ICLR 2024 Papers — Page 23
International Conference on Learning Representations · 2260 papers
ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation
Jiaming Liu (Peking University), Shanghang Zhang (Peking University)
ClassificationSegmentationDomain AdaptationTransformerImage
🎯 What it does: This paper proposes a Visual Domain Adapter (ViDA) and an Adaptive Knowledge Allocation strategy (HKA), which effectively alleviates error accumulation and catastrophic forgetting by inserting low-rank and high-rank adapters into pre-trained models to simultaneously capture domain-shared and domain-specific knowledge in the Continuous Test-Time Adaptation (CTTA) task.
Video Decomposition Prior: Editing Videos Layer by Layer
Gaurav Shrivastava (University of Maryland), Abhinav Shrivastava (University of Maryland)
RestorationSegmentationOptimizationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: A video decomposition prior (VDP) framework based on video layering is proposed, which uses U-Net to predict RGB layers and alpha layers from the RGB representation of input videos and optical flow, achieving video decomposition and editing.
Video Language Planning
Yilun Du (Google Deepmind), Jonathan Tompson
GenerationOptimizationRobotic IntelligenceTransformerReinforcement LearningVision Language ModelImageVideoText
🎯 What it does: This paper proposes a forward tree search planning method (VLP) that combines visual language models with text-to-video models, capable of generating long-term video plans from images and natural language goals, and transforming video plans into robot control through goal-conditioned policies or inverse dynamics models, achieving automation of complex long-term tasks.
Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation
Kimia Hamidieh (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
Representation LearningContrastive LearningImage
🎯 What it does: This paper studies the representation learning bias caused by spurious correlations in self-supervised learning and proposes a Late-layer Transformation-based View Generation (LATETVG) method for pruning regularization in the representation space to reduce the impact of spurious correlations.
ViLMA: A Zero-Shot Benchmark for Linguistic and Temporal Grounding in Video-Language Models
Ilker Kesen (Koç University), Erkut Erdem (Hacettepe University)
TransformerVision Language ModelVideoTextBenchmark
🎯 What it does: Designed and released the VILMA zero-shot video-language model benchmark, using adversarial foiling to evaluate the model's fine-grained spatiotemporal language understanding.
Vision Transformers Need Registers
Timothée Darcet (University of Grenoble Alpes), Piotr Bojanowski (Meta)
Object DetectionSegmentationDepth EstimationTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This study investigates the high norm token artifacts that occur in Vision Transformers and proposes the addition of learnable register tokens in the input sequence to eliminate these artifacts, thereby enhancing the performance of dense prediction and unsupervised object discovery.
Vision-by-Language for Training-Free Compositional Image Retrieval
Shyamgopal Karthik (Tubingen AI Center University of Tubingen), Zeynep Akata (University of Trento)
RetrievalTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: A training-free zero-shot compositional image retrieval framework CIReVL is proposed, which first generates descriptions of query images using a vision-language model, then utilizes a large language model to generate descriptions of target images based on textual modification instructions, and finally employs CLIP for text-image retrieval.
Vision-Language Foundation Models as Effective Robot Imitators
Xinghang Li (Tsinghua University), Tao Kong (ByteDance Research)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A policy head is added to the existing open-source vision-language model OpenFlamingo, and through imitation learning with a small amount of language-annotated robot demonstration data, a model called RoboFlamingo is obtained, which can directly execute robot manipulation commands.
Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning
Juan Rocamonde (FAR AI), David Lindner (ETH Zurich)
TransformerReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A framework utilizing pre-trained visual-language models (such as CLIP) as zero-shot reward models (VLM-RM) is proposed and implemented, which is used to drive reinforcement learning agents to complete various tasks.
Visual Data-Type Understanding does not emerge from scaling Vision-Language Models
Vishaal Udandarao (University of Cambridge), Matthias Bethge (University of Tübingen)
RecognitionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper studies the task of Visual Data-Type Identification, constructing two datasets: SyntheticTypeIdent and NaturalTypeIdent, and systematically evaluates the zero-shot performance of 39 visual language models (VLMs) on this task.
Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis
Hubert Siuzdak (Gemelo AI)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkGenerative Adversarial NetworkAudio
🎯 What it does: This paper presents Vocos, a speech synthesis model that directly generates complex STFT coefficients using GANs and achieves upsampling through inverse STFT, completely avoiding traditional transposed convolutions.
VONet: Unsupervised Video Object Learning With Parallel U-Net Attention and Object-wise Sequential VAE
Haonan Yu (Horizon Robotics), Wei Xu (Horizon Robotics)
Object DetectionSegmentationGenerationTransformerAuto EncoderVideo
🎯 What it does: This paper proposes an unsupervised video object learning framework called VONet, which can simultaneously generate attention masks for multiple objects while maintaining temporal consistency.
VQ-TR: Vector Quantized Attention for Time Series Forecasting
Kashif Rasul (Morgan Stanley), Yuriy Nevmyvaka (Morgan Stanley)
TransformerTime Series
🎯 What it does: This paper proposes a vector quantization-based Transformer architecture, VQ-TR, for efficient probabilistic time series forecasting.
VQGraph: Rethinking Graph Representation Space for Bridging GNNs and MLPs
Ling Yang (Peking University), Jure Leskovec (Stanford University)
Computational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: Learn discrete encoding of local subgraphs of nodes through an improved VQ-VAE structure, and use this codebook as a benchmark to distill the structural knowledge of the teacher GNN into an MLP without adjacency queries, achieving efficient inference;
Waxing-and-Waning: a Generic Similarity-based Framework for Efficient Self-Supervised Learning
Sheng Li (University of Pittsburgh), Geng Yuan (University of Georgia)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The SIMWNW framework is proposed to reduce computation by identifying and removing similar image patches and feature map regions in self-supervised learning.
Weaker MVI Condition: Extragradient Methods with Multi-Step Exploration
Yifeng Fan (Zhejiang University of Technology), Bo Chen (Zhejiang University of Technology)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a multi-step extrapolation limit gradient algorithm (n-step EG) and an adaptive maximum distance limit gradient (MDEG) for solving non-convex-non-concave minimax problems under weak Minty variational inequalities (weak MVI);
Weakly Supervised Virus Capsid Detection with Image-Level Annotations in Electron Microscopy Images
Hannah Kniesel (Ulm University), Pedro Hermosilla (Vienna University of Technology)
Object DetectionTransformerImage
🎯 What it does: A weakly supervised virus capsid detection method based solely on image-level labels is proposed, utilizing a pre-trained classifier and gradient optimization of a Gaussian mask to achieve particle localization.
Weakly-supervised Audio Separation via Bi-modal Semantic Similarity
Tanvir Mahmud (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
Data SynthesisContrastive LearningMultimodalityAudio
🎯 What it does: A weakly supervised language-conditioned audio separation framework is proposed, which utilizes a pre-trained multimodal audio-text embedding model (CLAP) to provide weak supervision signals for single-source extraction, enabling target source separation without the need for single-source audio data during the training phase.
Weatherproofing Retrieval for Localization with Generative AI and Geometric Consistency
Yannis Kalantidis (NAVER LABS Europe), Gabriela Csurka (NAVER LABS Europe)
GenerationData SynthesisRetrievalAutonomous DrivingContrastive LearningImage
🎯 What it does: Use text prompt-driven generative models to expand the training set, and improve the retrieval step of visual localization by combining geometric consistency filtering.
WebArena: A Realistic Web Environment for Building Autonomous Agents
Shuyan Zhou (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: We constructed WebArena, a reproducible and fully functional web environment, and released 812 long-term task benchmarks based on real websites to evaluate the performance of LLMs in natural language-driven web interactions.
What Algorithms can Transformers Learn? A Study in Length Generalization
Hattie Zhou (Apple), Preetum Nakkiran (Apple)
TransformerLarge Language ModelSequential
🎯 What it does: Explores the length generalization ability of Transformers in algorithmic tasks and proposes the RASP language framework to predict which tasks can generalize.
What does automatic differentiation compute for neural networks?
Sejun Park (Korea University), Wonyeol Lee (NAVER AI Lab)
Convolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: The study investigates the results of automatic differentiation (AD) in neural networks that include non-differentiable functions (such as ReLU and maxpool), proving that AD always outputs the Clarke subderivative when specific conditions are met;
What does the Knowledge Neuron Thesis Have to do with Knowledge?
Jingcheng Niu (University of Toronto), Gerald Penn (University of Toronto)
Large Language ModelText
🎯 What it does: Evaluate the Knowledge Neuron (KN) theory and its model editing methods based on MLP weights (KN Edit, ROME), and expand the evaluation scope to synthetic grammatical phenomena, introducing two new editing evaluation metrics: symmetry and synonymy.
What Makes a Good Prune? Maximal Unstructured Pruning for Maximal Cosine Similarity
Gabryel Mason-Williams (Queen Mary University), Fredrik Dahlqvist (Queen Mary University)
Convolutional Neural NetworkImage
🎯 What it does: This study investigates the effect of global unstructured magnitude pruning (L1) while maintaining maximum cosine similarity and proposes determining the optimal pruning ratio through the curvature of parameter distribution.
What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning
Wei Liu (ShanghaiTech University), Junxian He (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: A comprehensive study on data selection in instruction tuning is conducted, proposing the EVOL COMPLEXITY/QUALITY and REPR FILTER strategies. Based on this, the DEITA model is trained, achieving alignment performance comparable to traditional datasets of 100K–300K scale with only 6K samples.
What Matters to You? Towards Visual Representation Alignment for Robot Learning
Thomas Tian, Andrea Bajcsy (Carnegie Mellon University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningVideo
🎯 What it does: This paper proposes the RAPL method, which aligns visual representations based on human preferences for video demonstrations and constructs a reward function through optimal transport, enabling robots to learn behaviors that align with user preferences.
What's in a Prior? Learned Proximal Networks for Inverse Problems
Zhenghan Fang (Johns Hopkins University), Jeremias Sulam (Johns Hopkins University)
RestorationCompressionConvolutional Neural NetworkImageComputed Tomography
🎯 What it does: This paper proposes a Learning-based Proximal Operator Network (LPN), which achieves an exact proximal operator for any non-convex regularization function by parameterizing the network as the gradient of a convex function. It learns the logarithmic prior of the data using proximal matching loss in an unsupervised manner, and subsequently embeds LPN into the Plug-and-Play (PnP) framework to solve inverse problems, providing convergence guarantees.
What's In My Big Data?
Yanai Elazar (Allen Institute for AI), Jesse Dodge (Allen Institute for AI)
Large Language ModelText
🎯 What it does: This paper constructs the WIMBD platform, combining counting and search capabilities, to conduct sixteen reproducible analyses on ten large language model training corpora, revealing issues such as duplication, low quality, PII leakage, toxicity, and benchmark contamination.
When can transformers reason with abstract symbols?
Enric Boix-Adserà (Apple), Joshua M. Susskind
TransformerLarge Language ModelText
🎯 What it does: This study investigates the learning and generalization capabilities of the Transformer in abstract symbolic relational reasoning tasks (template tasks) and demonstrates that it can learn abstract relationships under gradient descent and with sufficiently large samples; at the same time, it proves that traditional MLPs cannot accomplish such reasoning.
When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations
Aleksandar Petrov (University of Oxford), Adel Bibi (University of Oxford)
TransformerPrompt Engineering
🎯 What it does: The paper conducts a theoretical analysis of context-based fine-tuning methods for pre-trained Transformers (prompting, soft prompting, prefix-tuning), exploring their expressive capabilities and limitations, and validates conclusions through experiments with small Transformers.
When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method
Biao Zhang (Google DeepMind), Orhan Firat (Google DeepMind)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The system studied the scaling laws of LLM fine-tuning, exploring the impact of model size, pre-training data, fine-tuning data, and parameter-efficient fine-tuning methods on performance.
When Semantic Segmentation Meets Frequency Aliasing
Linwei Chen (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper systematically analyzes three types of challenging errors (false positives, merging errors, and displacement errors) of boundary pixels in semantic segmentation, finding that they are related to aliasing phenomena during the downsampling process. It proposes an aliasing score based on equivalent sampling rate calculations. Subsequently, it designs a de-aliasing filter (DAF) and a frequency mixing module (FreqMix) to accurately eliminate or reconstruct high-frequency components to mitigate the negative impact of aliasing on segmentation.
When should we prefer Decision Transformers for Offline Reinforcement Learning?
Prajjwal Bhargava (Meta AI), Amy Zhang (University of Texas)
TransformerReinforcement LearningTabular
🎯 What it does: A systematic evaluation of three mainstream algorithms in offline reinforcement learning (CQL, BC, DT) was conducted under different conditions of data quality, task complexity, reward sparsity, and randomness, and a guideline for algorithm selection in various scenarios was provided.
Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in DNNs
Qihan Ren (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
ImagePoint CloudTabular
🎯 What it does: It is proven that under specific conditions, trained deep neural networks only encode a small number of sparse interaction primitives, thereby explaining the emergence of symbolic concepts.
Whittle Index with Multiple Actions and State Constraint for Inventory Management
Chuheng Zhang (Microsoft Research Asia), Jiang Bian (Microsoft Research Asia)
OptimizationReinforcement LearningTabular
🎯 What it does: An extended Whittle index (WIMS) and its deep reinforcement learning implementation WIMSN are proposed for multi-SKU inventory management under global inventory capacity constraints.
Whole-Song Hierarchical Generation of Symbolic Music Using Cascaded Diffusion Models
Ziyu Wang (New York University), Gus Xia (Mohamed bin Zayed University of Artificial Intelligence)
GenerationData SynthesisDiffusion modelAudio
🎯 What it does: A framework for generating full-score symbolic music based on a hierarchical language is proposed, which achieves automatic generation of entire pieces through a four-level hierarchical language (Form, Reduced Lead Sheet, Lead Sheet, Accompaniment) and a cascade diffusion model.
Why is SAM Robust to Label Noise?
Christina Baek (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
OptimizationImage
🎯 What it does: This study investigates the robustness mechanism of Sharpness-Aware Minimization (SAM) in the presence of label noise.
WildChat: 1M ChatGPT Interaction Logs in the Wild
Wenting Zhao (Cornell University), Yuntian Deng (Allen Institute for Artificial Intelligence)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Collected and publicly released over 1 million real user multi-turn dialogue logs with ChatGPT, forming a multilingual, multi-turn, anonymous dataset with geographical information called WILDCHAT.
WildFusion: Learning 3D-Aware Latent Diffusion Models in View Space
Katja Schwarz (University of Tübingen), Karsten Kreis (University of Tübingen)
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: We designed and implemented WildFusion, a 3D-aware image synthesis framework based on latent space diffusion models, capable of generating high-quality, view-consistent 3D results from single-view images without pose information or multi-view supervision.
Win-Win: Training High-Resolution Vision Transformers from Two Windows
Vincent Leroy (Naver Labs Europe), Philippe Weinzaepfel (Naver Labs Europe)
SegmentationAutonomous DrivingTransformerOptical FlowImageVideo
🎯 What it does: A 'Win-Win' training strategy is proposed: randomly masking most patches on high-resolution images while retaining N windows (N≥2), using only the tokens from these windows during training, and performing a single forward pass on the complete image during testing.
Window Attention is Bugged: How not to Interpolate Position Embeddings
Daniel Bolya (Georgia Tech), Christoph Feichtenhofer (Meta)
Object DetectionSegmentationTransformerImageVideo
🎯 What it does: Identified and fixed a bug that occurred during high-resolution fine-tuning when using window attention and absolute position embeddings, and resolved the issue by introducing the 'absolute win' position embedding scheme.
WizardCoder: Empowering Code Large Language Models with Evol-Instruct
Ziyang Luo (Microsoft), Daxin Jiang (Microsoft)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Code Evol-Instruct scheme, which automatically evolves the existing Code Alpaca instruction data, and then uses the evolved data for fine-tuning StarCoder and CodeLlama with fine-grained instructions to generate the WizardCoder series models.
WizardLM: Empowering Large Pre-Trained Language Models to Follow Complex Instructions
Can Xu (Microsoft), Daxin Jiang (Microsoft)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes the Evol-Instruct method, which utilizes LLM to automatically evolve and expand initial instruction data, generating diverse and controllable difficulty open-domain instructions. The WizardLM model trained with these instructions significantly improves performance on multiple LLM benchmarks.
Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models
Pablo Pernias, Marc Aubreville (Technische Hochschule Ingolstadt)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageText
🎯 What it does: A three-stage architecture called Wurstchen is proposed, which achieves efficient text-to-image diffusion generation using low-dimensional dense latent representations.
Xformer: Hybrid X-Shaped Transformer for Image Denoising
Jiale Zhang (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationTransformerImage
🎯 What it does: A Transformer network called Xformer with an X-shaped structure is proposed for image denoising.
YaRN: Efficient Context Window Extension of Large Language Models
Bowen Peng (Nous Research), Enrico Shippole (EleutherAI)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the YaRN method, which improves the interpolation of RoPE positional encoding and temperature scaling to extend the context window of LLMs from thousands to hundreds of thousands without the need for large amounts of data or training steps.
Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML
Robin van de Water (Hasso Plattner Institute), Patrick Rockenschaub (Amsterdam UMC)
Hyperparameter SearchRecurrent Neural NetworkTransformerTabularBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: A scalable ICU prediction benchmark framework YAIB has been developed, unifying data extraction, task definition, preprocessing, and model training.
You Only Query Once: An Efficient Label-Only Membership Inference Attack
YUTONG WU, Tianwei Zhang (Nanyang Technological University)
Adversarial AttackImageTabular
🎯 What it does: A novel membership inference attack called YOQO is proposed, which queries the label only once to determine whether the target sample belongs to the training set by generating query samples in an improved region.
Zero and Few-shot Semantic Parsing with Ambiguous Inputs
Elias Stengel-Eskin (University of North Carolina at Chapel Hill), Benjamin Van Durme (Johns Hopkins University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the AMP framework and dataset, specifically designed to address the ambiguity issues in semantic parsing, and designs zero-shot and few-shot evaluation tasks to assess the performance of large language models when faced with ambiguous sentences.
Zero Bubble (Almost) Pipeline Parallelism
Penghui Qi (Sea AI Lab), Min Lin (Sea AI Lab)
TransformerLarge Language ModelText
🎯 What it does: A scheduling strategy is proposed to achieve near-zero pipeline bubbles under synchronous training, which splits backpropagation into input gradients and parameter gradients, and provides an automatic scheduling algorithm.
Zero-Mean Regularized Spectral Contrastive Learning: Implicitly Mitigating Wrong Connections in Positive-Pair Graphs
Xiong Zhou (Harbin Institute of Technology), Xiangyang Ji (Tsinghua University)
Domain AdaptationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a zero-mean regularization method for spectral contrastive learning (SpeCL), which relaxes the orthogonal constraints on negative samples by adding a tunable constant τ to the negative sample term and uniformly reduces the weights of positive sample graphs, thereby weakening incorrect positive sample connections and enhancing the robustness of unsupervised domain adaptation and noisy label supervision tasks.
Zero-Shot Continuous Prompt Transfer: Generalizing Task Semantics Across Language Models
Zijun Wu (University of Alberta), Lili Mou (University of Alberta)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A zero-shot continuous prompt transfer method for cross-lingual models is proposed, utilizing encode-then-search relative space encoding and target model search.
Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models
Kevin Black (University of California), Sergey Levine (University of California)
Robotic IntelligenceDiffusion modelImageVideo
🎯 What it does: This paper proposes the SuSIE method, which combines a pre-trained image editing diffusion model with robot video fine-tuning to generate intermediate sub-goals, and then achieves zero-shot language-driven robotic operations through low-level goal accomplishment strategies.
Zero-Shot Robustification of Zero-Shot Models
Dyah Adila (University of Wisconsin Madison), Frederic Sala (University of Wisconsin Madison)
ClassificationLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: A zero-shot robustification method called ROBOSHOT is proposed, which utilizes harmful and beneficial concept embedding vectors generated by a language model to project and amplify the embedding space of a pre-trained model in an unsupervised manner, thereby enhancing the robustness of zero-shot classification.
ZeRO++: Extremely Efficient Collective Communication for Large Model Training
Guanhua Wang (Microsoft), Yuxiong He (Snowflake)
Large Language ModelReinforcement LearningText
🎯 What it does: This paper proposes ZeRO++, which significantly reduces communication volume and improves the throughput of large model training in ZeRO-3 by using low-precision weight quantization, hierarchical weight partitioning, and quantized gradient communication.
ZeroFlow: Scalable Scene Flow via Distillation
Kyle Vedder (University of Pennsylvania), James Hays (Georgia Tech)
Autonomous DrivingOptimizationKnowledge DistillationFlow-based ModelPoint Cloud
🎯 What it does: We propose ZeroFlow, a zero-label scalable scene flow estimation framework that generates pseudo-labels through unlabeled optimization and distills them to a feedforward model;
Zeroth-Order Optimization Meets Human Feedback: Provable Learning via Ranking Oracles
Zhiwei Tang (Chinese University of Hong Kong), Tsung-Hui Chang (Chinese University of Hong Kong)
OptimizationTabularSequential
🎯 What it does: A provable zero-order optimization algorithm ZO-RankSGD is proposed, which uses a ranking oracle to optimize black-box functions.
Zipformer: A faster and better encoder for automatic speech recognition
Zengwei Yao (Xiaomi Corporation), Daniel Povey (Xiaomi Corporation)
RecognitionComputational EfficiencyTransformerAudio
🎯 What it does: A new ASR encoder called Zipformer is proposed, which improves the U-Net structure, block structure, normalization, activation functions, and optimizers, enhancing speed and performance.
ZipIt! Merging Models from Different Tasks without Training
George Stoica (Georgia Tech), Judy Hoffman (Georgia Tech)
ClassificationRecognitionGraph Neural NetworkImage
🎯 What it does: A method called ZipIt! is proposed for merging models of different tasks without additional training.
Zoology: Measuring and Improving Recall in Efficient Language Models
Simran Arora (Stanford University), Christopher Re
Convolutional Neural NetworkTransformerLarge Language ModelText
🎯 What it does: This paper focuses on the non-attentive gated convolutional language model, systematically evaluating and improving its associative recall performance on real data, and further explaining the gap between it and Transformers.