptical FlowMultimodalityTime Series

🎯 What it does: A deep learning model based on satellite spatiotemporal context, CrossViViT, is proposed for predicting solar irradiance (GHI) time series one day in advance, achieving multi-quantile predictions to output confidence intervals.

Improving Adversarial Robustness via Information Bottleneck Distillation

Huafeng Kuang (Xiamen University), Rongrong Ji (Xiamen University)

Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: Improve the adversarial robustness of deep networks through Information Bottleneck Distillation (IBD) combined with adversarial training.

Improving Adversarial Transferability via Intermediate-level Perturbation Decay

Qizhang Li (Harbin Institute of Technology), Hao Chen (University of California Davis)

Adversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This paper proposes a single-stage intermediate layer attack method called ILPD, which significantly enhances the transferability of attacks on unknown target models by attenuating perturbations in the feature space, aligning the perturbation direction with the guidance, and increasing the perturbation magnitude.

Improving CLIP Training with Language Rewrites

Lijie Fan (Google Research), Yonglong Tian (Google Research)

RetrievalRepresentation LearningTransformerLarge Language ModelContrastive LearningImageTextMultimodality

🎯 What it does: The contextual learning ability of large language models is used to generate multiple rewritten versions of each image description in CLIP training, forming text data augmentation. During training, original or rewritten text is randomly selected to pair with images for contrastive learning.

Improving Compositional Generalization using Iterated Learning and Simplicial Embeddings

Yi Ren (University of British Columbia), Aaron Courville (Université de Montréal)

GenerationRepresentation LearningGraph Neural NetworkSupervised Fine-TuningGraph

🎯 What it does: This paper proposes a framework that combines Iterated Learning and Simplicial Embeddings (SEM) to enhance the performance of deep networks in compositional generalization tasks.

Improving Diffusion-Based Image Synthesis with Context Prediction

Ling Yang (Peking University), Bin CUI

RestorationGenerationData SynthesisTransformerDiffusion modelImageText

🎯 What it does: This paper proposes a context prediction framework named CONPREDIFF, which utilizes an additional context decoder during the training phase of the diffusion model to force each pixel/feature/token to predict its local neighborhood context, thereby improving the quality and diversity of image generation.

Improving Few-Shot Generalization by Exploring and Exploiting Auxiliary Data

Alon Albalak (University of California, Santa Barbara), William Yang Wang (University of California, Santa Barbara)

Meta LearningTransformerReinforcement LearningText

🎯 What it does: This paper proposes a few-shot learning framework utilizing auxiliary data (FLAD), which enhances the generalization performance of the target task by dynamically selecting appropriate auxiliary datasets during the training process.

Improving Graph Matching with Positional Reconstruction Encoder-Decoder Network

Yixiao Zhou (Peking University), Xiaoqing Lyu (Peking University)

RecognitionObject DetectionOptimizationGraph Neural NetworkContrastive LearningGraph

🎯 What it does: This paper proposes an end-to-end graph matching network PREGM, which learns the spatial location information of keypoint graphs through a Position Reconstruction Encoder-Decoder (PR-EnDec), thereby enhancing the semantic keypoint matching performance.

Improving Language Plasticity via Pretraining with Active Forgetting

Yihong Chen (University College London), Mikel Artetxe (Reka AI)

TransformerLarge Language ModelText

🎯 What it does: A proactive forgetting mechanism is proposed, which resets the word embedding layer every K steps during pre-training to enhance the linguistic plasticity of the language model.

Improving neural network representations using human similarity judgments

Lukas Muttenthaler (Google DeepMind), Simon Kornblith (Google DeepMind)

Anomaly DetectionRepresentation LearningTransformerContrastive LearningImage

🎯 What it does: This paper introduces a linear transformation (gLocal transform) that aligns the global representation of neural networks with human similarity judgments while preserving the local structure of the original representation, thereby enhancing the performance of downstream tasks.

Improving Robustness with Adaptive Weight Decay

Amin Ghiasi (Apple), Reza Ardekani (Apple)

ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This paper proposes Adaptive Weight Decay (AWD), which dynamically adjusts λ_wd in each iteration based on the ratio of the magnitudes of the cross-entropy gradient and the weight decay gradient, thereby reducing robustness overfitting and enhancing robustness in adversarial training.

Improving Self-supervised Molecular Representation Learning using Persistent Homology

Yuankai Luo (Beihang University), Veronika Thost (Massachusetts Institute of Technology IBM Watson Artificial Intelligence Lab IBM Research)

Representation LearningDrug DiscoveryGraph Neural NetworkAuto EncoderContrastive LearningGraph

🎯 What it does: A self-supervised learning framework based on Persistent Homology is proposed, which improves the molecular embedding space by providing multi-scale views and distance constraints through Persistence Images, and presents two methods: Topological Fingerprints AutoEncoder (TAE) and Topological Distance Contrastive Loss (TDL).

Improving the Knowledge Gradient Algorithm

Le Yang (City University of Hong Kong), Chin Pang Ho (City University of Hong Kong)

OptimizationDrug DiscoveryTabular

🎯 What it does: Proposes and evaluates an improved version of the Knowledge Gradient (iKG) algorithm for optimal arm identification under a fixed budget and its variants (ϵ-good arm identification and feasible arm identification).

Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear Learners

Rachel Emily Redberg (University of California Santa Barbara), Yu-Xiang Wang (University of California Santa Barbara)

OptimizationSafty and PrivacyTabular

🎯 What it does: In this paper, the authors redesigned the objective perturbation mechanism, provided a more compact (ε,δ)-DP and RDP analysis, and extended its applicability through gradient clipping and approximate minimization techniques; they also proposed a private convex optimization algorithm that achieves sublinear time complexity (O(n log n)).

In Defense of Softmax Parametrization for Calibrated and Consistent Learning to Defer

Yuzhou Cao (Nanyang Technological University), Bo An (Nanyang Technological University)

ClassificationConvolutional Neural NetworkImage

🎯 What it does: In the Learning to Defer framework, we propose an asymmetric softmax consistency approximation loss that allows the model to decide whether to defer to an expert while also outputting calibrated classification probabilities and expert accuracy.

In-Context Impersonation Reveals Large Language Models' Strengths and Biases

Leonard Salewski (University of Tübingen), Zeynep Akata (University of Tübingen)

TransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodality

🎯 What it does: The study investigates the impact of having LLMs assume different identities (age, profession, race, gender) on their reasoning, decision-making, and visual description capabilities.

In-Context Learning Unlocked for Diffusion Models

Zhendong Wang (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)

SegmentationGenerationDepth EstimationPrompt EngineeringVision Language ModelDiffusion modelImageMultimodality

🎯 What it does: Proposes and trains Prompt Diffusion, a diffusion model that accepts visual-language prompts, achieving contextual learning for six visual tasks and generalizing to unseen tasks.

Incentives in Federated Learning: Equilibria, Dynamics, and Mechanisms for Welfare Maximization

Aniket Murhekar (University of Illinois), Ruta Mehta (University of Illinois)

Federated LearningImage

🎯 What it does: This paper proposes a federated learning framework based on a budget-balanced mechanism, allowing participants to maximize individual utility while achieving Nash equilibrium through best response dynamics, thereby enhancing overall welfare.

Incentives in Private Collaborative Machine Learning

Rachael Hwee Ling Sim (National University of Singapore), Patrick Jaillet (Massachusetts Institute of Technology)

Federated LearningSafty and PrivacyTabular

🎯 What it does: A mechanism incentivized by differential privacy is proposed in collaborative machine learning, utilizing sufficient statistics (SS) from all parties for cooperation, measuring data value based on Bayesian surprise, and generating different model parameter sample rewards for each party.

Incentivized Communication for Federated Bandits

Zhepei Wei (University of Virginia), Hongning Wang (University of Virginia)

Federated Learning

🎯 What it does: Proposed and solved the incentive communication problem in federated Bandit learning, constructed the INC-FEDUCB framework, and designed two incentive mechanisms: payment-free and payment-efficient.

Incentivizing Honesty among Competitors in Collaborative Learning and Optimization

Florian E. Dorner (Max Planck Institute for Intelligent Systems), Martin Vechev (ETH Zurich)

OptimizationFederated LearningText

🎯 What it does: A competitive player game model is introduced in the federated learning framework to analyze its impact on model training and design mechanisms to incentivize all parties to maintain honest updates, thereby ensuring the benefits of collaborative learning.

Incomplete Multimodality-Diffused Emotion Recognition

Yuanzhi Wang (Nanjing University of Science and Technology), Zhen Cui (Nanjing University of Science and Technology)

RecognitionTransformerDiffusion modelScore-based ModelMultimodalityStochastic Differential Equation

🎯 What it does: A method for multimodal emotion recognition with missing integrity, named IMDer, is proposed. It utilizes a score-based diffusion model to recover the missing modality under the condition of observed modalities, and then inputs the complete modalities into a multimodal transformer for emotion prediction.

Inconsistency, Instability, and Generalization Gap of Deep Neural Network Training

Rie Johnson (RJ Research Consulting), Tong Zhang (Hong Kong University of Science and Technology)

OptimizationKnowledge DistillationConvolutional Neural NetworkImage

🎯 What it does: This paper proposes a theoretical framework that associates the generalization error of deep networks with the inconsistency and instability of model outputs, and provides a generalization upper bound that can be estimated on unlabeled data; it is subsequently validated through large-scale experiments showing that these two metrics are highly correlated with generalization error across various datasets and network architectures, and that inconsistency is a better predictor of generalization than traditional sharpness metrics; finally, consistency encouragement (co-distillation, SAM + consistency, ensemble, distillation) is integrated into training to further enhance performance.

Individual Arbitrariness and Group Fairness

Carol Xuan Long (Harvard University), Flavio Calmon

Tabular

🎯 What it does: This paper studies the significant increase in individual-level randomness (predictive multiplicity) of model predictions after implementing group fairness interventions in machine learning, and proposes a voting ensemble-based algorithm to reduce this multiplicity, ensuring more stable model predictions while maintaining fairness and accuracy.

Individualized Dosing Dynamics via Neural Eigen Decomposition

Stav Belogolovsky (Technion Israel Institute of Technology), Shie Mannor (Nvidia Research)

Drug DiscoveryRecurrent Neural NetworkTime SeriesBiomedical DataElectronic Health RecordsStochastic Differential Equation

🎯 What it does: A personalized continuous dose pharmacokinetic model NESDE based on neural feature decomposition has been developed, specifically for time series prediction in medical dose control;

Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

Kenneth Li (Harvard University), Martin Wattenberg (Harvard University)

GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText

🎯 What it does: This paper proposes an Inference-Time Intervention (ITI) technique that enhances the sincerity of large language models by fine-tuning the activation vectors of attention heads during the inference process.

Inferring Hybrid Neural Fluid Fields from Videos

Hong-Xing Yu (Stanford University), Jiajun Wu (Stanford University)

GenerationData SynthesisNeural Radiance FieldVideoPhysics Related

🎯 What it does: Inferring fluid density and velocity from sparse multi-view videos using a hybrid neural fluid field (HyFluid).

Inferring the Future by Imagining the Past

Kartik Chandra (Massachusetts Institute of Technology), Joshua B. Tenenbaum

OptimizationComputational EfficiencyReinforcement LearningGraph

🎯 What it does: Proposes an algorithm based on inverse planning for inferring the past and future from a single frame observation.

InfoCD: A Contrastive Chamfer Distance Loss for Point Cloud Completion

Fangzhou Lin (Worcester Polytechnic Institute), Venkatesh Saligrama (Boston University)

GenerationData SynthesisContrastive LearningPoint Cloud

🎯 What it does: A new contrastive Chamfer distance loss, InfoCD, is proposed for point cloud completion.

InfoPrompt: Information-Theoretic Soft Prompt Tuning for Natural Language Understanding

Junda Wu (University of California San Diego), Ricardo Henao (Duke University)

ClassificationRecognitionTransformerPrompt EngineeringContrastive LearningText

🎯 What it does: To address the initialization and information utilization issues in soft prompt tuning, a Prompt training framework based on information theory, called InfoPrompt, is proposed. It utilizes mutual information maximization to guide prompt learning to be more task-relevant.

Information Design in Multi-Agent Reinforcement Learning

Yue Lin (Chinese University of Hong Kong), Baoxiang Wang (Chinese University of Hong Kong)

Recommendation SystemReinforcement Learning

🎯 What it does: A Markov Signaling Game framework is proposed to study information design in Multi-Agent Reinforcement Learning (MARL); it designs a signaling gradient that considers the receiver's strategy chain and learnable extended obedience constraints, enabling the sender to influence the receiver's adaptive learning through information in a mixed-motivation environment.

Information Geometry of the Retinal Representation Manifold

Xuehao Ding (Stanford University), Stephen Baccus

Convolutional Neural NetworkImage

🎯 What it does: This study constructs a three-layer convolutional neural network (CNN) random encoding model with independent noise to model the responses of retinal populations to natural image stimuli, and analyzes the discernible features of the retinal representational manifold using information geometry (Fisher information metric);

Information Maximization Perspective of Orthogonal Matching Pursuit with Applications to Explainable AI

Aditya Chattopadhyay (Johns Hopkins University), Rene Vidal

ClassificationExplainability and InterpretabilityComputational EfficiencyContrastive LearningImage

🎯 What it does: This paper provides an in-depth comparison between Information Pursuit (IP) and Orthogonal Matching Pursuit (OMP), demonstrating that OMP is a special case of IP, and based on this, proposes the IP-OMP algorithm; subsequently, it applies IP-OMP to interpretable predictions in visual classification by combining CLIP text/image embeddings, generating sparse explanations based on semantic concepts.

Information Maximizing Curriculum: A Curriculum-Based Approach for Learning Versatile Skills

Denis Blessing (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)

Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningMixture of ExpertsSequential

🎯 What it does: An Information Maximization Curriculum (IMC) framework is proposed, which dynamically focuses on representable data subsets using curriculum weights and extends it to a mixture of experts model to address performance degradation caused by modality averaging.

Information Theoretic Lower Bounds for Information Theoretic Upper Bounds

Roi Livni (Tel Aviv University)

Optimization

🎯 What it does: In the framework of Stochastic Convex Optimization (SCO), the study investigates and proves the lower bound of mutual information required by learning algorithms, revealing the necessity of the relationship between information-theoretic generalization bounds and dimensions.

Information-guided Planning: An Online Approach for Partially Observable Problems

Matheus Aparecido Do Carmo Alves, Leandro Soriano Marcolino (Lancaster University)

OptimizationReinforcement Learning

🎯 What it does: An information entropy-guided POMCP algorithm (IB-POMCP) is proposed, which adaptively adjusts the exploration coefficient through entropy estimation in partially observable environments and improves action selection using the I-UCB function to achieve online planning for sparse reward problems.

Initialization Matters: Privacy-Utility Analysis of Overparameterized Neural Networks

Jiayuan Ye (National University of Singapore), Volkan Cevher (EPFL)

OptimizationSafty and PrivacyImageStochastic Differential Equation

🎯 What it does: This study analyzes the upper bound of KL privacy leakage for fully connected ReLU networks (and their linearized versions) trained with Langevin diffusion without using gradient clipping, and explores the impact of initialization methods, network width, and depth on the privacy-utility trade-off.

Initialization-Dependent Sample Complexity of Linear Predictors and Neural Networks

Roey Magen (Weizmann Institute of Science), Ohad Shamir (Weizmann Institute of Science)

🎯 What it does: This paper discusses the sample complexity of matrix parameter linear predictors and feedforward neural networks, providing size-independent upper and lower bounds based on the Frobenius norm and the distance to a reference matrix.

Injecting Multimodal Information into Rigid Protein Docking via Bi-level Optimization

Ruijia Wang (BioMap Research), Le Song (BioMap Research)

OptimizationProtein Structure PredictionTransformerMultimodalityBiomedical DataBenchmark

🎯 What it does: The BiDock model is proposed, which predicts rigid protein-protein docking through dual-layer optimization combined with cross-modal Transformer.

Inner Product-based Neural Network Similarity

Wei Chen (Purdue University), Qiang Qiu (Purdue University)

Federated LearningComputational EfficiencyRepresentation LearningConvolutional Neural NetworkImage

🎯 What it does: This paper proposes a similarity measure based on convolutional filter atomic subspaces to efficiently evaluate the representation similarity between isomorphic CNN models.

Inner-Outer Aware Reconstruction Model for Monocular 3D Scene Reconstruction

Yu-Kun Qiu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)

SegmentationDepth EstimationTransformerImage

🎯 What it does: This paper proposes an Inner-Outer Perception Reconstruction model (IOAR), which achieves more accurate 3D scene reconstruction from monocular images by distinguishing between three types of voxels: surface voxels, inner surface voxels, and outer surface voxels.

InsActor: Instruction-driven Physics-based Characters

Jiawei Ren (Nanyang Technological University), Ziwei Liu

GenerationRobotic IntelligenceTransformerDiffusion modelTextMultimodalityPhysics Related

🎯 What it does: Achieve human-instructed physical simulation character animation generation through a hierarchical diffusion model and skill discovery.

Inserting Anybody in Diffusion Models via Celeb Basis

Ge Yuan (Sun Yat-sen University), Huicheng Zheng (Sun Yat-sen University)

GenerationData SynthesisDiffusion modelImage

🎯 What it does: Quickly inject a new identity into a pre-trained Stable Diffusion using a single face photo, enabling arbitrary poses and interactive generation with other concepts.

InstanT: Semi-supervised Learning with Instance-dependent Thresholds

Muyang Li (University of Sydney), Tongliang Liu (University of Sydney)

ClassificationDomain AdaptationTransformerSupervised Fine-TuningImage

🎯 What it does: A pseudo-label selection method based on instance-dependent thresholds, called InstanT, is proposed, along with theoretical guarantees.

InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning

Wenliang Dai (Salesforce Research), Steven Hoi (Salesforce Research)

ClassificationRecognitionGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality

🎯 What it does: A framework for instruction tuning based on BLIP-2, called InstructBLIP, has been constructed to complete general visual language tasks using a multimodal language model.

Instructing Goal-Conditioned Reinforcement Learning Agents with Temporal Logic Objectives

Wenjie Qiu (Rutgers University), He Zhu (Rutgers University)

Robotic IntelligenceReinforcement LearningAgentic AI

🎯 What it does: This study investigates how to enable goal-conditioned reinforcement learning agents to complete tasks with zero samples upon receiving linear temporal logic (LTL) tasks.

Integration-free Training for Spatio-temporal Multimodal Covariate Deep Kernel Point Processes

YIXUAN ZHANG, Feng Zhou (Renmin University of China)

Score-based ModelMultimodalityTime Series

🎯 What it does: A point process model DKMPP that combines multimodal covariates with deep kernels has been designed and implemented, allowing for training without the need for integration.

Intensity Profile Projection: A Framework for Continuous-Time Representation Learning for Dynamic Networks

Alexander Modell (Imperial College London), Patrick Rubin-Delanchy (University of Bristol)

Representation LearningGraph Neural NetworkGraphTime Series

🎯 What it does: A continuous-time network representation learning framework (Intensity Profile Projection) is proposed, which can generate low-dimensional trajectories for each node that evolve over time.

Interaction Measures, Partition Lattices and Kernel Tests for High-Order Interactions

Zhaolu Liu (Imperial College London), Mauricio Barahona (Imperial College London)

TabularTime SeriesMagnetic Resonance Imaging

🎯 What it does: This paper proposes a high-order interaction measurement framework based on the theory of partition lattices, designs the Streitberg and Lancaster interaction tests, and implements non-parametric statistical tests through kernel mean embedding.

Interactive Multi-fidelity Learning for Cost-effective Adaptation of Language Model with Sparse Human Supervision

Jiaxin Zhang (Intuit AI Research), Sricharan Kumar (Intuit AI Research)

ClassificationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataFinance Related

🎯 What it does: Under a limited budget, an Interactive Multi-Precision Learning (IMFL) framework is proposed, which combines human and LLM (Large Language Model) annotation sources to iteratively fine-tune small-scale domain language models efficiently.

Interpretability at Scale: Identifying Causal Mechanisms in Alpaca

Zhengxuan Wu (Stanford University), Noah Goodman

Explainability and InterpretabilityLarge Language ModelTextFinance Related

🎯 What it does: This paper uses the Boundless DAS method to provide causal explanations for the Alpaca (7B) model, discovering that it internally uses two Boolean variables for left and right boundary checks when performing price tagging tasks.

Interpretable and Explainable Logical Policies via Neurally Guided Symbolic Abstraction

Quentin Delfosse (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)

OptimizationExplainability and InterpretabilityReinforcement LearningImage

🎯 What it does: This study proposes a reinforcement learning framework named NUDGE, which utilizes neural network-guided symbolic abstraction to generate interpretable logical policies, and achieves policy optimization through differentiable reasoning and PPO training.

Interpretable Graph Networks Formulate Universal Algebra Conjectures

Francesco Giannini (CINI), Pietro Barbiero (Università della Svizzera Italiana)

ClassificationExplainability and InterpretabilityGraph Neural NetworkReinforcement LearningGraph

🎯 What it does: This study proposes a complete AI framework that first automatically generates a finite lattice dataset suitable for training through algorithms, and constructs graphical representations for propositional properties in Universal Algebra (UA); subsequently, an interpretable Graph Neural Network (iGNN) layer is designed, which compresses node features into interpretable concepts using Gumbel-Softmax and makes decisions directly with a linear classifier; finally, classification experiments are conducted on the lattice dataset using this model to test and validate existing UA conjectures, while also discovering new subgraph patterns.

Interpretable Prototype-based Graph Information Bottleneck

Sangwoo Seo (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)

ClassificationExplainability and InterpretabilityGraph Neural NetworkContrastive LearningGraph

🎯 What it does: A novel explainable graph neural network framework called PGIB is developed, which combines prototype learning with the information bottleneck to automatically extract key subgraphs and provide visual explanations for graph classification tasks.

Interpretable Reward Redistribution in Reinforcement Learning: A Causal Approach

Yudi Zhang (Eindhoven University of Technology), Mykola Pechenizkiy (Eindhoven University of Technology)

Explainability and InterpretabilityRobotic IntelligenceReinforcement LearningSequential

🎯 What it does: This paper proposes a reward redistribution framework called Generative Return Decomposition (GRD) from a causal perspective, which decomposes delayed rewards by learning interpretable causal structures and generates interpretable proxy rewards and compact state representations to enhance policy learning.

Interpreting Unsupervised Anomaly Detection in Security via Rule Extraction

Ruoyu Li (Tsinghua University), Yong Yang (Tsinghua University)

Anomaly DetectionExplainability and InterpretabilityAuto EncoderTabular

🎯 What it does: This paper proposes a method for globally interpretable unsupervised anomaly detection models, utilizing rule extraction to construct interpretable and deployable rule sets.