π― What it does: This study investigates strategies for pruning the source dataset in transfer learning to enhance pre-training efficiency and downstream performance.
CodeObject TrackingDomain AdaptationRecurrent Neural NetworkSupervised Fine-TuningTime Series
π― What it does: This study investigates the data distribution drift caused by displacement of wearable flexible sensors at different wearing positions and proposes an unsupervised adaptive motion tracking network.
Self-Chained Image-Language Model for Video Localization and Question Answering
Shoubin Yu (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
CodeRecognitionRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoText
π― What it does: This paper proposes a Self-Chained Video Localization and Answering (SeViLA) framework based on a single image-language model BLIP-2, which first uses a Localizer to locate key frames with language perception, and then employs an Answerer to perform video question answering or event prediction on these frames, refining the Localizer through pseudo-labels generated from the answers.
π― What it does: This paper proposes a self-weighted multi-view contrastive learning framework SEM, which utilizes reconstruction regularization to alleviate the representation degradation problem in multi-view contrastive learning.
Semantic segmentation of sparse irregular point clouds for leaf/wood discrimination
Yuchen BAI, Florence Forbes (University of Grenoble Alpes)
CodeClassificationSegmentationPoint Cloud
π― What it does: A SOUL network based on PointNet++ is proposed for semantic segmentation of leaf and wood materials from sparse and heterogeneous point clouds generated by drone LiDAR.
π― What it does: This paper proposes a Semi-Implicit Denoising Diffusion Model (SIDDM), which decomposes the denoising distribution into implicit edge matching and explicit conditional matching (AFD), achieving large step sampling while maintaining high-quality samples.
π― What it does: A new semi-supervised contrastive learning framework CLSS is proposed, which utilizes the feature similarity matrix of unlabeled samples to recover ordinal relationships through spectral sequence algorithms, and applies it to contrastive learning and predictive supervision.
π― What it does: The SeqMatch method is proposed, which divides synthetic data into multiple subsets and optimizes them sequentially to address the issue of insufficient high-order feature distillation caused by excessive compression of low-level features.
π― What it does: A honeypot module is proposed, which inserts a small classifier at the lower layers of a pre-trained language model (PLM) to specifically absorb and cover backdoor information during the fine-tuning process, allowing the main network to only learn the original task.
π― What it does: This paper proposes SGFormer, a simplified Transformer model that utilizes a single-layer global linear attention combined with GCN to achieve efficient learning of large-scale graph node representations.
π― What it does: A zero-shot non-rigid shape matching method SNK is proposed, achieving shape matching through unsupervised feature mapping and decoder reconstruction directly on a single pair of shapes.
π― What it does: A Shared Adversarial Unlearning (SAU) method is proposed to eliminate backdoor attacks on deep neural networks by generating and removing shared adversarial samples.
Sharp Bounds for Generalized Causal Sensitivity Analysis
Dennis Frauen (LMU Munich), Stefan Feuerriegel (LMU Munich)
CodeFlow-based ModelTabularTime Series
π― What it does: This paper proposes a generalized marginal sensitivity model (GMSM) and derives sharp boundaries under various causal effects (average treatment effect, mediation effect, path effect, and distribution effect), providing scalable algorithms to estimate these boundaries from observational data.
π― What it does: This paper studies the impact of Sharpness-Aware Minimization (SAM) on the feature rank of deep networks, finding that SAM can significantly reduce the feature rank across different models (ResNet, ViT, MLP-Mixer, etc.) and different tasks (classification, contrastive learning);
ShiftAddViT: Mixture of Multiplication Primitives Towards Efficient Vision Transformer
Haoran You (Georgia Institute of Technology), Yingyan Celine Lin (Georgia Institute of Technology)
CodeClassificationComputational EfficiencyTransformerMixture of ExpertsImage
π― What it does: Reparameterize the pre-trained Vision Transformer by replacing multiplication operations with shifts and additions to construct the ShiftAddViT model.
SHOT: Suppressing the Hessian along the Optimization Trajectory for Gradient-Based Meta-Learning
JunHoo Lee (Seoul National University), Nojun Kwak (Seoul National University)
CodeOptimizationMeta LearningImage
π― What it does: This paper proposes an explicit suppression of the Hessian matrix on the optimization trajectory in gradient-based meta-learning, achieved by minimizing the distance between the target model and the reference model.
π― What it does: A framework for multimodal domain generalization, SimMMDG, is proposed by splitting each modality feature into shared and exclusive parts, and enhancing the model's generalization ability to unknown domains through supervised contrastive learning, distance constraints, and cross-modal translation.
π― What it does: This paper proposes SimMTM, a self-supervised pre-training framework for time series based on multiple masked sequence aggregation, aimed at improving the performance of prediction and classification tasks.
π― What it does: This paper proposes GraphACL, a framework for asymmetric contrastive learning that does not rely on graph data augmentation. It learns node representations by utilizing the one-hop neighbor context and two-hop homophily, making it applicable to both homophilic and heterophilic graphs.
Single-Call Stochastic Extragradient Methods for Structured Non-monotone Variational Inequalities: Improved Analysis under Weaker Conditions
Sayantan Choudhury (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
CodeOptimization
π― What it does: This paper proposes a new convergence analysis framework for the Single-Call Stochastic Extragradient methods (such as SPEG and SOG), providing convergence proofs for two types of structured non-monotone variational inequalities (quasi-strongly monotone and weak Minty) under weak assumptions (without the need for bounded variance or growth conditions), and unifying the consideration of arbitrary sampling (including uniform, importance sampling, and arbitrary mini-batch).
π― What it does: This paper designs a lightweight contrastive learning framework SACL-XD, which can train small networks from scratch and achieve high accuracy.
π― What it does: Designed the SmooSeg method, which utilizes smoothness priors and energy minimization to achieve unsupervised semantic segmentation.
SmoothHess: ReLU Network Feature Interactions via Stein's Lemma
Max Torop (Northeastern University), Jennifer Dy (Northeastern University)
CodeOptimizationAdversarial AttackConvolutional Neural NetworkGaussian SplattingImageTime SeriesBiomedical Data
π― What it does: Using Gaussian convolution on the output of ReLU networks to obtain a smooth function, we estimate its Hessian to capture the second-order interactions of features.
π― What it does: A text-to-image diffusion model has been implemented on mobile devices, completing in under 2 seconds using 8 denoising steps, with image quality comparable to Stable Diffusion v1.5.
π― What it does: Proposes the SOC framework to achieve video-level multimodal understanding, enhancing RVOS through semantic-assisted object clustering.
π― What it does: In scenarios where model parameters are inaccessible, a test-time data adaptation method called SODA based on zero-order optimization is proposed.
Softmax Output Approximation for Activation Memory-Efficient Training of Attention-based Networks
Changhyeon Lee (Ulsan National Institute of Science and Technology), Seulki Lee (Ulsan National Institute of Science and Technology)
CodeTransformerText
π― What it does: This paper proposes a method to approximate the softmax output during the training process of attention networks such as Transformer, reducing the activation memory usage.
Solving a Class of Non-Convex Minimax Optimization in Federated Learning
Xidong Wu (University of Pittsburgh), Heng Huang (University of Maryland)
CodeOptimizationFederated LearningImage
π― What it does: This paper studies non-convex-convex, non-convex-strongly convex, and non-convex-PL (Polyak-Lojasiewicz) minimax optimization problems in a federated learning environment, and proposes two improved algorithms, FedSGDA+ and FedSGDA-M.
π― What it does: This paper proposes a method for inverse physical solving that combines score matching with differentiable physical simulation, capable of inferring the initial state from the final state of the system and sampling the posterior distribution.
π― What it does: Using pre-trained Latent Diffusion Models to solve linear inverse problems (such as image inpainting, denoising, super-resolution, etc.)
SoTTA: Robust Test-Time Adaptation on Noisy Data Streams
Taesik Gong (Nokia Bell Labs), Sung-Ju Lee (KAIST)
CodeDomain AdaptationImage
π― What it does: This paper proposes a testing-time adaptive method for noisy samples (SoTTA), aimed at enhancing the model's robustness to noisy samples in real-world testing streams.
π― What it does: Using audio captured by a head-mounted microphone and human joint posture information, a multimodal network was constructed that can generate a three-dimensional sound field covering the human body, enabling spatial audio rendering at any spatial location.
π― What it does: This paper addresses the problem of unsupervised domain adaptation and proposes a framework based on spectral alignment, which captures cross-domain transferability while enhancing intra-domain discriminability.
π― What it does: A single-round participant fusion evaluation method called SPACE is proposed for efficiently assessing the contributions of each client in federated learning.
π― What it does: Designed and implemented a differentiable Sparse Modular Activation (SMA) mechanism, constructing the SeqBoat network, which utilizes SMA for dynamic sparse activation state space models (SSM) and gated attention units (GAU) to achieve efficient sequence modeling;
π― What it does: A sparse parameterization-based Epitomic dataset distillation framework SPEED is proposed, capable of synthesizing information-rich synthetic data in extremely low storage space.
π― What it does: This paper proposes and utilizes critical band masking experiments to compare the object recognition performance of humans and various deep networks under the influence of frequency domain noise, assessing their spatial frequency channel characteristics.
Spatially Resolved Gene Expression Prediction from Histology Images via Bi-modal Contrastive Learning
Ronald Xie (University of Toronto), Gary Bader
CodeRepresentation LearningData-Centric LearningConvolutional Neural NetworkContrastive LearningImageBiomedical Data
π― What it does: A dual-modal contrastive learning-based embedding framework called BLEEP is proposed, which can predict the spatial resolution of gene expression from H&E tissue slice images.
Sehoon Kim (University of California), Kurt Keutzer (University of California)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The Big Little Decoder (BiLD) framework is proposed, which generates text collaboratively using a small model (autoregressive) and a large model (non-autoregressive), significantly reducing inference latency through fallback and rollback strategies without altering the training process.
π― What it does: Proposes Spike-Driven Transformer, which integrates the spike-driven principle of SNN with Transformer, allowing the network to achieve attention and MLP computation solely through sparse addition.
π― What it does: This paper studies the phenomenon of spontaneous symmetry breaking in the generation process of diffusion models and utilizes this phenomenon to improve fast samplers.
SPQR: Controlling Q-ensemble Independence with Spiked Random Model for Reinforcement Learning
Dohyeok Lee (Seoul National University), Jungwoo Lee (Seoul National University)
CodeReinforcement LearningSequential
π― What it does: A Q-ensemble independence regularization method SPQR based on random matrix theory is proposed, which uses the KL divergence between the eigenvalue distribution and the Wigner semicircle distribution to reduce the correlation of the Q-ensemble and suppress estimation bias.
Spuriosity Didnβt Kill the Classifier: Using Invariant Predictions to Harness Spurious Features
Cian Eastwood (University of Edinburgh), Bernhard SchΓΆlkopf (Max Planck Institute for Intelligent Systems)
CodeDomain AdaptationImage
π― What it does: The Stable Feature Boosting (SFB) algorithm is proposed, which utilizes stable feature predictions to adjust and leverage unstable (artifact) features in an unsupervised manner to enhance domain generalization performance.
Squared Neural Families: A New Class of Tractable Density Models
Russell Tsuchida (Data61 CSIRO), Dino Sejdinovic (University of Adelaide)
CodeGenerationOptimizationFlow-based ModelTabular
π― What it does: A class of interpretable probability density models called Squared Neural Family (SNEFY) is proposed, which constructs distributions by normalizing the squared 2-norm of a single hidden layer neural network with respect to a base measure.
π― What it does: A Stabilized Neural Differential Equation (SNDE) is proposed, which enforces the solution trajectory to remain on any explicit constraint manifold by adding a penalty term based on the constrained Jacobian to the neural ODE.
π― What it does: This paper studies the instability of Stable Diffusion under slight perturbations of text prompts and proposes an automatic attack method called ATM based on Gumbel-Softmax, which can generate perturbations similar to the original prompts but lead to generation failures.
π― What it does: This paper presents a complete feature generation pipeline that treats the signature barcode of multiparameter persistent homology as a signature measure, and provides two stable vectorization methods (convolutional images and sliced Wasserstein kernels) to achieve efficient vectorization of multiparameter persistent homology.
π― What it does: A star-shaped denoising diffusion probabilistic model (SS-DDPM) is proposed, which can implement diffusion models on any exponential family distribution through a non-Markovian forward process that relies only on the marginal distribution.
π― What it does: An auxiliary self-supervised task is proposed that uses Fourier transform prediction in the frequency domain for infinite step state sequences to learn high-quality representations.
State-Action Similarity-Based Representations for Off-Policy Evaluation
Brahma S Pavse, Josiah P. Hanna (University of Wisconsin - Madison)
CodeReinforcement LearningTabular
π― What it does: A state-action similarity metric ROPE based on OPE is proposed, which learns an encoder from offline data and improves data efficiency in FQE.
CodeExplainability and InterpretabilityReinforcement LearningSequential
π― What it does: The StateMask explanation method is proposed, which learns a state masking network to randomize actions at non-critical moments without significantly affecting the final reward of the target DRL agent, thereby identifying and quantifying which states are most important for the final reward.
Statistical Knowledge Assessment for Large Language Models
Qingxiu Dong (Peking University), Lei Li (Carnegie Mellon University)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: A statistical method called KaRR is proposed to evaluate the reliability of factual answers generated by large language models under different prompts;
Statistically Valid Variable Importance Assessment through Conditional Permutations
Ahmad Chamma (Inria), Bertrand Thirion (Inria)
CodeBiomedical DataMagnetic Resonance Imaging
π― What it does: A conditional permutation importance method (CPI) is proposed and implemented to provide statistically valid variable importance assessments in the presence of correlated features.
StEik: Stabilizing the Optimization of Neural Signed Distance Functions and Finer Shape Representation
Huizong Yang (Georgia Institute of Technology), Anthony Yezzi (Georgia Institute of Technology)
CodeOptimizationPoint CloudMeshBenchmark
π― What it does: This study investigates the instability caused by the eikonal loss in neural SDF learning and proposes a new regularization method and a network structure using quadratic layers to stabilize optimization and improve detail reconstruction quality.
Congye Wang (Newcastle University), Chris J. Oates (Newcastle University)
CodeTabularBenchmark
π― What it does: The Stein Ξ -Importance Sampling method is proposed, which approximates the target distribution P by sampling from a Ξ -invariant Markov chain and post-processing through Stein discrepancy.
STEVE-1: A Generative Model for Text-to-Behavior in Minecraft
Shalev Lifshitz (University of Toronto), Sheila A. McIlraith
CodeGenerationRobotic IntelligenceTransformerPrompt EngineeringVision Language ModelVideoTextMultimodality
π― What it does: In the Minecraft game, STEVE-1 is constructed using a pre-trained VPT behavior model and the MineCLIP visual-text alignment model, enabling the generation of low-level control behaviors based on text or visual instructions.
π― What it does: A path integral stochastic optimal control method (PIPS) is proposed, which does not require prior specification of collective variables (CVs), for efficiently sampling the transition paths of molecular systems between two steady states.
Strategic Behavior in Two-sided Matching Markets with Prediction-enhanced Preference-formation
Stefania Ionescu (University of ZΓΌrich), Aniko Hannak
CodeTabularSequential
π― What it does: This paper proposes a framework for embedding predictive models into two-sided matching markets and defines the concept of 'adversarial interaction attack' for the first time, which refers to the returning party interacting with the matching object in a non-optimal way in the current round to disrupt future predictions and matching results. Through economic models and simplified analysis, it is demonstrated that such attacks can yield benefits for the returning party in certain situations, and it further illustrates the inequality and overall welfare decline caused by these attacks.
π― What it does: This paper proposes a structural pruning method for diffusion models called Diff-Pruning, which achieves the generation of lightweight models without retraining by retaining the core parameters of the pre-trained model and removing redundant weights.
Structure from Duplicates: Neural Inverse Graphics from a Pile of Objects
Tianhang Cheng (University of Illinois Urbana-Champaign), Shenlong Wang (University of Illinois Urbana-Champaign)
CodePose EstimationOptimizationSimultaneous Localization and MappingImage
π― What it does: This paper proposes inverse rendering in a single image containing repeated objects, utilizing multi-view information of the repeated objects to recover the geometry, material, and environmental lighting of the same object.
Structure Learning with Adaptive Random Neighborhood Informed MCMC
Xitong Liang (University College London), Jim Griffin (University College London)
CodeReinforcement LearningGraphBiomedical Data
π― What it does: A new MCMC sampler PARNI-DAG has been developed for complete Bayesian structure learning under observed data, sampling directly in the DAG space and achieving efficient mixing.
π― What it does: This paper proposes a Structure-Free Graph Compression (SFGC) method that compresses large-scale graph data into an unstructured set of nodes while implicitly encoding the graph structure information into node attributes.
Structured Neural Networks for Density Estimation and Causal Inference
Asic Q Chen (University of Toronto), Rahul G Krishnan (University of Toronto)
CodeFlow-based ModelTabular
π― What it does: A structured neural network (StrNN) is proposed, which implements explicit constraints on conditional independence through weight masking and integrates it into autoregressive flows (StrAF) and continuous flows (StrCNF) for density estimation and causal inference.
π― What it does: This paper designs and evaluates an S5 structured state space model that can reset hidden states through parallel scanning during training, aimed at addressing the intrinsic learning problem of long sequences in reinforcement learning.
Afra Amini (ETH Zurich), Ryan Cotterell (Johns Hopkins University)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper studies a gradient-based sampling method called Structural Voronoi Sampling (SVS) for sampling from discrete language models and achieving controlled text generation.
π― What it does: This paper studies a new type of label noise called Subclass Dominant Noise (SDN) and proposes a correction method called NoiseCluster for this noise.
π― What it does: A zero-training soft unified block pruning (SUBP) method is proposed for constructing 1ΓN sparse convolutional neural networks and achieving inference acceleration in a multi-threaded CPU environment.
π― What it does: Two variants based on reinforcement learning and graph neural networks (GRL-SVO(UP) and GRL-SVO(NUP)) have been designed and implemented to automatically suggest variable ordering in Cylindrical Algebraic Decomposition (CAD).
π― What it does: Designed and trained a Decision-Pretrained Transformer (DPT), which achieves reinforcement learning by allowing the Transformer to predict the optimal action during multi-task pre-training given a contextual dataset.
Meena Jagadeesan (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
CodeRecommendation SystemTabular
π― What it does: Construct and analyze a supply-side equilibrium model in recommendation systems, exploring the impact of multi-dimensional content vectors and production costs on content specialization and market competition.
Supported Value Regularization for Offline Reinforcement Learning
Yixiu Mao (Tsinghua University), Xiangyang Ji (Tsinghua University)
CodeReinforcement Learning
π― What it does: A Support Value Regularization (SVR) method is proposed, which penalizes the Q-values of out-of-distribution (OOD) actions that exceed the support of the behavior dataset in offline reinforcement learning, while retaining the standard Bellman update in the in-distribution (ID) region, thus addressing the issues of overestimation and extrapolation error.
Survival Permanental Processes for Survival Analysis with Time-Varying Covariates
Hideaki Kim (NTT Corporation)
CodeOptimizationTime SeriesBiomedical Data
π― What it does: This paper proposes a non-parametric Bayesian survival model based on permanent processesβSurvPPβto analyze survival data with time-varying covariates. It transforms infinite-dimensional optimization into finite-dimensional parameterization through a representation theorem, achieving MAP inference with linear time complexity.
π― What it does: The SutraNets method is proposed, which splits long sequences into low-frequency subsequences and uses subsequence autoregressive generation to address the issues of error accumulation and signal path in long sequence prediction.
π― What it does: An adaptive grid refinement method based on adaptive swarm reinforcement learning (ASMR) is proposed, treating each grid cell as an agent and utilizing graph neural networks to learn refinement strategies without the need for error estimation during inference.
π― What it does: A novel 4D Swin Transformer model, SwiFT, has been developed for end-to-end learning of brain spatiotemporal dynamics directly from high-dimensional fMRI volumes, as well as for biological and cognitive predictions.
Hyun Dong Lee (Stanford University), Scott Linderman
CodeVideoTime Series
π― What it does: The Switching Autoregressive Low-rank Tensor (SALT) model is proposed, which constrains the autoregressive tensor of ARHMM with low-rank tensor decomposition, combining the interpretability of ARHMM with the parameter efficiency of SLDS.
Switching Temporary Teachers for Semi-Supervised Semantic Segmentation
Jaemin Na (Ajou University), Wonjun Hwang (Ajou University)
CodeSegmentationDomain AdaptationImage
π― What it does: Proposes a Dual Teacher framework that uses dual temporary teachers to alternately train a single student model, alleviating the teacher-student coupling problem.
SyncTREE: Fast Timing Analysis for Integrated Circuit Design through a Physics-informed Tree-based Graph Neural Network
Yuting Hu (University at Buffalo), Jinjun Xiong (University at Buffalo)
CodeGraph Neural NetworkContrastive LearningGraph
π― What it does: A tree-structured graph neural network called SyncTREE is proposed for fast and accurate prediction of the timing (delay and slope) of integrated circuit interconnect RC trees.
π― What it does: This paper proposes a Synthetic Experience Replay (SYNTHER) method based on diffusion models for synthesizing and upsampling experience data in both offline and online reinforcement learning, aimed at enhancing sample efficiency and learning effectiveness.
π― What it does: Proposed the T2T framework: during the training phase, a discrete diffusion model is used to learn the high-quality solution distribution for each instance, while in the testing phase, instance-level gradient search is performed by incorporating the target gradient into the diffusion denoising process to improve the initial solution.
π― What it does: This study explores whether placing Batch Normalization after the activation function (Swap) when using bounded activation functions can improve model performance. It finds that this approach produces asymmetric saturation and high sparsity, allowing bounded activation functions to approximate ReLU, significantly increasing accuracy.
π― What it does: Two random feature methods are proposed to approximate the Tanimoto kernel (TMM) in molecular fingerprints and its new extended kernel TDP. Theoretical upper bounds on the error are provided and evaluated on real molecular data.
TART: A plug-and-play Transformer module for task-agnostic reasoning
Kush Bhatia (Stanford University), Christopher Re
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningImageTextAudio
π― What it does: This study investigates the performance gap between large language models (LLMs) in context learning and task-specific fine-tuning, and proposes a general reasoning module named TART, which enhances the model's reasoning ability by combining a Transformer trained on synthetic logistic regression tasks with the embeddings of any pre-trained model.
CodeClassificationRecognitionTransformerVision Language ModelImageMultimodality
π― What it does: The research improved the task arithmetic editing method of pre-trained models, enabling the model to achieve multi-task learning and task forgetting through addition and subtraction of task vectors.
π― What it does: This paper proposes a task-aware distributed source coding framework called NDPCA, which can dynamically adapt to any available bandwidth with a single model in multi-sensor networks, while maintaining high performance for downstream tasks such as denoising, robotic arm grasping, and satellite target detection after compression.
π― What it does: A task-aware world model learning framework TEMPO based on dual-layer optimization is proposed, which uses a meta-weight network to perform task-aware weighting of training samples;
Temperature Balancing, Layer-wise Weight Analysis, and Neural Network Training
Yefan Zhou (Dartmouth), Yaoqing Yang (Nanjing University)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: This paper proposes and implements TempBalance, a hierarchical learning rate scheduling method based on the Heavy-Tail Self-Regularization theory, which enhances training effectiveness by dynamically adjusting the 'temperature' of each layer.
TempME: Towards the Explainability of Temporal Graph Neural Networks via Motif Discovery
Jialin Chen (Yale University), Zhitao Ying
CodeExplainability and InterpretabilityGraph Neural NetworkGraphTime Series
π― What it does: This paper proposes TempME, an interpretable framework for temporal graph neural networks based on the information bottleneck, which utilizes temporal motifs to generate compact and consistent explanation subgraphs.
Temporal Causal Mediation through a Point Process: Direct and Indirect Effects of Healthcare Interventions
ΓaΔlar HΔ±zlΔ± (Aalto University), Pekka Marttinen (Aalto University)
CodeTime SeriesBiomedical DataElectronic Health Records
π― What it does: A dynamic causal mediation analysis framework based on point process is proposed to separate the direct and indirect effects of medical interventions on continuous time outcomes.
π― What it does: A multi-stage continual learning framework called Temporal Continual Learning (TCL) has been designed and implemented, introducing a Prior Compensation Factor (PCF) to alleviate knowledge forgetting, aimed at human motion prediction.
CodeDrug DiscoveryGraph Neural NetworkTabularPhysics Related
π― What it does: A Cartesian second-order tensor-based O(3)-equivariant message passing network, TensorNet, is proposed for learning molecular potential energy, forces, and other vector/tensor properties.
Test-Time Amendment with a Coarse Classifier for Fine-Grained Classification
Kanishk Jain (International Institute of Information Technology Hyderabad), Vineet Gandhi (University of TΓΌbingen)
CodeClassificationImage
π― What it does: A Hierarchical Ensembles (HiE) post-correction method is proposed: fine-grained and coarse-grained classifiers are trained separately, and during inference, the fine-grained predictions are multiplied by the corresponding parent class probabilities and normalized, thereby reducing error severity and improving Top-1 accuracy.
Test-Time Distribution Normalization for Contrastively Learned Visual-language Models
Yifei Zhou (University of California), Ser-Nam Lim (University of Central Florida)
CodeClassificationRetrievalVision Language ModelContrastive LearningImageText
π― What it does: A technique using Distribution Mean Normalization (DN) during the inference phase is proposed and evaluated to align with the InfoNCE objective of contrastive learning models (such as CLIP) during training, improving the performance of traditional dot product similarity.
π― What it does: For the task of video object segmentation (VOS) based on matching, a scheme for adaptive fine-tuning of the model during inference is proposed and validated, known as Test-time Training (TTT).