ICLR 2024 Papers — Page 15
International Conference on Learning Representations · 2260 papers
Neural-Symbolic Recursive Machine for Systematic Generalization
Qing Li (National Key Laboratory of General Artificial Intelligence), Siyuan Huang
Convolutional Neural NetworkTransformerText
🎯 What it does: Designed and implemented a Neural Symbolic Recursive Machine (NSR), which achieves systematic generalization by self-learning a Grounded Symbol System (GSS) from data, enabling joint modeling of perception, syntax, and semantics.
Neuro-Inspired Information-Theoretic Hierarchical Perception for Multimodal Learning
Xiongye Xiao (University of Southern California), Paul Bogdan (University of Southern California)
ClassificationCompressionRepresentation LearningTransformerLarge Language ModelMultimodality
🎯 What it does: Proposes the Information Theory Hierarchical Perception (ITHP) model, which uses the information bottleneck to compress the main modality into a latent state while retaining relevant information from the remaining modalities, achieving multimodal fusion and feature compression.
NeuroBack: Improving CDCL SAT Solving using Graph Neural Networks
Wenxi Wang (University of Texas at Austin), Risto Miikkulainen (University of Texas at Austin)
Graph Neural NetworkGraph
🎯 What it does: Using graph neural networks to predict the variable phases of most satisfying assignments in SAT formulas in advance, thereby initializing the phase selection of the CDCL solver to improve solving performance;
Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data
Antonis Antoniades (University of California), Spencer Smith
GenerationData SynthesisTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Construct and train a multimodal multi-task generative pre-trained Transformer (Neuroformer) to autoregressively generate neural activity and predict behavior from cellular-level neural discharge, visual stimuli, and behavioral data.
Neuron Activation Coverage: Rethinking Out-of-distribution Detection and Generalization
Yibing Liu (City University of Hong Kong), Shiqi Wang (City University of Hong Kong)
Domain AdaptationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes the neuron activation state and neuron activation coverage (NAC), and utilizes NAC for OOD detection (NAC-UE) and OOD generalization evaluation (NAC-ME).
Neuron-Enhanced AutoEncoder Matrix Completion and Collaborative Filtering: Theory and Practice
Jicong Fan (Chinese University of Hong Kong), Chris Ding
Recommendation SystemAuto EncoderTabular
🎯 What it does: A neural network enhanced autoencoder matrix completion method AEMC-NE is proposed for collaborative filtering and missing value completion;
Neurosymbolic Grounding for Compositional World Models
Atharva Sehgal (University of Texas at Austin), Swarat Chaudhuri (University of Texas at Austin)
Object DetectionRobotic IntelligenceGraph Neural NetworkVision Language ModelAuto EncoderWorld ModelMultimodality
🎯 What it does: Designed and implemented COSMOS, an object-centered world model that utilizes neural symbolic encoding to perform accurate reasoning in unseen visual composite scenes.
NeurRev: Train Better Sparse Neural Network Practically via Neuron Revitalization
Gen Li (Clemson University), Xiaolong Ma (Clemson University)
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: The NeurRev framework is proposed to revive neurons by identifying and pruning large negative weights that lead to neuron 'exhaustion', thereby enhancing the learning effectiveness of dynamic sparse training.
Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors
Ido Amos (Tel Aviv University), Ankit Gupta (IBM Research)
TransformerSupervised Fine-TuningMultimodalityTime SeriesSequentialBiomedical Data
🎯 What it does: This paper evaluates the true capabilities of Transformers and State-Space Models (SSMs) in modeling long-range dependencies by comparing the effects of training with random initialization on long sequence tasks versus self-supervised pretraining (Self-Pretraining, SPT) on task data.
New Insight of Variance reduce in Zero-Order Hard-Thresholding: Mitigating Gradient Error and Expansivity Contradictions
Xinzhe Yuan (Harbin Institute of Technology), Huan Xiong (Harbin Institute of Technology)
OptimizationAdversarial AttackImage
🎯 What it does: A class of zero-order hard threshold algorithms based on variance reduction (pM-SZHT and VR-SZHT) is proposed, which estimates variance by reducing historical gradients, eliminating the strict limitation of the number of random directions q in traditional SZOHT, significantly improving convergence rate and query complexity.
NfgTransformer: Equivariant Representation Learning for Normal-form Games
Siqi Liu (Google DeepMind), Nicolas Heess (Google DeepMind)
OptimizationRepresentation LearningTransformer
🎯 What it does: A deep network named NfgTransformer is proposed to learn the representation of regularized, interpretable, and equivariant regularized games (NFG), which can be used for various game theory tasks such as Nash equilibrium solving, maximum deviation profit estimation, and ranking.
Node2ket: Efficient High-Dimensional Network Embedding in Quantum Hilbert Space
Hao Xiong (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationRepresentation LearningGraph Neural NetworkGraphPhysics Related
🎯 What it does: This paper proposes a method for constructing high-dimensional network embeddings in quantum Hilbert space, utilizing the 'product state' generated by tensor products to achieve exponential embedding dimensions, and achieving O(p) training complexity through an inner product objective; it also introduces a parameter-sharing compressed variant called node2ket+; develops a Riemannian-Adagrad optimizer with normalization and positive inner product constraints; and implements the LIBN2K C++ library.
Noise Map Guidance: Inversion with Spatial Context for Real Image Editing
Hansam Cho (Korea University), Yonghyun Jeong (NAVER Cloud)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: A non-optimized Noise Map Guidance (NMG) inverse method is proposed, which directly retains spatial context using noise maps, thereby enhancing the editing quality of real images.
Noise-free Score Distillation
Oren Katzir (Tel Aviv University), Dani Lischinski (Hebrew University of Jerusalem)
GenerationData SynthesisDiffusion modelScore-based ModelImageTextPoint Cloud
🎯 What it does: An improved noise-free score distillation (NFSD) method is proposed for generating text-to-image/3D content, addressing the noise interference issue in traditional score distillation methods (SDS);
NoiseDiffusion: Correcting Noise for Image Interpolation with Diffusion Models beyond Spherical Linear Interpolation
PengFei Zheng, Bo Han (Hong Kong Baptist University)
Image TranslationRestorationDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A NoiseDiffusion method is proposed, which performs noise correction on natural images in the diffusion model before interpolation, eliminating artifacts while preserving the original image information.
Noisy Interpolation Learning with Shallow Univariate ReLU Networks
Nirmit Joshi (TTI Chicago), Nathan Srebro (TTI Chicago)
Tabular
🎯 What it does: This paper provides a rigorous analysis of noisy interpolation learning, with a particular focus on the overfitting behavior of univariate two-layer ReLU networks in regression.
NOLA: Compressing LoRA using Linear Combination of Random Basis
Soroush Abbasi Koohpayegani (University of California), Hamed Pirsiavash (University of California)
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: NOLA achieves efficient compression and fine-tuning of LLM weights by reparameterizing the low-rank matrices in LoRA as a linear combination of random bases.
Non-Exchangeable Conformal Risk Control
António Farinhas (Instituto de Telecomunicações), Andre Martins
ClassificationOptimizationTextTime Series
🎯 What it does: A non-exchangeable conformal risk control method is proposed, which can provide an upper bound on the expected value of any monotonic loss function in the case of non-exchangeable data, and degenerates to traditional CRC when the data is exchangeable.
Non-negative Contrastive Learning
Yifei Wang (Peking University), Yisen Wang (Peking University)
RetrievalExplainability and InterpretabilityRepresentation LearningContrastive LearningImage
🎯 What it does: Non-negative Contrastive Learning (NCL) is proposed, which enhances interpretability and sparsity by enforcing non-negativity of features in contrastive learning.
Nougat: Neural Optical Understanding for Academic Documents
Lukas Blecher (Meta AI), Robert Stojnic (Meta AI)
RecognitionGenerationTransformerLarge Language ModelText
🎯 What it does: The Nougat model and its dataset generation pipeline are proposed, capable of directly converting the PDF pages of academic papers into a lightweight markup language (similar to LaTeX), without relying on traditional OCR.
Novel Quadratic Constraints for Extending LipSDP beyond Slope-Restricted Activations
Patricia Pauli (Institute for Systems Theory and Automatic Control, University of Stuttgart), Bin Hu (University of Illinois Urbana-Champaign)
Image
🎯 What it does: This study investigates how to extend LipSDP to estimate the Lipschitz constants of neural networks using MaxMin, GroupSort, and Householder activation functions.
NuwaDynamics: Discovering and Updating in Causal Spatio-Temporal Modeling
Kun Wang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
RestorationDomain AdaptationTransformerContrastive LearningTime SeriesSequential
🎯 What it does: Proposes the NuwaDynamics framework, which discovers causal patches in self-supervised reconstruction tasks and enhances the model's robustness to different distributions through intervention patches.
Object centric architectures enable efficient causal representation learning
Amin Mansouri (Mila Quebec AI Institute), Yoshua Bengio (Mila Quebec AI Institute)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a method that combines causal representation learning with object-centric learning, utilizing a Slot Attention-based object segmentation network and a weakly supervised sparse perturbation learning framework to achieve identifiable attribute decoupling in multi-object scenes.
Object-Aware Inversion and Reassembly for Image Editing
Zhen Yang (Zhejiang University), Chunhua Shen (Zhejiang University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: Proposes an Object-aware Inversion and Recombination (OIR) framework for multi-target text-driven image editing, which automatically determines the optimal inversion steps for each editing pair and achieves fine editing through separate inversion, editing, and recombination.
Object-Centric Learning with Slot Mixture Module
Daniil Kirilenko (Universita della Svizzera italiana), Aleksandr Panov
Representation LearningTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the Slot Mixture Module (SMM), a slot attention mechanism based on Gaussian Mixture Models (GMM) to improve object-centric representation learning.
Octavius: Mitigating Task Interference in MLLMs via LoRA-MoE
Zeren Chen (Shanghai AI Laboratory), Jing Shao (Shanghai AI Laboratory)
TransformerLarge Language ModelMixture of ExpertsMultimodalityPoint Cloud
🎯 What it does: This paper proposes the Octavius framework, which combines LoRA and Mixture-of-Experts (MoE) to construct a multimodal large language model decoder to address the issue of task interference.
OctoPack: Instruction Tuning Code Large Language Models
Niklas Muennighoff, Shayne Longpre
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper constructs a 4TB COMMITPACK dataset by utilizing the natural structure of Git commit messages and fine-tunes LLMs for code instructions. It proposes a multilingual evaluation benchmark, HUMANEVALPACK, covering code repair, explanation, and synthesis, and generates and releases two commercially usable code LLMs, OCTOCODER and OCTOGEEX, which perform best on this benchmark.
ODE Discovery for Longitudinal Heterogeneous Treatment Effects Inference
Krzysztof Kacprzyk (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
OptimizationExplainability and InterpretabilityDrug DiscoveryTime SeriesBiomedical DataOrdinary Differential Equation
🎯 What it does: This paper proposes a method for transforming ordinary differential equation (ODE) discovery into a framework for inferring long-term heterogeneous treatment effects, and implements the INSITE method based on this framework, which can predict treatment effects in an interpretable manner and is robust to irregular sampling without using neural networks.
ODEFormer: Symbolic Regression of Dynamical Systems with Transformers
Stéphane d'Ascoli (École Polytechnique Fédérale de Lausanne), Niki Kilbertus (Helmholtz Munich)
TransformerTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a Transformer-based ODEFormer model that can directly infer the symbolic expressions of multi-dimensional ordinary differential equations from a single noisy and irregularly sampled observation trajectory.
ODICE: Revealing the Mystery of Distribution Correction Estimation via Orthogonal-gradient Update
Liyuan Mao (Shanghai Jiao Tong University), Xianyuan Zhan (Institute for AI Industry Research)
Reinforcement LearningTabularBenchmark
🎯 What it does: The DICE (Distributionally Corrected Estimation) method has been researched and improved, introducing Orthogonal-Gradient Update in offline reinforcement learning and offline imitation learning.
Off-Policy Primal-Dual Safe Reinforcement Learning
Zifan Wu (Sun Yat-sen University), Dong Wang (Meituan)
OptimizationSafty and PrivacyReinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes an offline-prioritized master-slave safety reinforcement learning method called CAL, which integrates conservative policy optimization and local policy convexification.
Offline Data Enhanced On-Policy Policy Gradient with Provable Guarantees
Yifei Zhou (University of California), Wen Sun (Cornell University)
Reinforcement LearningImage
🎯 What it does: A new hybrid reinforcement learning algorithm is proposed, which integrates online policy updates based on natural policy gradients with offline data fitting for policy evaluation, thereby achieving the joint utilization of online and offline data.
Offline RL with Observation Histories: Analyzing and Improving Sample Complexity
Joey Hong (University of California Berkeley), Sergey Levine (University of California Berkeley)
Convolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This study investigates the sample complexity issue of offline reinforcement learning when using observation history in partially observable environments, and proposes a method to enhance the efficiency of offline RL by learning a compressed representation of observation history through bisimulation loss.
OMNI: Open-endedness via Models of human Notions of Interestingness
Jenny Zhang (University of British Columbia), Jeff Clune (University of British Columbia)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The OMNI framework is proposed, which utilizes pre-trained foundation models (FM) to estimate human interest levels, thereby filtering tasks that are both learnable and interesting in open learning.
OmniControl: Control Any Joint at Any Time for Human Motion Generation
Yiming Xie (Northeastern University), Huaizu Jiang (Northeastern University)
GenerationPose EstimationTransformerDiffusion modelVideo
🎯 What it does: We propose OmniControl, which can perform spatial control at any joint and any time point under text prompts, generating realistic full-body movements.
OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models
Wenqi Shao (OpenGVLab), Ping Luo (The University of Hong Kong)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A post-training quantization method named OmniQuant is proposed, which can improve the accuracy of large language models while maintaining low computation and memory usage.
On Accelerating Diffusion-Based Sampling Processes via Improved Integration Approximation
Guoqiang Zhang (University of Exeter), W. Bastiaan Kleijn (Victoria University of Wellington)
GenerationData SynthesisOptimizationComputational EfficiencyDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This paper studies the acceleration of ODE sampling for diffusion models through improved Integral Approximation (IIA), optimizing the step size coefficients of existing samplers (DDIM, DPM-Solver++, EDM) to achieve more accurate integral approximations.
On Adversarial Training without Perturbing all Examples
Max Losch (Max Planck Institute for Informatics), Bernt Schiele (CISPA Helmholtz Center for Information Security)
ClassificationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes Subset Adversarial Training (SAT), which generates adversarial samples only on a subset A of the training set, while the remaining subset B is not attacked, exploring the robustness transfer effects of this approach across different categories, samples, and downstream tasks.
On Bias-Variance Alignment in Deep Models
Lin Chen (Google Research), Sanjiv Kumar (Google Research)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies the bias-variance characteristics of deep learning model ensembles and finds that the squared bias of correctly classified samples aligns with the variance at the sample level.
On Differentially Private Federated Linear Contextual Bandits
Xingyu Zhou (Wayne State University), Sayak Ray Chowdhury (Microsoft Research)
Federated LearningSafty and PrivacyTabular
🎯 What it does: This paper addresses the Federated Linear Contextual Bandit problem in cross-datacenter settings, proposing a general algorithmic framework that satisfies differential privacy (DP), and systematically evaluates the convergence and communication costs under two privacy models: silo-level LDP and shuffle-DP.
On Diffusion Modeling for Anomaly Detection
Victor Livernoche (McGill University), Siamak Ravanbakhsh (McGill University)
Anomaly DetectionDiffusion modelImageTextTabularBenchmark
🎯 What it does: Utilizing the time posterior distribution of the diffusion model to detect anomalies, the Diffusion Time Estimation (DTE) method is derived by simplifying DDPM;
On Double Descent in Reinforcement Learning with LSTD and Random Features
David Brellmann (ENSTA Paris), Goran Frehse (ENSTA Paris)
Reinforcement LearningSequential
🎯 What it does: Analyzes the theoretical performance of regularized LSTD (TD learning) based on random features in the dual asymptotic limit where the ratio of parameter dimensions to the number of visited states (i.e., model complexity) approaches infinity, revealing the double descent phenomenon;
On Error Propagation of Diffusion Models
Yangming Li (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This study investigates the error propagation problem in diffusion models (DM), proposing a theoretical framework and designing regularization methods to reduce cumulative errors and improve generation quality.
On gauge freedom, conservativity and intrinsic dimensionality estimation in diffusion models
Christian Horvat (University of Bern), Jean-Pascal Pfister (University of Bern)
Diffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper introduces the theory of 'conformal freedom' and proves that conservative fields are neither a necessary nor sufficient condition for achieving exact sampling and density estimation in diffusion models. Exactness can only be guaranteed when the error term satisfies specific partial differential equations. Furthermore, it is demonstrated that conservative fields are beneficial for inferring local information of data manifolds (such as intrinsic dimension).
On Harmonizing Implicit Subpopulations
Feng Hong (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
ClassificationObject DetectionData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study addresses the issue of implicit subgroup imbalance and proposes the Scatter and Harmonize (SHE) method, which automatically discovers hidden subgroups through optimal data partitioning based on information theory and achieves balanced predictions for subgroups during training.
On Penalty Methods for Nonconvex Bilevel Optimization and First-Order Stochastic Approximation
Jeongyeol Kwon, Robert D Nowak
Optimization
🎯 What it does: A first-order stochastic approximation algorithm for non-convex bilevel optimization based on penalty methods is proposed, along with the gradient formula and approximation error analysis.
On Representation Complexity of Model-based and Model-free Reinforcement Learning
Hanlin Zhu (University of California Berkeley), Stuart Russell (University of California Berkeley)
Reinforcement LearningSequential
🎯 What it does: This study investigates the representation complexity of different functions (transition, reward, Q-function) in model-based and model-free reinforcement learning, proving that in certain classes of MDPs, transitions and rewards can be implemented using constant-depth polynomial-size circuits, while the optimal Q-function requires exponential-size circuits. Experiments in the MuJoCo environment further validate that the Q-function is more difficult to approximate with small networks compared to transitions and rewards.
On Stationary Point Convergence of PPO-Clip
Ruinan Jin (Chinese University of Hong Kong), Baoxiang Wang (Chinese University of Hong Kong)
OptimizationReinforcement Learning
🎯 What it does: A rigorous theoretical analysis of the convergence of the PPO-Clip algorithm is conducted, proving that under certain conditions of smoothness, bounded rewards, and learning rates satisfying the Robbins-Monro condition, PPO-Clip can converge to a stationary point of the value function, and provides the convergence rate of the gradient norm;
On the Analysis of GAN-based Image-to-Image Translation with Gaussian Noise Injection
Chaohua Shi (Xidian University), Xinbo Gao (Chongqing University of Posts and Telecommunications)
Image TranslationGenerationGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes injecting Gaussian noise during the training phase of a GAN-based image-to-image translation model and elucidates its mechanism for enhancing robustness through theoretical analysis.
On the Effect of Batch Size in Byzantine-Robust Distributed Learning
Yi-Rui Yang (Nanjing University), Wu-Jun Li (Nanjing University)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: The study investigates the impact of batch size on model accuracy and convergence in Byzantine Robust Distributed Learning (BRDL) with a fixed total gradient computation, and proposes a new algorithm called ByzSGDnm.
On the Expressivity of Objective-Specification Formalisms in Reinforcement Learning
Rohan Subramani (Columbia University), Joar Max Viktor Skalse (University of Oxford)
Reinforcement Learning
🎯 What it does: A systematic expressiveness comparison of 17 forms of objective specifications in reinforcement learning (such as Markov rewards, linear temporal logic, reward machines, multi-objective RL, etc.) was conducted, constructing a Hasse diagram and revealing the expressive limitations of each form.
On the Fairness ROAD: Robust Optimization for Adversarial Debiasing
Vincent Grari (Stanford University), Marcin Detyniecki (Polish Academy of Science)
OptimizationAdversarial AttackGenerative Adversarial NetworkTabular
🎯 What it does: The ROAD framework is proposed, combining distributionally robust optimization and adversarial learning to achieve local fairness and high accuracy on unknown subpopulations.
On the Foundations of Shortcut Learning
Katherine Hermann, Michael Curtis Mozer
ClassificationData SynthesisExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This study investigates how deep learning models prioritize features with higher availability over core features with stronger predictability during the training process, and evaluates the impact of availability and predictability on shortcut learning by constructing a controllable generative framework system.
On the Generalization and Approximation Capacities of Neural Controlled Differential Equations
Linus Bleistein (Inria Paris), Agathe Guilloux (Inria Paris)
Recurrent Neural NetworkTime SeriesFinance RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The paper provides the first rigorous statistical theoretical analysis of the generalization and approximation capabilities of Neural Controlled Differential Equations (NCDE) in learning irregular time series. It presents generalization bounds dependent on the smoothness of the time series and decomposes the total risk into sampling bias, approximation bias, and estimation bias, further providing an explicit impact of the sampling interval on the error.
On the generalization capacity of neural networks during generic multimodal reasoning
Takuya Ito (IBM Research), Murray Campbell (IBM Research)
Recurrent Neural NetworkTransformerMultimodalityBenchmark
🎯 What it does: A configurable multimodal question-answering benchmark gCOG was constructed, and the performance of various benchmark neural networks was evaluated on three types of OOD generalization (interference, system combination, productive combination).
On the Hardness of Constrained Cooperative Multi-Agent Reinforcement Learning
Ziyi Chen (University of Maryland), Heng Huang (University of Maryland)
OptimizationReinforcement Learning
🎯 What it does: This paper analyzes and addresses the issue of strong duality failure in constrained cooperative multi-agent reinforcement learning (C-MARL), proposing a decoupled primal-dual algorithm and comparing the convergence of two types of algorithms.
On the hardness of learning under symmetries
Bobak Kiani, Melanie Weber (Harvard University)
Convolutional Neural NetworkGraph Neural NetworkGraph
🎯 What it does: This paper proves that even with the inclusion of known symmetries (equivariance/invariance) in neural networks, there still exists exponential learning difficulty under common optimization methods such as gradient descent, through the analysis of the statistical query (CSQ) model.
On the Hardness of Online Nonconvex Optimization with Single Oracle Feedback
Ziwei Guan (Ohio State University), Yingbin Liang (Ohio State University)
Optimization
🎯 What it does: In online non-convex optimization with single operator feedback (exact gradient, stochastic gradient, or function value) obtained at once, a local regret evaluation of the original (non-smooth) objective function is proposed, along with corresponding lower bounds and optimal algorithms.
On the Humanity of Conversational AI: Evaluating the Psychological Portrayal of LLMs
Jen-tse Huang (Chinese University of Hong Kong), Michael Lyu
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the PsychoBench framework to evaluate five LLMs (text-davinci-003, ChatGPT, GPT-4, LLaMA-2-7B, LLaMA-2-13B) on 13 clinical psychological scales (personality, relationships, motivation, emotions).
On the Joint Interaction of Models, Data, and Features
Yiding Jiang (Carnegie Mellon University), J Zico Kolter
ClassificationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: By constructing interaction tensors and a new feature learning framework, a quantitative analysis of feature distribution and model prediction behavior in deep learning is conducted, explaining the Generalization Disagreement Equality (GDE) phenomenon.
On the Learnability of Watermarks for Language Models
Chenchen Gu (Stanford University), Tatsunori Hashimoto (Stanford University)
Knowledge DistillationAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper studies the learnability of watermarking in language models, proposing that through knowledge distillation, models can naturally generate watermarked text based solely on weights, thus validating the feasibility of learning watermarks.
On the Limitations of Temperature Scaling for Distributions with Overlaps
Muthu Chidambaram (Duke University), Rong Ge (Duke University)
ClassificationData SynthesisConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper studies the calibration limitations of temperature scaling when there is overlap in class distributions and proposes that using Mixup (including the extended d-Mixup) during the training phase can significantly improve the model's calibration performance.
On the Markov Property of Neural Algorithmic Reasoning: Analyses and Methods
Montgomery Bohde (Texas A&M University), Shuiwang Ji (Texas A&M University)
Graph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes ForgetNet and its improved version G-ForgetNet, aimed at eliminating the dependence on historical embeddings in neural algorithm reasoning models, ensuring consistency with the Markov property of algorithm reasoning; by introducing a gating mechanism and regularization, G-ForgetNet achieves more stable gradients in the early training phase and gradually converges to the Markov mode;
On the Over-Memorization During Natural, Robust and Catastrophic Overfitting
Runqi Lin (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: By studying the memory effect of deep networks, a unified concept of 'over-memorization' is proposed, and a Distraction Over-Memorization (DOM) framework is designed to prevent the model from over-memorizing high-confidence training samples, thereby alleviating overfitting in three scenarios: natural training, robust training, and catastrophic overfitting.
On the Parameterization of Second-Order Optimization Effective towards the Infinite Width
Satoki Ishikawa (Tokyo Institute of Technology), Ryo Karakida (Artificial Intelligence Research Center AIST)
OptimizationConvolutional Neural NetworkRecurrent Neural NetworkImageText
🎯 What it does: This study investigates how to parameterize second-order optimization (K-FAC and Shampoo) for stable feature learning in the context of infinitely wide networks.
On the Posterior Distribution in Denoising: Application to Uncertainty Quantification
Hila Manor (Technion - Israel Institute of Technology), Tomer Michaeli (Technion - Israel Institute of Technology)
RestorationSegmentationDiffusion modelImage
🎯 What it does: This paper derives a recursive relationship between high-order posterior central moments and high-order posterior mean derivatives, and utilizes a pre-trained denoiser to achieve training-free posterior uncertainty visualization and marginal distribution estimation.
On the Power of the Weisfeiler-Leman Test for Graph Motif Parameters
Matthias Lanzinger (Vienna University of Technology), Pablo Barcelo
Graph Neural NetworkGraph
🎯 What it does: Accurately characterize the WL dimension of graph primitive parameters and provide WL dimension determination algorithms for subgraph counting, induced subgraph counting, and k-graph block counting.
On the Provable Advantage of Unsupervised Pretraining
Jiawei Ge (Princeton University), Chi Jin (Princeton University)
OptimizationRepresentation LearningContrastive Learning
🎯 What it does: This paper proposes a general theoretical framework for rigorously analyzing the statistical advantages of unsupervised pre-training for downstream tasks, and presents a two-stage algorithm based on maximum likelihood pre-training and empirical risk minimization.
On the Reliability of Watermarks for Large Language Models
John Kirchenbauer (University of Maryland), Tom Goldstein (University of Maryland)
Large Language ModelText
🎯 What it does: This paper studies and evaluates the reliability of watermarks in large language models, exploring their performance when text is rewritten by humans, machine rewriting, or embedded in large texts;
On the Role of Discrete Tokenization in Visual Representation Learning
Tianqi Du (Peking University), Yisen Wang (Peking University)
Representation LearningTransformerContrastive LearningImage
🎯 What it does: This paper studies the role of discrete tokenization in self-supervised masked image modeling and proposes a clustering-based discrete tokenization method called ClusterMIM.
On the Role of General Function Approximation in Offline Reinforcement Learning
Chenjie Mao (Shanghai Artificial Intelligence Laboratory), Xuelong Li (China Telecom)
Reinforcement Learning
🎯 What it does: This paper studies the problem of universal function approximation in offline reinforcement learning, clarifying the distinction between realizability and completeness assumptions. It demonstrates the necessity of the completeness assumption through the construction of information-theoretic lower bounds, further proving that even if Q* realizability is satisfied under partially covered data, robust policy improvement cannot be achieved.
On the Scalability and Memory Efficiency of Semidefinite Programs for Lipschitz Constant Estimation of Neural Networks
Zi Wang (University of Wisconsin Madison), Somesh Jha (University of Illinois Urbana Champaign)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a method that transforms the semi-definite programming (SDP) estimation of the network Lipschitz constant into an eigenvalue optimization problem (EP-LipSDP) and implements a GPU-friendly first-order subgradient solver called LipDiff. It also introduces techniques such as Lanczos approximation, sparse matrix multiplication, and analytical initialization, enabling efficient computation of Lipschitz upper bounds on large networks (e.g., AlexNet on ImageNet).
On the Stability of Expressive Positional Encodings for Graphs
Yinan Huang (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A stable and expressive position information encoding method based on graph Laplacian eigenvectors, called SPE, is proposed to address the non-uniqueness and instability issues of traditional encodings.
On the Stability of Iterative Retraining of Generative Models on their own Data
Quentin Bertrand (Université de Montréal), Gauthier Gidel (Université de Montréal)
GenerationData SynthesisDiffusion modelFlow-based ModelImage
🎯 What it does: The study investigates the stability of iterative retraining of generative models on a mixed dataset (real data + data generated by the model itself).
On the Variance of Neural Network Training with respect to Test Sets and Distributions
Keller Jordan (Independent researcher)
ClassificationOptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: This study quantifies the performance variance of neural network training under different random seeds, test set partitions, and distribution shifts, and proposes an independent error assumption along with a variance estimation formula based on class calibration.
On the Vulnerability of Adversarially Trained Models Against Two-faced Attacks
Shengjie Zhou (Chongqing University), Lei Feng (Nanyang Technological University)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates a novel attack on adversarially trained models during the model validation phase—dual-sided attack, demonstrating that it can lead to misleadingly high adversarial robustness during validation.
On Trajectory Augmentations for Off-Policy Evaluation
Ge Gao (North Carolina State University), Min Chi (North Carolina State University)
Robotic IntelligenceReinforcement Learning from Human FeedbackAuto EncoderTabularTime SeriesSequentialElectronic Health Records
🎯 What it does: The OAT method is proposed to address the issue of insufficient offline trajectory coverage in human-involved tasks for OPE. By mining potential sub-trajectories and utilizing an improved VAE-MDP to generate new sub-trajectories, it enhances trajectory data and improves the evaluation accuracy of OPE by merging them with the original trajectories.
On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes
Rishabh Agarwal (Mila), Olivier Bachem (Google DeepMind)
Knowledge DistillationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A general knowledge distillation framework GKD is proposed, which utilizes the output sequences generated by the student (on-policy) and allows for the flexible selection of divergences such as KL or JSD to address the training-inference distribution mismatch problem in autoregressive model distillation.
One For All: Towards Training One Graph Model For All Classification Tasks
Hao Liu (Washington University in St. Louis), Muhan Zhang (Peking University)
ClassificationMeta LearningGraph Neural NetworkLarge Language ModelPrompt EngineeringGraph
🎯 What it does: A unified graph model called One for All (OFA) has been developed, which can be used for various graph tasks such as node, edge, and graph classification, and supports cross-domain, few-shot, and zero-shot learning.
One Forward is Enough for Neural Network Training via Likelihood Ratio Method
Jinyang Jiang (Peking University), Yijie Peng (Peking University)
Domain AdaptationOptimizationConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkSpiking Neural NetworkImageText
🎯 What it does: A unified likelihood ratio estimation method (ULR) is proposed, allowing gradient estimation to be completed with a single forward pass, thus eliminating the need for backpropagation.
One Step of Gradient Descent is Provably the Optimal In-Context Learner with One Layer of Linear Self-Attention
Arvind V. Mahankali, Tengyu Ma
OptimizationTransformerTabular
🎯 What it does: This paper theoretically proves that a single-layer linear self-attention Transformer, when trained on a synthetic noise linear regression task, has a global optimal solution corresponding to a prediction function that is equivalent to performing one gradient descent on the regression data (or preprocessed gradient descent when the covariance is non-unit).
One-hot Generalized Linear Model for Switching Brain State Discovery
Chengrui Li (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
Time Series
🎯 What it does: A one-hot encoded hidden Markov model generalized linear model (one-hot HMM-GLM) is proposed, which estimates the functional connectivity that varies over time across multiple states by decomposing the weight matrix into a discrete adjacency matrix and a positive intensity matrix, and sharing a Gumbel-Softmax prior for the adjacency matrix.
One-shot Active Learning Based on Lewis Weight Sampling for Multiple Deep Models
Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics), Ying-Peng Tang (Nanyang Technological University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A one-time active learning method is proposed, which can select the most valuable unlabeled samples for multiple deep models simultaneously without retraining the models;
One-shot Empirical Privacy Estimation for Federated Learning
Galen Andrew (Google), Vinith Menon Suriyakumar
Federated LearningSafty and PrivacyRecurrent Neural NetworkTextSequential
🎯 What it does: A one-time method is proposed to estimate the model's differential privacy loss ε in real-time during federated learning training.
Online Continual Learning for Interactive Instruction Following Agents
Byeonghwi Kim (Yonsei University), Jonghyun Choi (Seoul National University)
Knowledge DistillationRobotic IntelligenceReinforcement LearningTextMultimodality
🎯 What it does: This paper proposes a robot for interactive instruction following that can continuously learn new behaviors and new environments after deployment, addressing the limitations of traditional prior data learning scenarios.
Online GNN Evaluation Under Test-time Graph Distribution Shifts
Xin Zheng (Monash University), Shirui Pan (Griffith University)
Graph Neural NetworkGraph
🎯 What it does: In an online deployment environment, an online GNN evaluation framework is proposed to estimate the generalization error of a trained GNN on unlabeled, distribution-drift test graphs.
Online Information Acquisition: Hiring Multiple Agents
Federico Cacciamani (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper studies how to design incentive-compatible mechanisms for agents in multi-agent information acquisition scenarios, which can solve optimal mechanisms offline and learn approximately optimal mechanisms online, ultimately achieving no-regret learning with polynomial time complexity.
Online Stabilization of Spiking Neural Networks
Yaoyu Zhu (Peking University), Zhaofei Yu (Peking University)
Spiking Neural NetworkImage
🎯 What it does: Proposes Online Spike Re-normalization (OSR) and Online Threshold Stabilizer (OTS), achieving memory efficiency and biological interpretability in SNN online training without using future information.
Only Pay for What Is Uncertain: Variance-Adaptive Thompson Sampling
Aadirupa Saha (Apple), Branislav Kveton (Amazon)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: A Variational Adaptive Thompson Sampling (VarTS) algorithm suitable for Gaussian rewards and unknown heteroscedasticity is proposed, along with a prior-dependent upper bound on the Bayes expected regret loss.
Open the Black Box: Step-based Policy Updates for Temporally-Correlated Episodic Reinforcement Learning
Ge Li (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes Temporally-Correlated Episodic RL (TCE), a reinforcement learning framework that combines temporal correlation with periodic exploration, balancing the trajectory smoothness of ERL with the data efficiency of SRL.
Open-ended VQA benchmarking of Vision-Language models by exploiting Classification datasets and their semantic hierarchy
Simon Ging (University of Freiburg), Thomas Brox (University of Freiburg)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a method to convert traditional classification datasets into an open visual question answering (oVQA) benchmark, and follows up with questions at the semantic hierarchy level to eliminate answer ambiguity.
OpenChat: Advancing Open-source Language Models with Mixed-Quality Data
Guan Wang (Tsinghua University), Yang Liu (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the OpenChat framework, using Conditioned-RLFT for RL-free training on mixed quality data to enhance the performance of open-source LLMs.
OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views
Francis Engelmann (ETH Zurich), Federico Tombari (Google)
SegmentationVision Language ModelNeural Radiance FieldImage
🎯 What it does: Developed the OpenNeRF method, which utilizes neural radiance fields (NeRF) and pixel-level CLIP features to achieve open-set 3D semantic segmentation. It can generate continuous 3D representations from RGB images (with optional depth) with known camera poses and perform zero-shot segmentation for any text or image queries.
OpenTab: Advancing Large Language Models as Open-domain Table Reasoners
Kezhi Kong (University of Maryland), George Karypis (Amazon Web Services)
RetrievalTransformerLarge Language ModelTabularRetrieval-Augmented Generation
🎯 What it does: A framework for open-domain table reasoning called OPENTAB is proposed, which generates and executes SQL using LLM, achieving a complete link from retrieval to answering.
OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text
Keiran Paster (University of Toronto), Jimmy Ba (University of Toronto)
Text
🎯 What it does: The OpenWebMath dataset has been proposed and made public, containing 14.7 billion high-quality mathematical web texts, including LaTeX formulas and other technical content.
Optimal criterion for feature learning of two-layer linear neural network in high dimensional interpolation regime
Keita Suzuki (University of Tokyo and Preferred Networks), Taiji Suzuki (University of Tokyo and RIKEN AIP)
OptimizationRepresentation LearningTabular
🎯 What it does: This paper theoretically analyzes the statistical properties of feature learning in two-layer linear neural networks with multiple outputs in a high-dimensional interpolation environment and proposes a new criterion for achieving feature learning in high-dimensional settings.
OPTIMAL ROBUST MEMORIZATION WITH RELU NEURAL NETWORKS
Lijia Yu (Institute of Software Chinese Academy of Sciences Academy of Mathematics and Systems Science Chinese Academy of Sciences), Lijun Zhang (Institute of Software Chinese Academy of Sciences University of Chinese Academy of Sciences)
OptimizationTabular
🎯 What it does: This study investigates the robust memorization problem of neural networks on finite classification datasets, providing the complexity, necessary conditions, and theoretical methods for explicitly constructing optimal robust networks.
Optimal Sample Complexity for Average Reward Markov Decision Processes
Shengbo Wang (Stanford University), Peter Glynn (Stanford University)
OptimizationReinforcement Learning
🎯 What it does: This paper addresses the open problem of sample complexity in policy learning for maximizing the long-term average reward of uniformly ergodic Markov Decision Processes (MDPs) under the assumption of generative models.
Optimal Sample Complexity of Contrastive Learning
Noga Alon (Princeton University), Grigory Yaroslavtsev (George Mason University)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper studies the sample complexity of contrastive learning under the PAC learning framework, providing optimal upper and lower bounds for arbitrary distances, ℓp (integer p), and tree metrics.
Optimal Sketching for Residual Error Estimation for Matrix and Vector Norms
Yi Li (Nanyang Technological University), David Woodruff
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper studies the estimation of residual errors (low-rank approximation residuals and ℓp residuals) of matrices and vectors through linear sketches, and provides corresponding space/time complexity analysis.