π― What it does: A gray-box Bayesian optimization method (RCGP) is proposed, which achieves low-budget and efficient optimization of hyperparameters for reinforcement learning algorithms by embedding the generalized logistic function estimation of the reward curve into the Gaussian process feature space.
π― What it does: This paper proposes a Greedy Actor-Critic algorithm based on Conditional Cross-Entropy (Conditional CEM), which enhances the policy of the original MDP by taking the top percentile of actions at each state and performing maximum likelihood updates on the actor.
GReTo: Remedying dynamic graph topology-task discordance via target homophily
Zhengyang Zhou (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeGraph Neural NetworkGraphTime Series
π― What it does: A GNN for dynamic graph regression, called GReTo, is proposed, which can correct the topology-task inconsistency issue through target homogeneity.
π― What it does: This paper proposes the Gromov-Wasserstein Autoencoder (GWAE), which achieves unsupervised representation learning by minimizing the GW distance.
Grounding Graph Network Simulators using Physical Sensor Observations
Jonas LinkerhΓ€gner (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
CodeGraph Neural NetworkPoint CloudMesh
π― What it does: A new graph network simulator (GGNS) is proposed, which integrates point cloud observation information into grid simulation, achieving more accurate physical simulation under incomplete initial conditions.
π― What it does: The BOON method is proposed, which makes structural modifications on the neural operator kernel to explicitly satisfy the Dirichlet, Neumann, and periodic boundary conditions of PDEs, thereby improving the accuracy of the solutions.
π― What it does: By using the spherical latent representations obtained from contrastive learning as latent variables for the energy model, we jointly train the energy model and contrastive learning to improve generation quality and training stability.
Guiding Safe Exploration with Weakest Preconditions
Greg Anderson (University of Texas at Austin), Isil Dillig (University of Texas at Austin)
CodeSafty and PrivacyReinforcement LearningSequential
π― What it does: A neural-symbolic method called SPICE is implemented, which constructs interpretable safety shields using weak preconditions to ensure action safety during training.
Harnessing Mixed Offline Reinforcement Learning Datasets via Trajectory Weighting
Zhang-Wei Hong (Massachusetts Institute of Technology), Romain Laroche
CodeReinforcement LearningTabular
π― What it does: This paper proposes a method of trajectory weighting to improve policy learning in offline reinforcement learning datasets with mixed reward distributions.
π― What it does: This paper proposes integrating the Hebbian rule with a new gradient-based plasticity rule into recurrent neural networks (RNNs), enabling them to continuously update weights in an unsupervised environment through self-generated targets, thereby enhancing memory and rapid learning capabilities.
π― What it does: Proposed and implemented a multi-layer SoftHebb algorithm that performs deep feature extraction using feedback-free, unsupervised Hebbian learning, achieving high-accuracy classification on standard visual benchmarks.
π― What it does: Proposes HiCLIP, which incorporates hierarchy-aware attention into the visual and language branches of CLIP to automatically learn the hierarchical structure of images and text;
Hidden Markov Transformer for Simultaneous Machine Translation
Shaolei Zhang (University of Chinese Academy of Sciences), Yang Feng (Institute of Computing Technology, Chinese Academy of Sciences)
CodeTransformerText
π― What it does: This paper proposes the Hidden Markov Transformer (HMT), which unifies the problem of 'when to start translating' in machine translation with the translation itself, and learns the optimal translation timing through an HMM structure.
π― What it does: This study focuses on few-shot knowledge graph completion and proposes a hierarchical relationship learning framework called HiRe, which utilizes three layers of relational information (entity layer, triple layer, and context layer) to jointly learn meta-relation representations.
π― What it does: This paper proposes the Hidden Time Markov Decision Process (HiT-MDP) and proves its equivalence to the traditional SMDP options framework; based on this, it constructs the Maximum Entropy Option Policy Gradient (MOPG) algorithm, using transformers to learn option embeddings and optimize policies within this framework.
π― What it does: This paper proposes HiViT, a simplified and efficient hierarchical visual Transformer that combines the simplicity of ViT with the hierarchical features of Swin, making it particularly suitable for occluded image modeling.
HotProtein: A Novel Framework for Protein Thermostability Prediction and Editing
Tianlong Chen (University of Texas at Austin), Adam Klivans (University of Texas at Austin)
CodeProtein Structure PredictionGraph Neural NetworkContrastive LearningBiomedical Data
π― What it does: A large-scale protein thermal stability dataset called HotProtein is proposed, along with a complete learning framework for thermal stability prediction and editing.
π― What it does: This study investigates semi-supervised learning assisted by pseudo-labelers, demonstrating that semi-supervised pre-training combined with linear probing can achieve nearly zero test error on a simplified data model and a two-layer convolutional neural network, while pure supervised learning can only achieve constant-level test error.
How Informative is the Approximation Error from Tensor Decomposition for Neural Network Compression?
Jetze Schuurmans (Delft University of Technology), Julian Kooij (Delft University of Technology)
CodeCompressionConvolutional Neural NetworkImage
π― What it does: A systematic experimental study was conducted on the correlation between the approximation error using tensor decomposition in neural network compression and the performance of the compressed model.
π― What it does: This paper evaluates the contribution of data augmentation to model performance by quantifying the 'equivalent sample size' of augmented views, and explores the role of augmentation under different data scales, distribution shifts, model widths, and in comparison with explicit invariant networks. It also investigates the mechanism of augmentation as a source of randomness in the training process and demonstrates that it can lead to a flatter loss landscape.
π― What it does: A representation learning framework named CIDER is proposed, which utilizes two types of losses (diversity and compactness) in the hyperspherical embedding space to enhance the performance of out-of-distribution (OOD) sample detection.
π― What it does: A Motion Diffusion Model (MDM) based on Transformer is proposed, utilizing an unsupervised diffusion generative model to achieve various conditional and unconditional human motion generation and editing.
Human-Guided Fair Classification for Natural Language Processing
Florian E. Dorner (Max Planck Institute for Intelligent Systems), Martin Vechev (ETH Zurich)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A complete process for individual fairness constraints based on text pair generation is proposed to guide the fair training of text classifiers.
π― What it does: A theoretical framework was constructed to estimate the unobserved 'oracle' accuracy using the consistency among multiple observed human annotators, providing a lower bound on the accuracy of machine learning models, thus enabling a formal determination of whether the model surpasses average human performance.
Hungry Hungry Hippos: Towards Language Modeling with State Space Models
Daniel Y Fu, Christopher Re (University at Buffalo)
CodeTransformerLarge Language ModelText
π― What it does: This paper studies the application of state space models (SSM) in language modeling and proposes a new H3 layer and FLASHCONV acceleration algorithm to enhance model performance and hardware efficiency.
π― What it does: A hybrid reinforcement learning framework, Hybrid RL, is proposed, and a Hybrid Q-Learning (Hy-Q) algorithm based on fitted Q iteration is designed, with its efficiency verified both theoretically and experimentally.
HypeR: Multitask Hyper-Prompted Training Enables Large-Scale Retrieval Generalization
ZeFeng Cai, Daxin Jiang (Microsoft Corporation)
CodeRetrievalTransformerPrompt EngineeringText
π― What it does: HYPER is proposedβa multi-task hyper-prompt training mechanism that utilizes a Query Conditional Prompt Synthesizer (QPS) and Contrastive Prompt Regularization (CPR) to enable neural retrievers to uniformly handle queries across different tasks and domains, achieving cross-domain knowledge transfer through multi-task training.
π― What it does: The HYSP model is proposed, utilizing self-supervised learning and self-paced learning in hyperbolic space for representation learning of skeletal actions.
π― What it does: This study investigates the identifiability of multimodal contrastive learning, proving that under different generative mechanisms and modality-specific latent variables, contrastive learning can achieve block-level identifiability for shared content factors.
ILA-DA: Improving Transferability of Intermediate Level Attack with Data Augmentation
Chiu Wai Yan (Hong Kong University of Science and Technology), Dit-Yan Yeung (Hong Kong University of Science and Technology)
CodeAdversarial AttackImage
π― What it does: Improved the transferability of Intermediate Layer Attacks (ILA) by introducing three data augmentation techniques during the ILA process to generate more aggressive adversarial samples.
π― What it does: Proposes Fast Weight Painter (FPA), which applies the outer product learning rule of rapid weight generation to image generation, gradually stacking rank-one updates to obtain the final image.
ImaginaryNet: Learning Object Detectors without Real Images and Annotations
Minheng Ni (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
CodeObject DetectionLarge Language ModelDiffusion modelContrastive LearningImage
π― What it does: This paper proposes a learning paradigm for training object detectors entirely based on synthetic imagesβImaginary-Supervised Object Detection (ISOD), which does not rely on any real images or manual annotations.
Imbalanced Semi-supervised Learning with Bias Adaptive Classifier
Renzhen Wang (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
CodeClassificationSupervised Fine-TuningImage
π― What it does: A pseudo-label learning framework L2AC is proposed to adapt to class imbalance, using a bias-adaptive classifier to suppress training bias in pseudo-label semi-supervised learning and improve the performance of minority classes.
π― What it does: The PIG framework is proposed, which utilizes graph planning to generate subgoals and distills the subgoal policies into the target policy through self-imitation learning, significantly improving the sample efficiency of goal-conditioned RL.
π― What it does: This paper introduces diffusion models into behavior cloning, constructing a conditional distribution from observation to action, and proposes various architectures (Basic MLP, MLP Sieve, Transformer) and reliable sampling schemes (Diffusion-X, Diffusion-KDE) for mimicking human behavior in continuous or mixed action spaces.
Jiangyuan Li (Texas A&M University), Raymond K. W. Wong (Texas A&M University)
CodeOptimizationTabular
π― What it does: This paper constructs a new 'diagonal grouped linear neural network' reparameterization and proves that gradient descent will naturally converge to a solution with a group sparse structure without explicit regularization.
π― What it does: The paper discusses how to learn the optimal strategy from 'unachievable' experts in situations where experts have more information than learners, and proposes the ELF Distillation method.
π― What it does: In deep regression tasks, the differences in feature learning between regression and classification are analyzed from the perspective of mutual information. It is found that using MSE for regression tends to lead to low feature entropy. An 'ordinal entropy' regularization term is proposed to maintain both ordinal relationships and high entropy in the feature space, and this regularization is applied as a plugin to various regression models.
Improving the imputation of missing data with Markov Blanket discovery
Yang Liu (Queen Mary University of London), Anthony Constantinou
CodeTabular
π― What it does: A feature selection method based on Markov Blanket (MBFS) is proposed, which is embedded in MissForest to form a new imputation algorithm (MBMF) aimed at improving the filling of missing data.
InCoder: A Generative Model for Code Infilling and Synthesis
Daniel Fried (Carnegie Mellon University), Mike Lewis (Facebook AI Research)
CodeGenerationAI Code AssistantTransformerLarge Language ModelText
π― What it does: This paper presents InCoder, a unified generative code model capable of both program synthesis (left-to-right generation) and code completion (masking and filling in), trained on a large-scale code corpus.
π― What it does: This paper proposes and evaluates a non-targeted poisoning attack for unsupervised contrastive learningβContrastive Poisoning (CP), which can significantly reduce the performance of linear probes in CL models and can also attack supervised models based on CL learning.
Individual Privacy Accounting with Gaussian Differential Privacy
Antti Koskela (Nokia Bell Labs University of Helsinki), Antti Honkela (University of Helsinki)
CodeSafty and PrivacyGaussian SplattingBiomedical DataElectronic Health Records
π― What it does: This paper proposes an individual privacy measurement framework based on Gaussian Differential Privacy (GDP), which can progressively track and limit the privacy loss of each data point in a fully adaptive algorithm sequence.
π― What it does: This paper explores and quantifies the 'inequality phenomenon' that occurs during lβ adversarial training, where the model relies on a small number of extremely high-weight features for predictions; it also validates the robustness flaws caused by this phenomenon through noise and occlusion attacks.
π― What it does: Proposes and verifies that the mutual information (MI) between input and hidden representations in Dropout neural networks with continuous noise is finite, enabling information plane (IP) analysis;
π― What it does: In the context of federated learning, the recovery of instance-level labels for batch samples is achieved by analyzing the shared gradients.
π― What it does: This paper proposes an interactive portrait harmonization framework that allows users to select any reference area in the background to guide the matching of tones and brightness between the foreground and background.
Interpretable Geometric Deep Learning via Learnable Randomness Injection
Siqi Miao (Georgia Institute of Technology), Pan Li (Purdue University)
CodeExplainability and InterpretabilityGraph Neural NetworkGaussian SplattingPoint Cloud
π― What it does: This paper proposes a learnable randomness injection (LRI) mechanism that allows geometric deep learning models to directly generate interpretable point importance scores while maintaining high predictive performance.
π― What it does: A theoretical framework based on Layer Variational Analysis (LVA) is proposed to explain and implement transfer learning and domain adaptation in deep networks, and an optimal first-order weight update formula is provided through this framework; its effectiveness is validated on three types of tasks (1D time series regression, speech enhancement, and image deblurring).
Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization
Rajkumar Ramamurthy (Fraunhofer IAIS), Yejin Choi (Paul G. Allen School of Computer Science, University of Washington)
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: An open-source RL4LMs library, GRUE evaluation benchmark, and NLPO algorithm are proposed for aligning large language models (LLMs) with human preferences through reinforcement learning (RL);
π― What it does: This paper systematically evaluates the effectiveness of synthetic images generated by the current state-of-the-art text-to-image generation model (GLIDE) in image recognition tasks, covering three scenarios: zero-shot, few-shot learning, and large-scale pre-training, and proposes various strategies to enhance the quality and diversity of synthetic data.
Is the Performance of My Deep Network Too Good to Be True? A Direct Approach to Estimating the Bayes Error in Binary Classification
Takashi Ishida (University of Tokyo), Masashi Sugiyama (RIKEN)
CodeClassificationTransformerImage
π― What it does: A method for directly estimating Bayesian error in binary classification problems is proposed, aimed at evaluating classifier performance and detecting overfitting in the test set.
Iterative Circuit Repair Against Formal Specifications
Matthias Cosler (CISPA Helmholtz Center for Information Security), Bernd Finkbeiner (CISPA Helmholtz Center for Information Security)
CodeTransformerSequential
π― What it does: A deep learning method is proposed that utilizes Transformer to repair sequential circuits to meet given Linear Temporal Logic (LTL) specifications.
π― What it does: An Iterative Patch Selection (IPS) method is proposed to select the most discriminative patches in high-resolution images without using gradients, and aggregates the selected patches using a cross-attention Transformer, enabling training and inference of large images within limited GPU memory.
π― What it does: Proposes the RAVEn framework, which uses self-supervised cross-modal learning to jointly learn visual and auditory speech representations from raw video and audio, and fine-tunes on visual and auditory speech recognition tasks.
Alexander Korotin (Skolkovo Institute of Science and Technology), Evgeny Burnaev (Skolkovo Institute of Science and Technology)
CodeImage TranslationOptimizationImage
π― What it does: This paper proposes a neural optimal transport (NOT) method based on kernel weak quadratic cost to learn one-to-one or one-to-many stochastic transport plans.
KnowDA: All-in-One Knowledge Mixture Model for Data Augmentation in Low-Resource NLP
Yufei Wang (Macquarie University), Daxin Jiang (Microsoft Corporation)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Developed KnowDA, a Seq2Seq generation model based on multi-task pre-training (KoMT) for data augmentation in low-resource NLP tasks;
π― What it does: In traditional label propagation (LPA), prior information (such as weak labeler predictions) is incorporated to provide new error bounds, and a method is proposed to integrate multi-source noise information into the graph by adding 'dongle' nodes, ultimately generating more accurate pseudo-labels in weakly supervised scenarios.
Tuomas Oikarinen (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
CodeExplainability and InterpretabilityLarge Language ModelContrastive LearningImage
π― What it does: An automated concept bottleneck model framework that does not require concept labels is proposed, which can transform any network into an interpretable CBM;
Phillip Rust (University of Copenhagen), Desmond Elliott (University of Copenhagen)
CodeTransformerLarge Language ModelAuto EncoderText
π― What it does: PIXEL is proposed, a language model that renders text as images and uses Vision Transformer for masked autoencoding, bypassing the traditional vocabulary bottleneck.
Language models are multilingual chain-of-thought reasoners
Freda Shi (Google Research), Jason Wei (Google Research)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: This paper studies multilingual chain-of-thought reasoning, proposing and constructing the MGSM benchmark to evaluate the arithmetic reasoning capabilities of large language models in ten different languages.
Language Models are Realistic Tabular Data Generators
Vadim Borisov (University of Tuebingen), Gjergji Kasneci (Technical University of Munich)
CodeGenerationData SynthesisTransformerLarge Language ModelTabular
π― What it does: A method is proposed for text encoding of tabular data and generating high-quality synthetic tabular data using Transformer-based LLMs (such as GPT-2).
Language Models Can Teach Themselves to Program Better
Patrick Haluptzok (Microsoft Research), Adam Tauman Kalai (Microsoft Research)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Using a language model to self-generate programming problems and answers, verifying correctness through a Python interpreter, and then fine-tuning the model to significantly improve its performance on programming problems.
Large Language Models are Human-Level Prompt Engineers
Yongchao Zhou (University of Toronto), Jimmy Ba (University of Toronto)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Using large language models to automatically generate and select natural language prompts (Instructions), treating prompt engineering as a black-box optimization and program synthesis problem;
π― What it does: This paper proposes Latent Bottlenecked Attentive Neural Processes (LBANPs), a neural process model that maintains high performance while having query complexity independent of the number of context points.
π― What it does: A learning algorithm-independent data value assessment framework LAVA is proposed, utilizing category-based Wasserstein distance as a proxy for validating performance, and assigning values to individual samples through sensitivity analysis of OT distance.
Xinjie Zhang (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
CodeCompressionImage
π― What it does: A learning-driven distributed multi-view image compression framework LDMIC is proposed, which adopts an independent encoding and joint decoding mode, utilizing cross-view attention to achieve global mutual view information fusion.
Learnable Topological Features For Phylogenetic Inference via Graph Neural Networks
Cheng Zhang (Peking University)
CodeGraph Neural NetworkGraph
π― What it does: A learnable topological feature based on graph neural networks is proposed for the systematic representation of tree structures without domain-specific knowledge.
π― What it does: This paper proposes a learning index framework that can dynamically adjust the error upper bound Ξ΅. Through theoretical derivation, it connects Ξ΅ with local data features (mean Β΅, variance Ο), and automatically selects an appropriate Ξ΅ based on local distribution during learning in each linear segment, thereby enhancing the space-time trade-off.
π― What it does: This paper proposes the FETCH framework, which automates feature engineering based on reinforcement learning and constructs a data-driven feature search process by directly using the raw dataset as the MDP state.
π― What it does: A progress regression model is trained using expert demonstrations (only observations, no actions, no rewards) to extract monotonic progress information during the game process, and this progress is used as auxiliary rewards to drive agents to efficiently explore and complete tasks in sparse reward environments like NetHack.
π― What it does: This study investigates the cutting plane selection problem in Mixed Integer Linear Programming (MILP) and proposes a hierarchical sequence model based on reinforcement learning that can simultaneously learn which cutting planes to select, how many to select, and the order in which to add them.
π― What it does: Learn fair graph representations through an automated graph data augmentation method (Graphair) to reduce biases caused by sensitive attributes.
Learning Fast and Slow for Online Time Series Forecasting
Quang Pham (Institute for Infocomm Research Agency for Science Technology and Research), Steven Hoi
CodeRecurrent Neural NetworkTime Series
π― What it does: This study investigates fast and slow learning in online time series forecasting, proposing the FSNet framework to enhance the adaptability and memory capabilities of deep networks in data streams.
Learning Group Importance using the Differentiable Hypergeometric Distribution
Thomas M. Sutter (ETH Zurich), Julia E Vogt
CodeImage
π― What it does: A differentiable hypergeometric distribution is proposed for end-to-end learning of subgroup importance (such as cluster size and number of shared factors).
π― What it does: A multi-resolution molecular representation framework HMR based on molecular surface Laplace-Beltrami eigenfunctions is proposed, and a harmonic information propagation and functional mapping mechanism is designed based on this framework.
Learning Hierarchical Protein Representations via Complete 3D Graph Networks
Limei Wang (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeRepresentation LearningProtein Structure PredictionGraph Neural NetworkGraphBiomedical Data
π― What it does: Construct a 3D representation of proteins and design the ProNet hierarchical graph network to achieve protein representation learning at different granularities from amino acids, backbone to all-atom.
π― What it does: A parameter-free hyper label model (Hyper Label Model) is proposed, which can complete label aggregation in a single forward pass and is suitable for programmatic weak supervision scenarios.
π― What it does: This paper proposes a text-guided StyleGAN image editing method based on dictionary learning called Multi2One, which achieves multi-channel interactive editing direction learning and inference while maintaining real-time inference speed.
π― What it does: A learning-based iterative neural optimizer (LISO) is proposed and trained to find low bit error rate and visually natural steganographic images in image steganography.
π― What it does: An end-to-end regularized label encoding learning method (RLEL) is proposed for binary classification encoding in regression problems.
Learning Locality and Isotropy in Dialogue Modeling
Han Wu (City University of Hong Kong), Linqi Song (City University of Hong Kong)
CodeGenerationRetrievalTransformerLarge Language ModelContrastive LearningText
π― What it does: A dialogue representation calibration method called SimDRC is proposed, which optimizes the contextual representation of dialogue models using locality and isotropy constraints.
Learning Math Reasoning from Self-Sampled Correct and Partially-Correct Solutions
Ansong Ni (Yale University), Jianfeng Gao (Microsoft Research)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a method that allows pre-trained language models to self-sample multiple correct or partially correct solutions during training, using these solutions as multiple objectives for learning, thereby improving the quality of answers in multi-step mathematical reasoning tasks.
π― What it does: Train a multi-layer perceptron (MLP) without message passing on graphs through knowledge distillation, enabling it to capture graph structural information and be robust to feature noise.
Learning Object-Language Alignments for Open-Vocabulary Object Detection
Chuang Lin (Monash University), Jianfei Cai (Monash University)
CodeObject DetectionVision Language ModelContrastive LearningImageText
π― What it does: This paper proposes an open vocabulary object detection model VLDet, which directly learns the alignment between image regions and text words from image-text pair data without the need for manual annotation.
Learning on Large-scale Text-attributed Graphs via Variational Inference
Jianan Zhao (Mila - Quebec AI Institute), Jian Tang (Mila - Quebec AI Institute)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkLarge Language ModelTextGraph
π― What it does: The research focuses on learning node representations on Text Attribute Graphs (TAG) and proposes the GLEM framework, which achieves scalable integration of language models and graph neural networks through variational EM, addressing the computational and memory bottlenecks of traditional end-to-end training.
Learning Proximal Operators to Discover Multiple Optima
Lingxiao Li (Massachusetts Institute of Technology), Justin Solomon (Massachusetts Institute of Technology)
CodeObject DetectionOptimizationImagePoint Cloud
π― What it does: Learn and train proximal operators so that for any parameterized non-convex optimization problem, one can quickly obtain multiple local optimal solutions by iterating from a random initial point for only a few steps, and generalize to unseen problems.
Learning ReLU networks to high uniform accuracy is intractable
Julius Berner (University of Vienna), Felix Voigtlaender (Catholic University of EichstΓ€tt-Ingolstadt)
CodeRecurrent Neural Network
π― What it does: This paper quantifies the number of training samples required by any learning algorithm to guarantee high uniform accuracy on ReLU neural networks with a given architecture, proving that under general assumptions, the required sample size grows exponentially with the network depth and input dimension.
Learning to Compose Soft Prompts for Compositional Zero-Shot Learning
Nihal V. Nayak (Brown University), Stephen Bach
CodeRecognitionObject DetectionTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: By using learnable soft word embeddings for the attribute and object vocabulary of CLIP, we train composable prompts to enhance the performance of large-scale vision-language models in compositional zero-shot learning.
π― What it does: This study investigates the Cross Exchange (NCE) operation based on Graph Neural Networks to efficiently solve various min-max Vehicle Routing Problems (min-max VRP).
π― What it does: A spatially correlated noise model based on Markov processes is proposed, and a label correction method is designed based on this model.
Learning to Solve Constraint Satisfaction Problems with Recurrent Transformer
Zhun Yang (Arizona State University), Joohyung Lee (Samsung Research)
CodeTransformerImageText
π― What it does: The study uses a recursive Transformer to solve constraint satisfaction problems (CSP), particularly Sudoku and visual Sudoku, and implements sample-efficient and semi-supervised learning by directly embedding discrete constraints into the model using the STE method.
π― What it does: A fully self-supervised object localization and identity decoupling network called Loci is proposed, which learns the 'what' and 'where' of objects through slot coding, and achieves object permanence and interaction through predictive coding, gated recursion, and multi-head attention.
π― What it does: A neural-symbolic integration pipeline (NASR) is proposed, consisting of a neural solver, a mask predictor, and a symbolic solver, capable of enforcing hard constraints during inference.
π― What it does: Proposes the Auxiliary Activation Learning algorithm, which uses auxiliary activations to replace the input activations in forward propagation, thereby reducing the activation data that needs to be stored during training and significantly lowering memory usage.
Learning with Logical Constraints but without Shortcut Satisfaction
Zenan Li (Nanjing University), Jian L\"{u}
CodeClassificationOptimizationImage
π― What it does: This paper proposes a new framework that seamlessly integrates logical constraints into deep neural networks, avoiding the shortcut satisfaction problem encountered in traditional methods.