Author name cluster

Peihao Wang

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

23 papers

2 author rows

AAAI Conference 2026 Conference Paper

Oscillation Inversion: Training-Free Image and Video Enhancement Through Oscillated Latents in Large Flow Models

Yan Zheng
Zhenxiao Liang
Xiaoyan Cong
Yi Yang
Lanqing Guo
Yuehao Wang
Peihao Wang
Zhangyang Wang

We explore the oscillatory behavior observed in inversion methods applied to large-scale flow models, including text-to-image and text-to-video. By employing an augmented fixed-point-inspired iterative approach to invert real-world images, we observe that the solution does not achieve convergence, instead oscillating between distinct clusters. Through both experiments on synthetic data, text-to-image and text-to-video, we demonstrate that these oscillating clusters exhibit notable semantic coherence. We offer theoretical insights, showing that this behavior arises from oscillatory dynamics in flow models. Building on this understanding, we introduce a simple and fast distribution transfer technique that facilitates training-free image and video editing/enhancement. Furthermore, we provide quantitative results demonstrating the effectiveness of our method on tasks such as image enhancement, editing, and reconstruction. Notably, our approach enables the transformation of image-only enhancers and editors into lightweight, video-capable tools—without additional training—highlighting its practical versatility and impact.

PDF Details DOI

NeurIPS Conference 2025 Conference Paper

Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights

Zhiyuan Liang
Dongwen Tang
Yuhao Zhou
Xuanlei Zhao
Mingjia Shi
Wangbo Zhao
Zekai Li
Peihao Wang

Modern Parameter-Efficient Fine-Tuning (PEFT) methods such as low-rank adaptation (LoRA) reduce the cost of customizing large language models (LLMs), yet still require a separate optimization run for every downstream dataset. We introduce \textbf{Drag-and-Drop LLMs (\textit{DnD})}, a prompt-conditioned parameter generator that eliminates per-task training by mapping a handful of unlabeled task prompts directly to LoRA weight updates. A lightweight text encoder distills each prompt batch into condition embeddings, which are then transformed by a cascaded hyper-convolutional decoder into the full set of LoRA matrices. Once trained in a diverse collection of prompt-checkpoint pairs, DnD produces task-specific parameters in seconds, yielding i) up to \textbf{12, 000$\times$} lower overhead than full fine-tuning, ii) average gains up to \textbf{30\%} in performance over the strongest training LoRAs on unseen common-sense reasoning, math, coding, and multimodal benchmarks, and iii) robust cross-domain generalization improving \textbf{40\%} performance without access to the target data or labels. Our results demonstrate that prompt-conditioned parameter generation is a viable alternative to gradient-based adaptation for rapidly specializing LLMs. We open source \href{https: //jerryliang24. github. io/DnD}{our project} in support of future research.

PDF Details

NeurIPS Conference 2025 Conference Paper

Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models

Haoyu Wang
Peihao Wang
Mufei Li
Shikun Liu
Siqi Miao
Zhangyang "Atlas" Wang
Pan Li

Modern large language models (LLMs) are inherently auto-regressive, requiring input to be serialized into flat sequences regardless of their structural dependencies. This serialization hinders the model’s ability to leverage structural inductive biases, especially in tasks such as retrieval-augmented generation (RAG) and reasoning on data with native graph structures, where inter-segment dependencies are crucial. We introduce Graph-KV with the potential to overcome this limitation. Graph-KV leverages the KV-cache of text segments as condensed representations and governs their interaction through structural inductive biases. In this framework, ''target'' segments selectively attend only to the KV-caches of their designated ''source'' segments, rather than all preceding segments in a serialized sequence. This approach induces a graph-structured block mask, sparsifying attention and enabling a message-passing-like step within the LLM. Furthermore, strategically allocated positional encodings for source and target segments reduce positional bias and context window consumption. We evaluate Graph-KV across three scenarios: (1) seven RAG benchmarks spanning direct inference, multi-hop reasoning, and long-document understanding; (2) Arxiv-QA, a novel academic paper QA task with full-text scientific papers structured as citation ego-graphs; and (3) paper topic classification within a citation network. By effectively reducing positional bias and harnessing structural inductive biases, Graph-KV substantially outperforms baselines, including standard costly sequential encoding, across various settings.

PDF Details

ICML Conference 2025 Conference Paper

Rethinking Addressing in Language Models via Contextualized Equivariant Positional Encoding

Jiajun Zhu
Peihao Wang
Ruisi Cai
Jason D. Lee
Pan Li 0005
Zhangyang Wang

Transformers rely on both content-based and position-based addressing mechanisms to make predictions, but existing positional encoding techniques often diminish the effectiveness of position-based addressing. Many current methods enforce rigid patterns in attention maps, limiting the ability to model long-range dependencies and adapt to diverse tasks. Additionally, most positional encodings are learned as general biases, lacking the specialization required for different instances within a dataset. To address this, we propose con T extualized equivari A nt P osition E ncoding ( TAPE ), a novel framework that enhances positional embeddings by incorporating sequence content across layers. TAPE introduces dynamic, context-aware positional encodings, overcoming the constraints of traditional fixed patterns. We show that TAPE can provably facilitate LLM reasoning ability by emulating a broader class of algorithms. By enforcing permutation and orthogonal equivariance, TAPE ensures the stability of positional encodings during updates, improving long-context ability. Our method can be easily integrated into pre-trained transformers, offering parameter-efficient fine-tuning with minimal overhead. Extensive experiments show that TAPE achieves superior performance in language modeling, arithmetic reasoning, and long-context retrieval tasks compared to existing positional embedding techniques.

Details

NeurIPS Conference 2025 Conference Paper

SAS: Simulated Attention Score

Chuanyang Zheng
Jiankai Sun
Yihang Gao
Yuehao Wang
Peihao Wang
Jing Xiong
Liliang Ren
Hao Cheng

The attention mechanism is a core component of the Transformer architecture. Various methods have been developed to compute attention scores, including multi-head attention (MHA), multi-query attention, group-query attention and so on. We further analyze the MHA and observe that its performance improves as the number of attention heads increases, provided the hidden size per head remains sufficiently large. Therefore, increasing both the head count and hidden size per head with minimal parameter overhead can lead to significant performance gains at a low cost. Motivated by this insight, we introduce Simulated Attention Score (SAS), which maintains a compact model size while simulating a larger number of attention heads and hidden feature dimension per head. This is achieved by projecting a low-dimensional head representation into a higher-dimensional space, effectively increasing attention capacity without increasing parameter count. Beyond the head representations, we further extend the simulation approach to feature dimension of the key and query embeddings, enhancing expressiveness by mimicking the behavior of a larger model while preserving the original model size. To control the parameter cost, we also propose Parameter-Efficient Attention Aggregation (PEAA). Comprehensive experiments on a variety of datasets and tasks demonstrate the effectiveness of the proposed SAS method, achieving significant improvements over different attention variants.

PDF Details

ICLR Conference 2025 Conference Paper

Understanding and Mitigating Bottlenecks of State Space Models through the Lens of Recency and Over-smoothing

Peihao Wang
Ruisi Cai
Yuehao Wang
Jiajun Zhu
Pragya Srivastava
Zhangyang Wang
Pan Li 0005

Structured State Space Models (SSMs) have emerged as alternatives to transformers. While SSMs are often regarded as effective in capturing long-sequence dependencies, we rigorously demonstrate that they are inherently limited by strong recency bias. Our empirical studies also reveal that this bias impairs the models' ability to recall distant information and introduces robustness issues. Our scaling experiments then discovered that deeper structures in SSMs can facilitate the learning of long contexts. However, subsequent theoretical analysis reveals that as SSMs increase in depth, they exhibit another inevitable tendency toward over-smoothing, e.g., token representations becoming increasingly indistinguishable. This *fundamental dilemma* between recency and over-smoothing hinders the scalability of existing SSMs. Inspired by our theoretical findings, we propose to *polarize* two channels of the state transition matrices in SSMs, setting them to zero and one, respectively, simultaneously addressing recency bias and over-smoothing. Experiments demonstrate that our polarization technique consistently enhances the associative recall accuracy of long-range tokens and unlocks SSMs to benefit further from deeper architectures. All source codes are released at https://github.com/VITA-Group/SSM-Bottleneck.

Details

NeurIPS Conference 2024 Conference Paper

$\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design

Ruisi Cai
Yeonju Ro
Geon-Woo Kim
Peihao Wang
Babak E. Bejnordi
Aditya Akella
Zhangyang Wang

The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and performance. Despite their benefits, MoE models face significant challenges during inference, including inefficient memory management and suboptimal batching, due to misaligned design choices between the model architecture and the system policies. Furthermore, the conventional approach of training MoEs from scratch is increasingly prohibitive in terms of cost. In this paper, we propose a novel framework $\textit{Read-ME}$ that transforms pre-trained dense LLMs into smaller MoE models (in contrast to ``upcycling" generalist MoEs), avoiding the high costs of ground-up training. Our approach employs activation sparsity to extract experts. To compose experts, we examine the widely-adopted layer-wise router design and show its redundancy, and thus we introduce the pre-gating router decoupled from the MoE backbone that facilitates system-friendly pre-computing and lookahead scheduling, enhancing expert-aware batching and caching. Our codesign therefore addresses critical gaps on both the algorithmic and system fronts, establishing a scalable and efficient alternative for LLM inference in resource-constrained settings. $\textit{Read-ME}$ outperforms other popular open-source dense models of similar scales, achieving improvements of up to 10. 1\% on MMLU, and improving mean end-to-end latency up to 6. 1\%. Codes are available at: \url{https: //github. com/VITA-Group/READ-ME}.

PDF Details DOI

NeurIPS Conference 2024 Conference Paper

$\texttt{Model-GLUE}$: Democratized LLM Scaling for A Large Model Zoo in the Wild

Xinyu Zhao
Guoheng Sun
Ruisi Cai
Yukun Zhou
Pingzhi Li
Peihao Wang
Bowen Tan
Yexiao He

As Large Language Models (LLMs) excel across tasks and specialized domains, scaling LLMs based on existing models has gained significant attention, which is challenged by potential performance drop when combining disparate models. Various techniques have been proposed to aggregate pre-trained LLMs, including model merging, Mixture-of-Experts, and stacking. Despite their merits, a comprehensive comparison and synergistic application of them to a diverse model zoo is yet to be adequately addressed. In light of this research gap, this paper introduces $\texttt{Model-GLUE}$, a holistic LLM scaling guideline. First, our work starts with a benchmarking of existing LLM scaling techniques, especially selective merging, and variants of mixture. Utilizing the insights from the benchmark results, we formulate a strategy for the selection and aggregation of a heterogeneous model zoo characterizing different architectures and initialization. Our methodology involves clustering mergeable models, selecting a merging strategy, and integrating model clusters through model-level mixture. Finally, evidenced by our experiments on a diverse Llama-2-based model zoo, $\texttt{Model-GLUE}$ shows an average performance enhancement of 5. 61\%, achieved without additional training. Codes are available at https: //github. com/Model-GLUE/Model-GLUE.

PDF Details DOI

NeurIPS Conference 2024 Conference Paper

Large Spatial Model: End-to-end Unposed Images to Semantic 3D

Zhiwen Fan
Jian Zhang
Wenyan Cong
Peihao Wang
Renjie Li
Kairun Wen
shijie zhou
Achuta Kadambi

Reconstructing and understanding 3D structures from a limited number of images is a classical problem in computer vision. Traditional approaches typically decompose this task into multiple subtasks, involving several stages of complex mappings between different data representations. For example, dense reconstruction using Structure-from-Motion (SfM) requires transforming images into key points, optimizing camera parameters, and estimating structures. Following this, accurate sparse reconstructions are necessary for further dense modeling, which is then input into task-specific neural networks. This multi-stage paradigm leads to significant processing times and engineering complexity. In this work, we introduce the Large Spatial Model (LSM), which directly processes unposed RGB images into semantic radiance fields. LSM simultaneously estimates geometry, appearance, and semantics in a single feed-forward pass and can synthesize versatile label maps by interacting through language at novel views. Built on a general Transformer-based framework, LSM predicts global geometry via pixel-aligned point maps. To improve spatial attribute regression, we adopt local context aggregation with multi-scale fusion, enhancing the accuracy of fine local details. To address the scarcity of labeled 3D semantic data and enable natural language-driven scene manipulation, we incorporate a pre-trained 2D language-based segmentation model into a 3D-consistent semantic feature field. An efficient decoder parameterizes a set of semantic anisotropic Gaussians, allowing supervised end-to-end learning. Comprehensive experiments on various tasks demonstrate that LSM unifies multiple 3D vision tasks directly from unposed images, achieving real-time semantic 3D reconstruction for the first time.

PDF Details DOI

ICLR Conference 2024 Conference Paper

Polynomial Width is Sufficient for Set Representation with High-dimensional Features

Peihao Wang
Shenghao Yang 0002
Shu Li
Zhangyang Wang
Pan Li 0005

Set representation has become ubiquitous in deep learning for modeling the inductive bias of neural networks that are insensitive to the input order. DeepSets is the most widely used neural network architecture for set representation. It involves embedding each set element into a latent space with dimension $L$, followed by a sum pooling to obtain a whole-set embedding, and finally mapping the whole-set embedding to the output. In this work, we investigate the impact of the dimension $L$ on the expressive power of DeepSets. Previous analyses either oversimplified high-dimensional features to be one-dimensional features or were limited to complex analytic activations, thereby diverging from practical use or resulting in $L$ that grows exponentially with the set size $N$ and feature dimension $D$. To investigate the minimal value of $L$ that achieves sufficient expressive power, we present two set-element embedding layers: (a) linear + power activation (LP) and (b) linear + exponential activations (LE). We demonstrate that $L$ being $\operatorname{poly}(N, D)$ is sufficient for set representation using both embedding layers. We also provide a lower bound of $L$ for the LP embedding layer. Furthermore, we extend our results to permutation-equivariant set functions and the complex field.

Details

ICML Conference 2023 Conference Paper

Data Efficient Neural Scaling Law via Model Reusing

Peihao Wang
Rameswar Panda
Zhangyang Wang

The number of parameters in large transformers has been observed to grow exponentially. Despite notable performance improvements, concerns have been raised that such a growing model size will run out of data in the near future. As manifested in the neural scaling law, modern learning backbones are not data-efficient. To maintain the utility of the model capacity, training data should be increased proportionally. In this paper, we study the neural scaling law under the previously overlooked data scarcity regime, focusing on the more challenging situation where we need to train a gigantic model with a disproportionately limited supply of available training data. We find that the existing power laws underestimate the data inefficiency of large transformers. Their performance will drop significantly if the training set is insufficient. Fortunately, we discover another blessing - such a data-inefficient scaling law can be restored through a model reusing approach that warm-starts the training of a large model by initializing it using smaller models. Our empirical study shows that model reusing can effectively reproduce the power law under the data scarcity regime. When progressively applying model reusing to expand the model size, we also observe consistent performance improvement in large transformers. We release our code at: https: //github. com/VITA-Group/Data-Efficient-Scaling.

Details

ICLR Conference 2023 Conference Paper

Equivariant Hypergraph Diffusion Neural Operators

Peihao Wang
Shenghao Yang 0002
Yunyu Liu
Zhangyang Wang
Pan Li 0005

Hypergraph neural networks (HNNs) using neural networks to encode hypergraphs provide a promising way to model higher-order relations in data and further solve relevant prediction tasks built upon such higher-order relations. However, higher-order relations in practice contain complex patterns and are often highly irregular. So, it is often challenging to design an HNN that suffices to express those relations while keeping computational efficiency. Inspired by hypergraph diffusion algorithms, this work proposes a new HNN architecture named ED-HNN, which provably approximates any continuous equivariant hypergraph diffusion operators that can model a wide range of higher-order relations. ED-HNN can be implemented efficiently by combining star expansions of hypergraphs with standard message passing neural networks. ED-HNN further shows great superiority in processing heterophilic hypergraphs and constructing deep models. We evaluate ED-HNN for node classification on nine real-world hypergraph datasets. ED-HNN uniformly outperforms the best baselines over these nine datasets and achieves more than 2%$\uparrow$ in prediction accuracy over four datasets therein. Our code is available at: https://github.com/Graph-COM/ED-HNN.

Details

ICLR Conference 2023 Conference Paper

Is Attention All That NeRF Needs?

Mukund Varma T.
Peihao Wang
Xuxi Chen
Tianlong Chen 0001
Subhashini Venugopalan
Zhangyang Wang

We present Generalizable NeRF Transformer (GNT), a transformer-based architecture that reconstructs Neural Radiance Fields (NeRFs) and learns to render novel views on the fly from source views. While prior works on NeRFs optimize a scene representation by inverting a handcrafted rendering equation, GNT achieves neural representation and rendering that generalizes across scenes using transformers at two stages. (1) The view transformer leverages multi-view geometry as an inductive bias for attention-based scene representation, and predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views. (2) The ray transformer renders novel views using attention to decode the features from the view transformer along the sampled points during ray marching. Our experiments demonstrate that when optimized on a single scene, GNT can successfully reconstruct NeRF without an explicit rendering formula due to the learned ray renderer. When trained on multiple scenes, GNT consistently achieves state-of-the-art performance when transferring to unseen scenes and outperform all other methods by ~10% on average. Our analysis of the learned attention maps to infer depth and occlusion indicate that attention enables learning a physically-grounded rendering. Our results show the promise of transformers as a universal modeling tool for graphics. Please refer to our project page for video results: https://vita-group.github.io/GNT/

Details

ICLR Conference 2023 Conference Paper

Learning to Grow Pretrained Models for Efficient Transformer Training

Peihao Wang
Rameswar Panda
Lucas Torroba Hennigen
Philip Greengard
Leonid Karlinsky
Rogério Feris
David D. Cox
Zhangyang Wang

Scaling transformers has led to significant breakthroughs in many domains, leading to a paradigm in which larger versions of existing models are trained and released on a periodic basis. New instances of such models are typically trained completely from scratch, despite the fact that they are often just scaled-up versions of their smaller counterparts. How can we use the implicit knowledge in the parameters of smaller, extant models to enable faster training of newer, larger models? This paper describes an approach for accelerating transformer training by learning to grow pretrained transformers, where we learn to linearly map the parameters of the smaller model to initialize the larger model. For tractable learning, we factorize the linear transformation as a composition of (linear) width- and depth-growth operators, and further employ a Kronecker factorization of these growth operators to encode architectural knowledge. Extensive experiments across both language and vision transformers demonstrate that our learned Linear Growth Operator (LiGO) can save up to 50% computational cost of training from scratch, while also consistently outperforming strong baselines that also reuse smaller pretrained models to initialize larger models.

Details

ICLR Conference 2023 Conference Paper

NeRF-SOS: Any-View Self-supervised Object Segmentation on Complex Scenes

Zhiwen Fan
Peihao Wang
Yifan Jiang 0001
Xinyu Gong
Dejia Xu
Zhangyang Wang

Neural volumetric representations have shown the potential that Multi-layer Perceptrons (MLPs) can be optimized with multi-view calibrated images to represent scene geometry and appearance without explicit 3D supervision. Object segmentation can enrich many downstream applications based on the learned radiance field. However, introducing hand-crafted segmentation to define regions of interest in a complex real-world scene is non-trivial and expensive as it acquires per view annotation. This paper carries out the exploration of self-supervised learning for object segmentation using NeRF for complex real-world scenes. Our framework, called NeRF with Self-supervised Object Segmentation (NeRF-SOS), couples object segmentation and neural radiance field to segment objects in any view within a scene. By proposing a novel collaborative contrastive loss in both appearance and geometry levels, NeRF-SOS encourages NeRF models to distill compact geometry-aware segmentation clusters from their density fields and the self-supervised pre-trained 2D visual features. The self-supervised object segmentation framework can be applied to various NeRF models that both lead to photo-realistic rendering results and convincing segmentation maps for both indoor and outdoor scenarios. Extensive results on the LLFF, BlendedMVS, CO3Dv2, and Tank & Temples datasets validate the effectiveness of NeRF-SOS. It consistently surpasses other 2D-based self-supervised baselines and predicts finer object masks than existing supervised counterparts.

Details

NeurIPS Conference 2023 Conference Paper

Patch Diffusion: Faster and More Data-Efficient Training of Diffusion Models

Zhendong Wang
Yifan Jiang
Huangjie Zheng
Peihao Wang
Pengcheng He
Zhangyang "Atlas" Wang
Weizhu Chen
Mingyuan Zhou

Diffusion models are powerful, but they require a lot of time and data to train. We propose Patch Diffusion, a generic patch-wise training framework, to significantly reduce the training time costs while improving data efficiency, which thus helps democratize diffusion model training to broader users. At the core of our innovations is a new conditional score function at the patch level, where the patch location in the original image is included as additional coordinate channels, while the patch size is randomized and diversified throughout training to encode the cross-region dependency at multiple scales. Sampling with our method is as easy as in the original diffusion model. Through Patch Diffusion, we could achieve $\mathbf{\ge 2\times}$ faster training, while maintaining comparable or better generation quality. Patch Diffusion meanwhile improves the performance of diffusion models trained on relatively small datasets, $e. g. $, as few as 5, 000 images to train from scratch. We achieve outstanding FID scores in line with state-of-the-art benchmarks: 1. 77 on CelebA-64$\times$64, 1. 93 on AFHQv2-Wild-64$\times$64, and 2. 72 on ImageNet-256$\times$256. We share our code and pre-trained models at https: //github. com/Zhendong-Wang/Patch-Diffusion.

PDF Details

AAAI Conference 2023 Conference Paper

PINAT: A Permutation INvariance Augmented Transformer for NAS Predictor

Shun Lu
Yu Hu
Peihao Wang
Yan Han
Jianchao Tan
Jixiang Li
Sen Yang
Ji Liu

Time-consuming performance evaluation is the bottleneck of traditional Neural Architecture Search (NAS) methods. Predictor-based NAS can speed up performance evaluation by directly predicting performance, rather than training a large number of sub-models and then validating their performance. Most predictor-based NAS approaches use a proxy dataset to train model-based predictors efficiently but suffer from performance degradation and generalization problems. We attribute these problems to the poor abilities of existing predictors to character the sub-models' structure, specifically the topology information extraction and the node feature representation of the input graph data. To address these problems, we propose a Transformer-like NAS predictor PINAT, consisting of a Permutation INvariance Augmentation module serving as both token embedding layer and self-attention head, as well as a Laplacian matrix to be the positional encoding. Our design produces more representative features of the encoded architecture and outperforms state-of-the-art NAS predictors on six search spaces: NAS-Bench-101, NAS-Bench-201, DARTS, ProxylessNAS, PPI, and ModelNet. The code is available at https://github.com/ShunLu91/PINAT.

PDF Details DOI

NeurIPS Conference 2022 Conference Paper

A Comprehensive Study on Large-Scale Graph Training: Benchmarking and Rethinking

Keyu Duan
Zirui Liu
Peihao Wang
Wenqing Zheng
Kaixiong Zhou
Tianlong Chen
Xia Hu
Zhangyang Wang

Large-scale graph training is a notoriously challenging problem for graph neural networks (GNNs). Due to the nature of evolving graph structures into the training process, vanilla GNNs usually fail to scale up, limited by the GPU memory space. Up to now, though numerous scalable GNN architectures have been proposed, we still lack a comprehensive survey and fair benchmark of this reservoir to find the rationale for designing scalable GNNs. To this end, we first systematically formulate the representative methods of large-scale graph training into several branches and further establish a fair and consistent benchmark for them by a greedy hyperparameter searching. In addition, regarding efficiency, we theoretically evaluate the time and space complexity of various branches and empirically compare them w. r. t GPU memory usage, throughput, and convergence. Furthermore, We analyze the pros and cons for various branches of scalable GNNs and then present a new ensembling training manner, named EnGCN, to address the existing issues. Remarkably, our proposed method has achieved new state-of-the-art (SOTA) performance on large-scale datasets. Our code is available at https: //github. com/VITA-Group/Large Scale GCN_Benchmarking.

PDF Details

ICLR Conference 2022 Conference Paper

Anti-Oversmoothing in Deep Vision Transformers via the Fourier Domain Analysis: From Theory to Practice

Peihao Wang
Wenqing Zheng
Tianlong Chen 0001
Zhangyang Wang

Vision Transformer (ViT) has recently demonstrated promise in computer vision problems. However, unlike Convolutional Neural Networks (CNN), it is known that the performance of ViT saturates quickly with depth increasing, due to the observed attention collapse or patch uniformity. Despite a couple of empirical solutions, a rigorous framework studying on this scalability issue remains elusive. In this paper, we first establish a rigorous theory framework to analyze ViT features from the Fourier spectrum domain. We show that the self-attention mechanism inherently amounts to a low-pass filter, which indicates when ViT scales up its depth, excessive low-pass filtering will cause feature maps to only preserve their Direct-Current (DC) component. We then propose two straightforward yet effective techniques to mitigate the undesirable low-pass limitation. The first technique, termed AttnScale, decomposes a self-attention block into low-pass and high-pass components, then rescales and combines these two filters to produce an all-pass self-attention matrix. The second technique, termed FeatScale, re-weights feature maps on separate frequency bands to amplify the high-frequency signals. Both techniques are efficient and hyperparameter-free, while effectively overcoming relevant ViT training artifacts such as attention collapse and patch uniformity. By seamlessly plugging in our techniques to multiple ViT variants, we demonstrate that they consistently help ViTs benefit from deeper architectures, bringing up to 1.1% performance gains "for free" (e.g., with little parameter overhead). We publicly release our codes and pre-trained models at https://github.com/VITA-Group/ViT-Anti-Oversmoothing.

Details

ICML Conference 2022 Conference Paper

Neural Implicit Dictionary Learning via Mixture-of-Expert Training

Peihao Wang
Zhiwen Fan
Tianlong Chen 0001
Zhangyang Wang

Representing visual signals by coordinate-based deep fully-connected networks has been shown advantageous in fitting complex details and solving inverse problems than discrete grid-based representation. However, acquiring such a continuous Implicit Neural Representation (INR) requires tedious per-scene training on tons of signal measurements, which limits its practicality. In this paper, we present a generic INR framework that achieves both data and training efficiency by learning a Neural Implicit Dictionary (NID) from a data collection and representing INR as a functional combination of wavelets sampled from the dictionary. Our NID assembles a group of coordinate-based subnetworks which are tuned to span the desired function space. After training, one can instantly and robustly acquire an unseen scene representation by solving the coding coefficients. To parallelly optimize a large group of networks, we borrow the idea from Mixture-of-Expert (MoE) to design and train our network with a sparse gating mechanism. Our experiments show that, NID can improve reconstruction of 2D images or 3D scenes by 2 orders of magnitude faster with up to 98% less input data. We further demonstrate various applications of NID in image inpainting and occlusion removal, which are considered to be challenging with vanilla INR. Our codes are available in https: //github. com/VITA-Group/Neural-Implicit-Dict.

Details

NeurIPS Conference 2022 Conference Paper

Old can be Gold: Better Gradient Flow can Make Vanilla-GCNs Great Again

AJAY JAISWAL
Peihao Wang
Tianlong Chen
Justin Rousseau
Ying Ding
Zhangyang Wang

Despite the enormous success of Graph Convolutional Networks (GCNs) in modeling graph-structured data, most of the current GCNs are shallow due to the notoriously challenging problems of over-smoothening and information squashing along with conventional difficulty caused by vanishing gradients and over-fitting. Previous works have been primarily focused on the study of over-smoothening and over-squashing phenomena in training deep GCNs. Surprisingly, in comparison with CNNs/RNNs, very limited attention has been given to understanding how healthy gradient flow can benefit the trainability of deep GCNs. In this paper, firstly, we provide a new perspective of gradient flow to understand the substandard performance of deep GCNs and hypothesize that by facilitating healthy gradient flow, we can significantly improve their trainability, as well as achieve state-of-the-art (SOTA) level performance from vanilla-GCNs. Next, we argue that blindly adopting the Glorot initialization for GCNs is not optimal, and derive a topology-aware isometric initialization scheme for vanilla-GCNs based on the principles of isometry. Additionally, contrary to ad-hoc addition of skip-connections, we propose to use gradient-guided dynamic rewiring of vanilla-GCNs with skip connections. Our dynamic rewiring method uses the gradient flow within each layer during training to introduce on-demand skip-connections adaptively. We provide extensive empirical evidence across multiple datasets that our methods improve gradient flow in deep vanilla-GCNs and significantly boost their performance to comfortably compete and outperform many fancy state-of-the-art methods. Codes are available at: https: //github. com/VITA-Group/GradientGCN.

PDF Details

NeurIPS Conference 2022 Conference Paper

Signal Processing for Implicit Neural Representations

Dejia Xu
Peihao Wang
Yifan Jiang
Zhiwen Fan
Zhangyang Wang

Implicit Neural Representations (INRs) encoding continuous multi-media data via multi-layer perceptrons has shown undebatable promise in various computer vision tasks. Despite many successful applications, editing and processing an INR remains intractable as signals are represented by latent parameters of a neural network. Existing works manipulate such continuous representations via processing on their discretized instance, which breaks down the compactness and continuous nature of INR. In this work, we present a pilot study on the question: how to directly modify an INR without explicit decoding? We answer this question by proposing an implicit neural signal processing network, dubbed INSP-Net, via differential operators on INR. Our key insight is that spatial gradients of neural networks can be computed analytically and are invariant to translation, while mathematically we show that any continuous convolution filter can be uniformly approximated by a linear combination of high-order differential operators. With these two knobs, INSP-Net instantiates the signal processing operator as a weighted composition of computational graphs corresponding to the high-order derivatives of INRs, where the weighting parameters can be data-driven learned. Based on our proposed INSP-Net, we further build the first Convolutional Neural Network (CNN) that implicitly runs on INRs, named INSP-ConvNet. Our experiments validate the expressiveness of INSP-Net and INSP-ConvNet in fitting low-level image and geometry processing kernels (e. g. blurring, deblurring, denoising, inpainting, and smoothening) as well as for high-level tasks on implicit fields such as image classification.

PDF Details

NeurIPS Conference 2021 Conference Paper

Delayed Propagation Transformer: A Universal Computation Engine towards Practical Control in Cyber-Physical Systems

Wenqing Zheng
Qiangqiang Guo
Hao Yang
Peihao Wang
Zhangyang Wang

Multi-agent control is a central theme in the Cyber-Physical Systems (CPS). However, current control methods either receive non-Markovian states due to insufficient sensing and decentralized design, or suffer from poor convergence. This paper presents the Delayed Propagation Transformer (DePT), a new transformer-based model that specializes in the global modeling of CPS while taking into account the immutable constraints from the physical world. DePT induces a cone-shaped spatial-temporal attention prior, which injects the information propagation and aggregation principles and enables a global view. With physical constraint inductive bias baked into its design, our DePT is ready to plug and play for a broad class of multi-agent systems. The experimental results on one of the most challenging CPS -- network-scale traffic signal control system in the open world -- show that our model outperformed the state-of-the-art expert methods on synthetic and real-world datasets. Our codes are released at: https: //github. com/VITA-Group/DePT.

PDF Details