π― What it does: A framework is proposed that first aligns and then generates, utilizing a latent space of shape-image-text alignment for 3D shape generation.
π― What it does: A general exploration framework based on masked input modeling, MIMEx, is proposed, which utilizes sequence-level masked autoencoders to generate intrinsic rewards that encourage agents to actively explore in sparse reward visual control environments.
π― What it does: This paper proposes the MIMONets framework, which can simultaneously handle multiple sets of inputs in a single forward pass, significantly improving inference throughput.
Mind the spikes: Benign overfitting of kernels and neural networks in fixed dimension
Moritz Haas (University of TΓΌbingen), Ingo Steinwart (University of Stuttgart)
CodeTabular
π― What it does: The paper demonstrates that by designing a 'spiky-smooth' kernel and activation function with sharp high-frequency components, benign overfitting can still be achieved in fixed dimensions, and provides corresponding theoretical consistency and inconsistency analyses.
Zeyu Sun (University of Michigan), Alfred Hero (University of Michigan)
CodeClassificationOptimizationTabular
π― What it does: A minimum risk recalibration theory based on MSE decomposition is proposed, along with a unified risk upper bound and optimal bin number.
Mitigating Source Bias for Fairer Weak Supervision
Changho Shin (University of Wisconsin Madison), Frederic Sala (University of Wisconsin Madison)
CodeTabularBenchmark
π― What it does: This study investigates the fairness bias introduced by label functions in weakly supervised training and proposes a source bias mitigation method based on adversarial fairness.
Mitigating Test-Time Bias for Fair Image Retrieval
Fanjie Kong (Duke University), Ricardo Henao (Duke University)
CodeRetrievalTransformerSupervised Fine-TuningVision Language ModelImage
π― What it does: This paper studies gender/race bias in image retrieval under neutral text queries and proposes a post-processing method called PBM to mitigate bias during testing.
Mitigating the Effect of Incidental Correlations on Part-based Learning
Gaurav Bhatt (University of British Columbia), Vineeth N. Balasubramanian
CodeClassificationRecognitionExplainability and InterpretabilityKnowledge DistillationTransformerImage
π― What it does: This paper proposes DPViT, which learns decoupled components of foreground and background through a mixture-of-parts approach, alleviating the decline in interpretability and generalization caused by incidental correlations.
π― What it does: A decorrelation objective based on LogDet divergence is proposed, applied to Graph Collaborative Filtering (GCF), to alleviate popularity bias caused by dimensional collapse and improve the recommendation performance for unpopular items.
π― What it does: A method for unsupervised domain adaptation model selection based on mixed samples, called MixVal, is proposed, which can evaluate and select the best UDA model using only unlabeled target domain data during the inference phase.
π― What it does: MixFormerV2 is proposed, a vision object tracking framework completely based on Transformer, which removes convolutional heads and complex scoring modules;
MoCa: Measuring Human-Language Model Alignment on Causal and Moral Judgment Tasks
Allen Nie (Stanford University), Tobias Gerstenberg (Stanford University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: By integrating causal and moral judgment factors from 24 cognitive science papers, a structured annotated causal and moral judgment challenge set was constructed, and various large language models' alignment performance was evaluated on this set.
π― What it does: This paper proposes a modality-agnostic self-supervised learning framework called MetaMAE based on Masked Autoencoders (MAE), treating MAE as a meta-learning task to further enhance cross-modal representation learning effectiveness.
π― What it does: This paper proposes the VALOR method, which significantly improves model performance in weakly labeled audio-visual event parsing tasks by generating fine-grained, modality-aware pseudo-labels using large contrastive pre-trained models CLIP and CLAP, and extends this framework to weakly supervised audio-visual event localization tasks.
Model Shapley: Equitable Model Valuation with Black-box Access
Xinyi Xu (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeClassificationImageBiomedical Data
π― What it does: A fair model valuation method based on black-box access, called Model Shapley, is proposed, which abstracts model predictions using the Dirichlet distribution and measures model value through Shapley values.
π― What it does: This paper explores and verifies that model sparsification achieved through weight pruning can significantly enhance the performance of approximate machine unlearning (MU). It proposes two new paradigms: 'prune first, then unlearn' and sparse-aware unlearning, and applies them to backdoor defense and transfer learning.
Model Spider: Learning to Rank Pre-Trained Models Efficiently
Yi-Kai Zhang (Nanjing University), Han-Jia Ye (Nanjing University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelImageTextMultimodality
π― What it does: This paper proposes the MODEL SPIDER method, which achieves efficient model selection by mapping pre-trained models and tasks to vector spaces and learning the similarity between the two.
Model-Based Reparameterization Policy Gradient Methods: Theory and Practical Algorithms
Shenao Zhang (Northwestern University), Tuo Zhao (Georgia Tech)
CodeReinforcement LearningSequential
π― What it does: Analyzed and proved the convergence of the model-based reparameterization policy gradient method (RP-PGM) and the sources of gradient variance/bias, and proposed the Spectral Normalization technique to suppress gradient explosion and improve learning efficiency.
π― What it does: This paper proposes a retrieval framework called Model-Enhanced Vector Index (MEVI), which integrates autoregressive sequence generation models and dual-tower dense retrieval models to achieve high recall and low-latency retrieval for large-scale corpora.
Model-Free Active Exploration in Reinforcement Learning
Alessio Russo (KTH Royal Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
CodeReinforcement Learning
π― What it does: A model-free active exploration algorithm MF-BPI (and its deep learning version DBMF-BPI) is proposed, which dynamically allocates sampling resources by estimating the variance of the Q-function and value function, thereby quickly finding near-optimal policies in unknown MDPs.
Modelling Cellular Perturbations with the Sparse Additive Mechanism Shift Variational Autoencoder
Michael Bereket (insitro), Theofanis Karaletsos (insitro)
CodeGenerationRepresentation LearningDrug DiscoveryAuto EncoderBiomedical Data
π― What it does: This paper proposes a Sparse Additive Mechanism Shift Variational Autoencoder (SAMS-VAE) for learning the latent representations of cells under different interventions and predicting gene expression.
π― What it does: Proposes the Modulated Neural ODE (MoNODE) framework, which adds time-invariant modulation factors (static and dynamic modulators) to the traditional NODE, achieving separation of dynamic states and inherent factors;
π― What it does: This paper proposes a sampling framework based on 'pseudo Gibbs sampling' and 'moment matching', utilizing the noise EBM trained through Denoising Score Matching (DSM) to sample directly from its implicit clean model.
Monitor-Guided Decoding of Code LMs with Static Analysis of Repository Context
Lakshya Agrawal, Sriram Rajamani
CodeGenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Proposes Monitoring-Driven Decoding (MGD), which enhances code quality by applying static analysis from the IDE to constrain code in real-time during the generation process.
π― What it does: This paper proposes a unified normalized depth objective and 3D cubic depth supervision to address the differences in focal length and pitch angle between monocular 3D detection on the vehicle side and the infrastructure side.
Biao Jiang (Fudan University), Tao Chen (Fudan University)
CodeGenerationPose EstimationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVideoText
π― What it does: Treating human motion as a foreign language, a unified MotionGPT model is constructed to accomplish multiple tasks (text-driven motion generation, motion description, motion prediction, etc.).
Vaisakh Shaj (Karlsruhe Institute Of Technology), Gerhard Neumann (Karlsruhe Institute Of Technology)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackRecurrent Neural NetworkTransformerWorld ModelTime SeriesSequential
π― What it does: A multi-time scale world model (MTS3) is proposed, which can learn environmental dynamics at different time granularities and make long-horizon predictions.
Multi-Agent Learning with Heterogeneous Linear Contextual Bandits
Anh Do (Johns Hopkins University), Raman Arora (Johns Hopkins University)
CodeReinforcement Learning
π― What it does: This paper studies the heterogeneous multi-agent linear contextual bandit problem and proposes a distributed learning algorithm H-LINUCB based on UCB, providing theoretical upper and lower bounds when the heterogeneity Ξ΅ is known.
π― What it does: A lightweight two-headed network based on SE(3) equivariance is proposed, which simultaneously performs rigid body segmentation and motion estimation under unsupervised conditions.
π― What it does: A multi-modal inverse constraint reinforcement learning (MMICRL) algorithm is proposed to unsupervisedly identify different types of experts in mixed expert demonstration data and estimate the corresponding constraint functions, thereby achieving multi-constraint imitation learning.
Yifan Xu (Institute of Automation, Chinese Academy of Sciences), Changsheng Xu (Institute of Automation, Chinese Academy of Sciences)
CodeObject DetectionTransformerVision Language ModelImageTextMultimodality
π― What it does: This paper proposes a multimodal query object detection framework, MQ-Det, which achieves unsupervised open-set detection under the condition of providing only text descriptions and visual instances, and efficiently performs incremental tuning on baseline language query detection models (such as GLIP, GroundingDINO).
π― What it does: This paper proposes a multi-scale diffusion denoised smoothing scheme, which utilizes the same diffusion model for random smoothing at different noise scales and selects the most reliable scale through cascading, balancing high accuracy with provable robustness.
π― What it does: A method is proposed that can automatically search for the best graph neural network architecture in multi-task graph learning (MTGCβ―3), while also learning the collaborative relationships between tasks.
CodeExplainability and InterpretabilityRecurrent Neural NetworkTransformerMultimodalityTime SeriesBiomedical DataElectronic Health Records
π― What it does: We propose MultiModN, a sequentially composable multimodal multitask modular network that maintains predictive performance even when different modalities are missing.
π― What it does: Replace all multiplications in Transformer training with piecewise affine operations (PAM), achieving a completely multiplication-free training process.
Multitask Learning with No Regret: from Improved Confidence Bounds to Active Learning
Pier Giuseppe Sessa (ETH Zurich), Andreas Krause (ETH Zurich)
CodeDrug DiscoveryTabularBiomedical Data
π― What it does: This study proposes a new multi-task learning framework that provides confidence intervals for multi-task regression without prior knowledge and designs an adaptive regret-free algorithm to optimize the learning of multiple tasks.
MuSe-GNN: Learning Unified Gene Representation From Multimodal Biological Graph Data
Tianyu Liu (Yale University), Hongyu Zhao (Yale University)
CodeRepresentation LearningDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningMultimodalityBiomedical Data
π― What it does: The MuSe-GNN model is proposed, which integrates single-cell and spatial transcriptomics multimodal data in a gene-centric manner to generate unified gene embeddings.
π― What it does: This paper proposes NAR-Former V2, a universal neural network representation learning framework that integrates Transformer and GNN, capable of handling both cell structure models and complete deep networks.
π― What it does: A robust natural actor-critic (RNAC) algorithm based on double sampling (DS) and integral probability metric (IPM) uncertainty sets is proposed for achieving robust reinforcement learning using function approximation in large state spaces.
π― What it does: A systematic improvement of the evaluation of active learning methods was conducted, and a large-scale experimental benchmark was established.
π― What it does: This paper constructs a HiHack dataset with hierarchical labels to study and implement a neural strategy for NetHack based on hierarchical behavior cloning, a Transformer-LSTM structure, and RL fine-tuning, significantly improving the performance of the neural model.
π― What it does: A neural combinatorial optimization model with a light encoding-heavy decoding (LEHD) structure is proposed, achieving constructive solutions for large-scale TSP/CVRP problems through a data-efficient 'Learn to construct partial solution' training scheme and a Random Reconstruction (RRC) mechanism.
Neural Data Transformer 2: Multi-context Pretraining for Neural Spiking Activity
Joel Ye (University of Pittsburgh), Robert Gaunt (University of Pittsburgh)
CodeSpiking Neural NetworkTransformerPrompt EngineeringTime SeriesBiomedical Data
π― What it does: This paper presents NDT2, a multi-context pre-trained Transformer model designed for decoding neural spike activity in brain-machine interfaces.
CodeTabularBiomedical DataElectronic Health RecordsBenchmark
π― What it does: Proposed the Neural Frailty Machine (NFM), a survival regression framework that utilizes multiple random effects to extend CoxPH and combines neural networks.
π― What it does: Designed and implemented a Neural Function Transformer (NFT) based on attention mechanisms to handle weight space and applied it to tasks such as implicit space learning (INR2ARRAY), INR classification, editing, and CNN generalization prediction.
Kiarash Zahirnia (Simon Fraser University), Oliver Schulte (McGill University)
CodeGenerationData SynthesisSafty and PrivacyGraph Neural NetworkAuto EncoderGraph
π― What it does: We propose GenStatβa deep graph generation model that is based solely on graph statistics rather than a complete adjacency matrix, designed to generate realistic synthetic graphs under the premise of local differential privacy (LDP).
Neural Image Compression: Generalization, Robustness, and Spectral Biases
Kelsey Lieberman (Duke University), Bhavya Kailkhura (Lawrence Livermore National Laboratory)
CodeCompressionAuto EncoderImageBenchmark
π― What it does: This paper proposes a benchmark dataset (CLIC-C, Kodak-C) and a spectral inspection tool for evaluating neural image compression (NIC) in out-of-distribution (OOD) scenarios. It investigates the generalization and robustness performance of traditional codecs versus NIC models under different compression rates, noise types, and intensities through experimental and theoretical analysis.
Junlin Wu (Washington University in St. Louis), Yevgeniy Vorobeychik (Washington University in St. Louis)
CodeOptimizationReinforcement Learning from Human FeedbackTime Series
π― What it does: A neural Lyapunov control learning framework for discrete-time nonlinear systems (DITL) is proposed, which can derive provably stable control strategies.
π― What it does: This paper proposes a neural heuristic method based on indicator enhancement (NHDE) for multi-objective combinatorial optimization problems, significantly improving the diversity and convergence of the Pareto front.
π― What it does: A lightweight post-processing defense method called Neural Polarizer is proposed, which filters trigger features and retains normal features by inserting a learnable linear transformation layer into the model under backdoor attack, thereby purifying contaminated samples.
Matthew Wallingford (University of Washington), Ali Farhadi (University of Washington)
CodeRetrievalDomain AdaptationPrompt EngineeringVision Language ModelImageText
π― What it does: Proposes the Neural Priming method, which utilizes data retrieval from the pre-trained model itself and fine-tuning (or dynamic retrieval during inference) to improve adaptability to distribution shifts and downstream tasks, without the need for additional labeled data.
π― What it does: This paper proposes a Neural Relation Graph based on the relationships of data in the feature embedding space to unify the identification of label noise and anomalous/out-of-domain samples, providing a visualization tool for interactive data diagnosis.
Liulei Li (University of Technology Sydney), Yi Yang (Zhejiang University)
CodeRecognitionObject DetectionTransformerVision Language ModelImage
π― What it does: Proposes the LOGICHOI framework, which rewrites self-attention as triplet reasoning attention in the interactive decoder of the Transformer, allowing the model to autonomously combine people, actions, and objects during the decoding phase and predict interactions; simultaneously embedding feasibility constraints (affordances and proxemics) expressed in first-order logic into continuous space to guide the learning and reasoning of the Transformer.
NeuralGF: Unsupervised Point Normal Estimation by Learning Neural Gradient Function
Qing Li (Tsinghua University), Zhizhong Han (Wayne State University)
CodeOptimizationPoint Cloud
π― What it does: An unsupervised neural gradient function is proposed to directly estimate consistently oriented normal vectors from raw point clouds, achieving global surface fitting through multi-step point movement.
π― What it does: An end-to-end neural symbolic learning framework is proposed, capable of simultaneously training neural network perception and explicit logical constraints under weak supervision, and enabling interaction between the two through symbolic binding.
π― What it does: This paper proposes the NeuroGF, an implicit representation based on neural networks for fast and accurate querying of geodesic distances and shortest paths between arbitrary points; it also unifies the encoding of mesh geometry and geodesic information; further extended to a general framework that supports various inputs such as point clouds and meshes.
CodeGraph Neural NetworkGraphTime SeriesPhysics Related
π― What it does: A graph neural network (NC) based on Newton-Cotes numerical integration is proposed, which directly predicts the time evolution of dynamic systems through multi-step velocity estimation and integration formulas.
π― What it does: A training method is proposed that enhances the generalization and stability of GANs under limited data by incorporating adaptive multiplicative noise into the discriminator and using consistency regularization.
No Train No Gain: Revisiting Efficient Training Algorithms For Transformer-based Language Models
Jean Kaddour (University College London), Matt Kusner
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper evaluates three types of efficient training algorithms (dynamic architecture, batch selection, efficiency optimizers) under a fixed computational budget (reference system time) and explores their impact on the pre-training and downstream performance of BERT and T5.
No-Regret Online Prediction with Strategic Experts
Omid Sadeghi (University of Washington), Maryam Fazel (University of Washington)
CodeReinforcement LearningTabular
π― What it does: This paper studies the online binary prediction problem where m experts can be selected in each round, and designs an algorithm that can incentivize truthful reporting while achieving no-regret even when experts may cheat.
π― What it does: This paper proposes and implements Noether Embedding (NE), which directly learns and retrieves temporal patterns (TR) through the embedding of event samples, and defines two tasks: TR detection and TR querying, which are experimentally evaluated on ICEWS14, ICEWS18, and GDELT.
π― What it does: An unsupervised time series anomaly detection framework based on point reconstruction and sequence reconstruction (NPSR) is proposed, which captures both point anomalies and contextual anomalies by introducing nominality scores and induced anomaly scores.
Non-autoregressive Machine Translation with Probabilistic Context-free Grammar
Shangtong Gui (Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Institute of Computing Technology, Chinese Academy of Sciences)
π― What it does: This paper proposes the PCFG-NAT model, which embeds probabilistic context-free grammar (right-heavy PCFG) into a non-autoregressive Transformer, utilizing hierarchical structures to capture semantic dependencies of the target language, addressing multimodal issues and improving translation quality.
π― What it does: This paper proposes a hybrid framework based on Deep Functional Maps (DFM) and non-rigid shape registration, which can non-rigidly align the source mesh to the target point cloud under the condition of no correspondence supervision, achieving high-quality point cloud correspondences.
Nonparametric Boundary Geometry in Physics Informed Deep Learning
Scott Alexander Cameron (Oxford University), Stephen J. Roberts (Oxford University)
CodeTransformerMeshPhysics Related
π― What it does: Designed and trained a neural operator capable of receiving triangular mesh boundary geometry and outputting corresponding PDE solutions, addressing the issue of fixed geometric parameterization in traditional PINNs.
Nonparametric Identifiability of Causal Representations from Unknown Interventions
Julius von KΓΌgelgen (Max Planck Institute for Intelligent Systems), Bernhard SchΓΆlkopf (Max Planck Institute for Intelligent Systems)
CodeFlow-based ModelAuto Encoder
π― What it does: This paper studies the identifiability problem of non-parametric causal representation learning under multiple environments with unknown intervention targets, proving that potential causal variables and their causal graphs can be theoretically recovered solely through perfect single-node interventions.
π― What it does: A multi-learner non-parametric teaching framework MINT is proposed, which can provide teaching samples for multiple learners (each learner learning a scalar target function) at once, addressing the issue that single-learner teaching cannot be directly extended to multi-learner scenarios.
Normalization Layers Are All That Sharpness-Aware Minimization Needs
Maximilian Mueller, Matthias Hein (University of TΓΌbingen)
CodeOptimizationTransformerImage
π― What it does: This paper explores the application of Sharpness-Aware Minimization (SAM) by perturbing only the parameters of normalization layers (SAM-ON) to enhance model generalization performance.
π― What it does: A Progressive Active Learning (PAL) method is proposed, which actively selects valuable pseudo-ID and pseudo-OOD instances to simultaneously improve the performance of ID classifiers and OOD detectors.
NPCL: Neural Processes for Uncertainty-Aware Continual Learning
Saurav Jha (University of New South Wales Sydney), Lina Yao (CSIRO Data61)
CodeTransformerImage
π― What it does: Proposes Neural Processes for Continual Learning (NPCL), which maps continual learning tasks to a hierarchical latent variable model and combines experience replay to achieve knowledge sharing and uncertainty estimation between tasks.
NuTrea: Neural Tree Search for Context-guided Multi-hop KGQA
Hyeong Kyu Choi (University of Wisconsin Madison), Hyunwoo J. Kim (Korea University)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes a tree search-based graph neural network, NuTrea, for multi-hop knowledge graph question answering, which can simultaneously consider forward information and backward subtree context during the search path, addressing the limitations of traditional GNNs that only focus on past information.
π― What it does: The NVFi framework is proposed, which utilizes multi-view videos to simultaneously learn the geometry, appearance, and decoupled velocity fields of 3D scenes, achieving future frame extrapolation, unsupervised semantic decomposition, and motion transfer.
π― What it does: Latent Slot Diffusion (LSD) is proposed in object-centric learning, combining diffusion models with slot attention to achieve unsupervised segmentation, attribute prediction, combinatorial generation, and editing of complex scenes.
π― What it does: This paper proposes a framework that combines a recursive model based on ODE (GRU-ODE) with model-free reinforcement learning to address continuous control problems under partially observable Markov decision processes (POMDP).
Offline Multi-Agent Reinforcement Learning with Implicit Global-to-Local Value Regularization
Xiangsen Wang (Beijing Jiaotong University), Xianyuan Zhan (Tsinghua University)
CodeReinforcement Learning
π― What it does: The OMIGA algorithm is proposed, which utilizes implicit global-to-local value regularization to address the offline multi-agent reinforcement learning problem.
π― What it does: This paper proposes a technique that only requires a single calibration to improve the model score matching loss and likelihood lower bound for existing diffusion probability models (DPM).
On Class Distributions Induced by Nearest Neighbor Graphs for Node Classification of Tabular Data
Federico Errica (NEC Laboratories Europe)
CodeClassificationGraph Neural NetworkTabular
π― What it does: This paper studies the use of k-NN graphs for node classification in tabular data without original graph structure and constructs a theoretical framework to evaluate its impact on deep graph networks.
On Masked Pre-training and the Marginal Likelihood
Pablo Moreno-MuΓ±oz (Technical University of Denmark), SΓΈren Hauberg (Technical University of Denmark)
CodeTransformerAuto EncoderImageText
π― What it does: Through theoretical derivation and empirical validation, it is proven that Masked Pre-Training (MPT) essentially randomizes the maximization of the model's marginal likelihood (LML), revealing the equivalence between its cumulative loss and LML;
On skip connections and normalisation layers in deep optimisation
Lachlan Ewen MacDonald, Simon Lucey (Australian Institute for Machine Learning)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: This paper constructs a general multi-layer parameterized system (MPS) theoretical framework, systematically analyzing the effects of batch normalization, weight normalization, and residual skip connections on network loss curvature and regularization. Within this framework, it is proven that gradient descent can converge to a global optimum in networks containing these structures (even if the global optimum is located at infinity). Additionally, the authors reveal the causal mechanism by which residual connections enhance the Jacobian singular value distribution, thereby accelerating training through experiments on singular value distribution.
π― What it does: The theoretical analysis of the ability of Graph Neural Networks (GNN) to capture interactions between vertices is conducted, and based on this, a sparsification algorithm called WIS (Walk Index-based Sparsification) is proposed, aiming to maintain the interaction modeling capability of GNN when edges are deleted.
π― What it does: This paper studies the adversarial robustness in out-of-distribution (OOD) models, finding that existing OOD methods are vulnerable to adversarial attacks. It subsequently proposes two theory-based algorithms (AT and RDANN) to enhance OOD adversarial robustness.
On the choice of Perception Loss Function for Learned Video Compression
Sadaf Salehkalaibar (University of Toronto), Ashish J Khisti
CodeCompressionFlow-based ModelVideo
π― What it does: This study investigates the impact of two perceptual loss functions on reconstruction quality in causal, low-latency video compression, and proposes the universality of MMSE representation.
CodeExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
π― What it does: By training the network using optimal transport dual loss under the 1-Lipschitz constraint, this study investigates the impact of its gradient on interpretability and proposes that the Saliency Map of OTNN can serve as a high-quality explanation.
Manli Shu (University of Maryland), Tom Goldstein (University of Maryland)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes an attack method based on Automated Data Poisoning (AutoPoison), which utilizes an oracle LLM to inject malicious examples into instruction tuning data to alter the executable behavior of large language models.
π― What it does: Analyzed and proved the upper bound of generalization error in the training process of diffusion models, revealing the impact of early stopping strategies and mode displacement on generalization, and conducted numerical validation on synthetic and MNIST data.
On the impact of activation and normalization in obtaining isometric embeddings at initialization
Amir Joudaki (ETH Zurich), Francis Bach (INRIA ENS PSL Paris)
CodeClassificationOptimizationImage
π― What it does: This study investigates how normalization layers (such as layer normalization) and nonlinear activation functions affect the isometry of the Gram matrix during the initialization of deep neural networks and quantifies its convergence rate.
On the Importance of Exploration for Generalization in Reinforcement Learning
Yiding Jiang (Carnegie Mellon University), Roberta Raileanu (Meta AI Research)
CodeReinforcement LearningTabularBenchmark
π― What it does: The research explores the impact of exploration strategies on reinforcement learning generalization in Contextual MDPs (CMDP) and proposes an exploration method through Ensemble Distribution Exploration (EDE) to guide agents in exploring high-uncertainty states in the training environment.
π― What it does: This paper studies the relationship between the performance of different language directions in multilingual neural machine translation (MNMT) and the sampling ratio, revealing the phenomenon of Pareto front collapse, and proposes a double power law model to predict the performance of each direction; based on this model, it solves the optimal sampling ratio to improve overall translation quality while maintaining training costs.
On the Powerfulness of Textual Outlier Exposure for Visual OoD Detection
Sangha Park (Seoul National University), Sungroh Yoon (Seoul National University)
CodeAnomaly DetectionTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: The research proposes using text as proxy data for Out-of-Distribution (OoD) detection through 'Textual Outlier Exposure' and trains a linear classifier in the CLIP visual-language embedding space.
π― What it does: This study proves that the representation layer of pre-trained models constrains the adversarial robustness of downstream linear predictors, providing corresponding theoretical upper bounds and criteria.
One Fits All: Power General Time Series Analysis by Pretrained LM
Tian Zhou (Alibaba Inc.), Rong Jin (Alibaba Inc.)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTime Series
π― What it does: This paper proposes a general time series analysis frameworkβFrozen Pretrained Transformer (FPT). It transfers Transformers (such as GPT-2, BERT, BEiT) that have been pretrained on large-scale text or image data to time series tasks, requiring only the redesign of input embeddings, instance normalization, and output layers, and fine-tuning these layers while keeping other self-attention and feedforward layers frozen. The framework is evaluated on seven major tasks: classification, anomaly detection, imputation, short-term/long-term forecasting, and few-shot/zero-shot prediction.
π― What it does: This paper proposes a unified cross-architecture knowledge distillation framework, OFA-KD, which allows different structured teacher models (CNN, Transformer, MLP) to effectively distill into a student model.
One-Pass Distribution Sketch for Measuring Data Heterogeneity in Federated Learning
Zichang Liu (Rice University), Anshumali Shrivastava (ThirdAI Corporation)
CodeFederated LearningImageText
π― What it does: A one-pass distribution sketch is proposed to quantify data heterogeneity in federated learning, and this sketch is used for client selection and cold start model retrieval.
OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling
YiFan Zhang, Tieniu Tan (University of Chinese Academy of Sciences)
CodeReinforcement LearningTime Series
π― What it does: A dual-stream online model called OneNet is proposed, which uses two predictors for cross-time and cross-variable, and dynamically integrates their outputs through online convex programming (OCP) to achieve time series forecasting in the context of concept drift.
Online Adaptive Policy Selection in Time-Varying Systems: No-Regret via Contractive Perturbations
Yiheng Lin (California Institute of Technology), Adam Wierman (California Institute of Technology)
CodeOptimizationReinforcement Learning
π― What it does: This study investigates the problem of online adaptive strategy selection in dynamically changing systems, proposing a gradient-based adaptive strategy selection algorithm (GAPS) and establishing a general online optimization analysis framework.