Yanghao Li

Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya et al. · meta-ai

Modern hierarchical vision transformers have added several vision-specific components in the pursuit of supervised classification performance. While these components lead to effective accuracies and attractive FLOP counts, the added complexity actually makes these transformers slower than their vanilla ViT counterparts. In this paper, we argue that this additional bulk is unnecessary. By pretraining with a strong visual pretext task (MAE), we can strip out all the bells-and-whistles from a state-of-the-art multi-stage vision transformer without losing accuracy. In the process, we create Hiera, an extremely simple hierarchical vision transformer that is more accurate than previous models while being significantly faster both at inference and during training. We evaluate Hiera on a variety of tasks for image and video recognition. Our code and models are available at https://github.com/facebookresearch/hiera.

Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li et al. · meta-ai

In this work, we explore regions as a potential visual analogue of words for self-supervised image representation learning. Inspired by Masked Autoencoding (MAE), a generative pre-training baseline, we propose masked region autoencoding to learn from groups of pixels or regions. Specifically, we design an architecture which efficiently addresses the one-to-many mapping between images and regions, while being highly effective especially with high-quality regions. When integrated with MAE, our approach (R-MAE) demonstrates consistent improvements across various pre-training datasets and downstream detection and segmentation benchmarks, with negligible computational overheads. Beyond the quantitative evaluation, our analysis indicates the models pre-trained with masked region autoencoding unlock the potential for interactive segmentation. The code is provided at https://github.com/facebookresearch/r-mae.

Chen Wei, Karttikeya Mangalam, Po-Yao Huang et al. · meta-ai

There has been a longstanding belief that generation can facilitate a true understanding of visual data. In line with this, we revisit generatively pre-training visual representations in light of recent interest in denoising diffusion models. While directly pre-training with diffusion models does not produce strong representations, we condition diffusion models on masked input and formulate diffusion models as masked autoencoders (DiffMAE). Our approach is capable of (i) serving as a strong initialization for downstream recognition tasks, (ii) conducting high-quality image inpainting, and (iii) being effortlessly extended to video where it produces state-of-the-art classification accuracy. We further perform a comprehensive study on the pros and cons of design choices and build connections between diffusion models and masked autoencoders.

Po-Yao Huang, Vasu Sharma, Hu Xu et al. · meta-ai, mit

We present Masked Audio-Video Learners (MAViL) to train audio-visual representations. Our approach learns with three complementary forms of self-supervision: (1) reconstruction of masked audio and video input data, (2) intra- and inter-modal contrastive learning with masking, and (3) self-training by reconstructing joint audio-video contextualized features learned from the first two objectives. Pre-training with MAViL not only enables the model to perform well in audio-visual classification and retrieval tasks but also improves representations of each modality in isolation, without using information from the other modality for fine-tuning or inference. Empirically, MAViL sets a new state-of-the-art on AudioSet (53.1 mAP) and VGGSound (67.1% accuracy). For the first time, a self-supervised audio-visual model outperforms ones that use external supervision on these benchmarks.

Yubin Hu, Yuze He, Yanghao Li et al.

Video semantic segmentation (VSS) is a computationally expensive task due to the per-frame prediction for videos of high frame rates. In recent work, compact models or adaptive network strategies have been proposed for efficient VSS. However, they did not consider a crucial factor that affects the computational cost from the input side: the input resolution. In this paper, we propose an altering resolution framework called AR-Seg for compressed videos to achieve efficient VSS. AR-Seg aims to reduce the computational cost by using low resolution for non-keyframes. To prevent the performance degradation caused by downsampling, we design a Cross Resolution Feature Fusion (CReFF) module, and supervise it with a novel Feature Similarity Training (FST) strategy. Specifically, CReFF first makes use of motion vectors stored in a compressed video to warp features from high-resolution keyframes to low-resolution non-keyframes for better spatial alignment, and then selectively aggregates the warped features with local attention mechanism. Furthermore, the proposed FST supervises the aggregated features with high-resolution features through an explicit similarity loss and an implicit constraint from the shared decoding layer. Extensive experiments on CamVid and Cityscapes show that AR-Seg achieves state-of-the-art performance and is compatible with different segmentation backbones. On CamVid, AR-Seg saves 67% computational cost (measured in GFLOPs) with the PSPNet18 backbone while maintaining high segmentation accuracy. Code: https://github.com/THU-LYJ-Lab/AR-Seg.

Karttikeya Mangalam, Haoqi Fan, Yanghao Li et al.

We present Reversible Vision Transformers, a memory efficient architecture design for visual recognition. By decoupling the GPU memory requirement from the depth of the model, Reversible Vision Transformers enable scaling up architectures with efficient memory usage. We adapt two popular models, namely Vision Transformer and Multiscale Vision Transformers, to reversible variants and benchmark extensively across both model sizes and tasks of image classification, object detection and video classification. Reversible Vision Transformers achieve a reduced memory footprint of up to 15.5x at roughly identical model complexity, parameters and accuracy, demonstrating the promise of reversible vision transformers as an efficient backbone for hardware resource limited training regimes. Finally, we find that the additional computational burden of recomputing activations is more than overcome for deeper models, where throughput can increase up to 2.3x over their non-reversible counterparts. Full code and trained models are available at https://github.com/facebookresearch/slowfast. A simpler, easy to understand and modify version is also available at https://github.com/karttikeya/minREV

Mengmeng Xu, Yanghao Li, Cheng-Yang Fu et al.

This paper deals with the problem of localizing objects in image and video datasets from visual exemplars. In particular, we focus on the challenging problem of egocentric visual query localization. We first identify grave implicit biases in current query-conditioned model design and visual query datasets. Then, we directly tackle such biases at both frame and object set levels. Concretely, our method solves these issues by expanding limited annotations and dynamically dropping object proposals during training. Additionally, we propose a novel transformer-based module that allows for object-proposal set context to be considered while incorporating query information. We name our module Conditioned Contextual Transformer or CocoFormer. Our experiments show the proposed adaptations improve egocentric query detection, leading to a better visual query localization system in both 2D and 3D configurations. Thus, we are able to improve frame-level detection performance from 26.28% to 31.26 in AP, which correspondingly improves the VQ2D and VQ3D localization scores by significant margins. Our improved context-aware query object detector ranked first and second in the VQ2D and VQ3D tasks in the 2nd Ego4D challenge. In addition to this, we showcase the relevance of our proposed model in the Few-Shot Detection (FSD) task, where we also achieve SOTA results. Our code is available at https://github.com/facebookresearch/vq2d_cvpr.

Mengmeng Xu, Cheng-Yang Fu, Yanghao Li et al.

The recently released Ego4D dataset and benchmark significantly scales and diversifies the first-person visual perception data. In Ego4D, the Visual Queries 2D Localization task aims to retrieve objects appeared in the past from the recording in the first-person view. This task requires a system to spatially and temporally localize the most recent appearance of a given object query, where query is registered by a single tight visual crop of the object in a different scene. Our study is based on the three-stage baseline introduced in the Episodic Memory benchmark. The baseline solves the problem by detection and tracking: detect the similar objects in all the frames, then run a tracker from the most confident detection result. In the VQ2D challenge, we identified two limitations of the current baseline. (1) The training configuration has redundant computation. Although the training set has millions of instances, most of them are repetitive and the number of unique object is only around 14.6k. The repeated gradient computation of the same object lead to an inefficient training; (2) The false positive rate is high on background frames. This is due to the distribution gap between training and evaluation. During training, the model is only able to see the clean, stable, and labeled frames, but the egocentric videos also have noisy, blurry, or unlabeled background frames. To this end, we developed a more efficient and effective solution. Concretely, we bring the training loop from ~15 days to less than 24 hours, and we achieve 0.17% spatial-temporal AP, which is 31% higher than the baseline. Our solution got the first ranking on the public leaderboard. Our code is publicly available at https://github.com/facebookresearch/vq2d_cvpr.

Christoph Feichtenhofer, Haoqi Fan, Yanghao Li et al.

This paper studies a conceptually simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos. We randomly mask out spacetime patches in videos and learn an autoencoder to reconstruct them in pixels. Interestingly, we show that our MAE method can learn strong representations with almost no inductive bias on spacetime (only except for patch and positional embeddings), and spacetime-agnostic random masking performs the best. We observe that the optimal masking ratio is as high as 90% (vs. 75% on images), supporting the hypothesis that this ratio is related to information redundancy of the data. A high masking ratio leads to a large speedup, e.g., > 4x in wall-clock time or even more. We report competitive results on several challenging video datasets using vanilla Vision Transformers. We observe that MAE can outperform supervised pre-training by large margins. We further report encouraging results of training on real-world, uncurated Instagram data. Our study suggests that the general framework of masked autoencoding (BERT, MAE, etc.) can be a unified methodology for representation learning with minimal domain knowledge.

Yanghao Li, Haoqi Fan, Ronghang Hu et al.

We present Fast Language-Image Pre-training (FLIP), a simple and more efficient method for training CLIP. Our method randomly masks out and removes a large portion of image patches during training. Masking allows us to learn from more image-text pairs given the same wall-clock time and contrast more samples per iteration with similar memory footprint. It leads to a favorable trade-off between accuracy and training time. In our experiments on 400 million image-text pairs, FLIP improves both accuracy and speed over the no-masking baseline. On a large diversity of downstream tasks, FLIP dominantly outperforms the CLIP counterparts trained on the same data. Facilitated by the speedup, we explore the scaling behavior of increasing the model size, data size, or training length, and report encouraging results and comparisons. We hope that our work will foster future research on scaling vision-language learning.

Haotian Zhang, Mingfei Gao, Zhe Gan et al.

We present MM1.5, a new family of multimodal large language models (MLLMs) designed to enhance capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. Building upon the MM1 architecture, MM1.5 adopts a data-centric approach to model training, systematically exploring the impact of diverse data mixtures across the entire model training lifecycle. This includes high-quality OCR data and synthetic captions for continual pre-training, as well as an optimized visual instruction-tuning data mixture for supervised fine-tuning. Our models range from 1B to 30B parameters, encompassing both dense and mixture-of-experts (MoE) variants, and demonstrate that careful data curation and training strategies can yield strong performance even at small scales (1B and 3B). Additionally, we introduce two specialized variants: MM1.5-Video, designed for video understanding, and MM1.5-UI, tailored for mobile UI understanding. Through extensive empirical studies and ablations, we provide detailed insights into the training processes and decisions that inform our final designs, offering valuable guidance for future research in MLLM development.

Tongda Xu, Qian Zhang, Yanghao Li et al.

We propose conditional perceptual quality, an extension of the perceptual quality defined in \citet{blau2018perception}, by conditioning it on user defined information. Specifically, we extend the original perceptual quality $d(p_{X},p_{\hat{X}})$ to the conditional perceptual quality $d(p_{X|Y},p_{\hat{X}|Y})$, where $X$ is the original image, $\hat{X}$ is the reconstructed, $Y$ is side information defined by user and $d(.,.)$ is divergence. We show that conditional perceptual quality has similar theoretical properties as rate-distortion-perception trade-off \citep{blau2019rethinking}. Based on these theoretical results, we propose an optimal framework for conditional perceptual quality preserving compression. Experimental results show that our codec successfully maintains high perceptual quality and semantic quality at all bitrate. Besides, by providing a lowerbound of common randomness required, we settle the previous arguments on whether randomness should be incorporated into generator for (conditional) perceptual quality compression. The source code is provided in supplementary material.

Shanzhi Yin, Tongda Xu, Yongsheng Liang et al.

With neural networks growing deeper and feature maps growing larger, limited communication bandwidth with external memory (or DRAM) and power constraints become a bottleneck in implementing network inference on mobile and edge devices. In this paper, we propose an end-to-end differentiable bandwidth efficient neural inference method with the activation compressed by neural data compression method. Specifically, we propose a transform-quantization-entropy coding pipeline for activation compression with symmetric exponential Golomb coding and a data-dependent Gaussian entropy model for arithmetic coding. Optimized with existing model quantization methods, low-level task of image compression can achieve up to 19x bandwidth reduction with 6.21x energy saving.

Tom Gunter, Zirui Wang, Chong Wang et al.

We present foundation language models developed to power Apple Intelligence features, including a ~3 billion parameter model designed to run efficiently on devices and a large server-based language model designed for Private Cloud Compute. These models are designed to perform a wide range of tasks efficiently, accurately, and responsibly. This report describes the model architecture, the data used to train the model, the training process, how the models are optimized for inference, and the evaluation results. We highlight our focus on Responsible AI and how the principles are applied throughout the model development.

Tongda Xu, Ziran Zhu, Dailan He et al.

Idempotence is the stability of image codec to re-compression. At the first glance, it is unrelated to perceptual image compression. However, we find that theoretically: 1) Conditional generative model-based perceptual codec satisfies idempotence; 2) Unconditional generative model with idempotence constraint is equivalent to conditional generative codec. Based on this newfound equivalence, we propose a new paradigm of perceptual image codec by inverting unconditional generative model with idempotence constraints. Our codec is theoretically equivalent to conditional generative codec, and it does not require training new models. Instead, it only requires a pre-trained mean-square-error codec and unconditional generative model. Empirically, we show that our proposed approach outperforms state-of-the-art methods such as HiFiC and ILLM, in terms of Fréchet Inception Distance (FID). The source code is provided in https://github.com/tongdaxu/Idempotence-and-Perceptual-Image-Compression.

MiniCPM Team, Chaojun Xiao, Yuxuan Li et al. · tencent-ai, tsinghua

This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Furthermore, we construct a hybrid reasoning model, MiniCPM4.1, which can be used in both deep reasoning mode and non-reasoning mode. Evaluation results demonstrate that MiniCPM4 and MiniCPM4.1 outperform similar-sized open-source models across benchmarks, with the 8B variants showing significant speed improvements on long sequence understanding and generation.

Runchu Tian, Yanghao Li, Yuepeng Fu et al. · tsinghua

Positional bias in large language models (LLMs) hinders their ability to effectively process long inputs. A prominent example is the "lost in the middle" phenomenon, where LLMs struggle to utilize relevant information situated in the middle of the input. While prior research primarily focuses on single pieces of relevant information, real-world applications often involve multiple relevant information pieces. To bridge this gap, we present LongPiBench, a benchmark designed to assess positional bias involving multiple pieces of relevant information. Thorough experiments are conducted with five commercial and six open-source models. These experiments reveal that while most current models are robust against the "lost in the middle" issue, there exist significant biases related to the spacing of relevant information pieces. These findings highlight the importance of evaluating and reducing positional biases to advance LLM's capabilities.

You Li, Chi Chen, Yanghao Li et al.

Latent visual reasoning aims to mimic human's imagination process by meditating through hidden states of Multimodal Large Language Models. While recognized as a promising paradigm for visual reasoning, the underlying mechanisms driving its effectiveness remain unclear. Motivated to demystify the true source of its efficacy, we investigate the validity of latent reasoning using Causal Mediation Analysis. We model the process as a causal chain: the input as the treatment, the latent tokens as the mediator, and the final answer as the outcome. Our findings uncover two critical disconnections: (a) Input-Latent Disconnect: dramatic perturbations on the input result in negligible changes to the latent tokens, suggesting that latent tokens do not effectively attend to the input sequence. (b) Latent-Answer Disconnect: perturbations on the latent tokens yield minimal impact on the final answer, indicating the limited causal effect latent tokens imposing on the outcome. Furthermore, extensive probing analysis reveals that latent tokens encode limited visual information and exhibit high similarity. Consequently, we challenge the necessity of latent reasoning and propose a straightforward alternative named CapImagine, which teaches the model to explicitly imagine using text. Experiments on vision-centric benchmarks show that CapImagine significantly outperforms complex latent-space baselines, highlighting the superior potential of visual reasoning through explicit imagination.

Weilin Zhao, Zihan Zhou, Zhou Su et al. · tsinghua

Long-sequence processing is a critical capability for modern large language models. However, the self-attention mechanism in the standard Transformer architecture faces severe computational and memory bottlenecks when processing long sequences. While trainable sparse attention methods offer a promising solution, existing approaches such as NSA introduce excessive extra parameters and disrupt the conventional \textit{pretrain-on-short, finetune-on-long} workflow, resulting in slow convergence and difficulty in acceleration. To overcome these limitations, we introduce dense-sparse switchable attention framework, termed as InfLLM-V2. InfLLM-V2 is a trainable sparse attention that seamlessly adapts models from short to long sequences. Specifically, InfLLM-V2 reuses dense attention parameters through parameter-free architecture modification, maintaining consistency between short and long sequence processing. Additionally, InfLLM-V2 ensures computational efficiency across all sequence lengths, by using dense attention for short inputs and smoothly transitioning to sparse attention for long sequences. To achieve practical acceleration, we further introduce an efficient implementation of InfLLM-V2 that significantly reduces the computational overhead. Our experiments on long-context understanding and chain-of-thought reasoning demonstrate that InfLLM-V2 is 4$\times$ faster than dense attention while retaining 98.1% and 99.7% of the performance, respectively. Based on the InfLLM-V2 framework, we have trained and open-sourced MiniCPM4.1 (https://huggingface.co/openbmb/MiniCPM4.1-8B), a hybrid reasoning model, providing a reproducible implementation for the research community.

Yanghao Li, Rui Qian, Bowen Pan et al.

Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam et al.

While today's video recognition systems parse snapshots or short clips accurately, they cannot connect the dots and reason across a longer range of time yet. Most existing video architectures can only process <5 seconds of a video without hitting the computation or memory bottlenecks. In this paper, we propose a new strategy to overcome this challenge. Instead of trying to process more frames at once like most existing methods, we propose to process videos in an online fashion and cache "memory" at each iteration. Through the memory, the model can reference prior context for long-term modeling, with only a marginal cost. Based on this idea, we build MeMViT, a Memory-augmented Multiscale Vision Transformer, that has a temporal support 30x longer than existing models with only 4.5% more compute; traditional methods need >3,000% more compute to do the same. On a wide range of settings, the increased temporal support enabled by MeMViT brings large gains in recognition accuracy consistently. MeMViT obtains state-of-the-art results on the AVA, EPIC-Kitchens-100 action classification, and action anticipation datasets. Code and models are available at https://github.com/facebookresearch/memvit.

Yanghao Li, Chao-Yuan Wu, Haoqi Fan et al.

In this paper, we study Multiscale Vision Transformers (MViTv2) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTv2s' pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViTv2 has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 58.7 boxAP on COCO object detection as well as 86.1% on Kinetics-400 video classification. Code and models are available at https://github.com/facebookresearch/mvit.

Haoqi Fan, Tullie Murrell, Heng Wang et al.

We introduce PyTorchVideo, an open-source deep-learning library that provides a rich set of modular, efficient, and reproducible components for a variety of video understanding tasks, including classification, detection, self-supervised learning, and low-level processing. The library covers a full stack of video understanding tools including multimodal data loading, transformations, and models that reproduce state-of-the-art performance. PyTorchVideo further supports hardware acceleration that enables real-time inference on mobile devices. The library is based on PyTorch and can be used by any training framework; for example, PyTorchLightning, PySlowFast, or Classy Vision. PyTorchVideo is available at https://pytorchvideo.org/

Haoqi Fan, Bo Xiong, Karttikeya Mangalam et al.

We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10x more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast

Yuntao Chen, Chenxia Han, Yanghao Li et al.

Object detection and instance recognition play a central role in many AI applications like autonomous driving, video surveillance and medical image analysis. However, training object detection models on large scale datasets remains computationally expensive and time consuming. This paper presents an efficient and open source object detection framework called SimpleDet which enables the training of state-of-the-art detection models on consumer grade hardware at large scale. SimpleDet supports up-to-date detection models with best practice. SimpleDet also supports distributed training with near linear scaling out of box. Codes, examples and documents of SimpleDet can be found at https://github.com/tusimple/simpledet .

Ruohong Zhang, Bowen Zhang, Yanghao Li et al. · cmu

Chain-of-thought (CoT) reasoning in vision language models (VLMs) is crucial for improving interpretability and trustworthiness. However, current training recipes lack robust CoT reasoning data, relying on datasets dominated by short annotations with minimal rationales. In this work, we show that training VLM on short answers does not generalize well to reasoning tasks that require more detailed responses. To address this, we propose a two-fold approach. First, we distill rationales from GPT-4o model to enrich the training data and fine-tune VLMs, boosting their CoT performance. Second, we apply reinforcement learning to further calibrate reasoning quality. Specifically, we construct positive (correct) and negative (incorrect) pairs of model-generated reasoning chains, by comparing their predictions with annotated short answers. Using this pairwise data, we apply the Direct Preference Optimization algorithm to refine the model's reasoning abilities. Our experiments demonstrate significant improvements in CoT reasoning on benchmark datasets and better generalization to direct answer prediction as well. This work emphasizes the importance of incorporating detailed rationales in training and leveraging reinforcement learning to strengthen the reasoning capabilities of VLMs.

Changxin Liu, Yanghao Li, Yuhao Yi et al.

Distributed learning has become the standard approach for training large-scale machine learning models across private data silos. While distributed learning enhances privacy preservation and training efficiency, it faces critical challenges related to Byzantine robustness and communication reduction. Existing Byzantine-robust and communication-efficient methods rely on full gradient information either at every iteration or at certain iterations with a probability, and they only converge to an unnecessarily large neighborhood around the solution. Motivated by these issues, we propose a novel Byzantine-robust and communication-efficient stochastic distributed learning method that imposes no requirements on batch size and converges to a smaller neighborhood around the optimal solution than all existing methods, aligning with the theoretical lower bound. Our key innovation is leveraging Polyak Momentum to mitigate the noise caused by both biased compressors and stochastic gradients, thus defending against Byzantine workers under information compression. We provide proof of tight complexity bounds for our algorithm in the context of non-convex smooth loss functions, demonstrating that these bounds match the lower bounds in Byzantine-free scenarios. Finally, we validate the practical significance of our algorithm through an extensive series of experiments, benchmarking its performance on both binary classification and image classification tasks.

Yanghao Li, Changxin Liu, Yuhao Yi

In collaborative and distributed learning, Byzantine robustness reflects a major facet of optimization algorithms. Such distributed algorithms are often accompanied by transmitting a large number of parameters, so communication compression is essential for an effective solution. In this paper, we propose Byz-DM21, a novel Byzantine-robust and communication-efficient stochastic distributed learning algorithm. Our key innovation is a novel gradient estimator based on a double-momentum mechanism, integrating recent advancements in error feedback techniques. Using this estimator, we design both standard and accelerated algorithms that eliminate the need for large batch sizes while maintaining robustness against Byzantine workers. We prove that the Byz-DM21 algorithm has a smaller neighborhood size and converges to $\varepsilon$-stationary points in $\mathcal{O}(\varepsilon^{-4})$ iterations. To further enhance efficiency, we introduce a distributed variant called Byz-VR-DM21, which incorporates local variance reduction at each node to progressively eliminate variance from random approximations. We show that Byz-VR-DM21 provably converges to $\varepsilon$-stationary points in $\mathcal{O}(\varepsilon^{-3 })$ iterations. Additionally, we extend our results to the case where the functions satisfy the Polyak-Łojasiewicz condition. Finally, numerical experiments demonstrate the effectiveness of the proposed method.

Ethan Li, Anders Boesen Lindbo Larsen, Chen Zhang et al. · apple-ml, cmu

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Yichen Zhang, Da Peng, Zonghao Guo et al.

A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.

Yanghao Li, Hanzi Mao, Ross Girshick et al.

We explore the plain, non-hierarchical Vision Transformer (ViT) as a backbone network for object detection. This design enables the original ViT architecture to be fine-tuned for object detection without needing to redesign a hierarchical backbone for pre-training. With minimal adaptations for fine-tuning, our plain-backbone detector can achieve competitive results. Surprisingly, we observe: (i) it is sufficient to build a simple feature pyramid from a single-scale feature map (without the common FPN design) and (ii) it is sufficient to use window attention (without shifting) aided with very few cross-window propagation blocks. With plain ViT backbones pre-trained as Masked Autoencoders (MAE), our detector, named ViTDet, can compete with the previous leading methods that were all based on hierarchical backbones, reaching up to 61.3 AP_box on the COCO dataset using only ImageNet-1K pre-training. We hope our study will draw attention to research on plain-backbone detectors. Code for ViTDet is available in Detectron2.

Yanghao Li, Saining Xie, Xinlei Chen et al.

Object detection is a central downstream task used to test if pre-trained network parameters confer benefits, such as improved accuracy or training speed. The complexity of object detection methods can make this benchmarking non-trivial when new architectures, such as Vision Transformer (ViT) models, arrive. These difficulties (e.g., architectural incompatibility, slow training, high memory consumption, unknown training formulae, etc.) have prevented recent studies from benchmarking detection transfer learning with standard ViT models. In this paper, we present training techniques that overcome these challenges, enabling the use of standard ViT models as the backbone of Mask R-CNN. These tools facilitate the primary goal of our study: we compare five ViT initializations, including recent state-of-the-art self-supervised learning methods, supervised initialization, and a strong random initialization baseline. Our results show that recent masking-based unsupervised learning methods may, for the first time, provide convincing transfer learning improvements on COCO, increasing box AP up to 4% (absolute) over supervised and prior self-supervised pre-training methods. Moreover, these masking-based initializations scale better, with the improvement growing as model size increases.

Kaiming He, Xinlei Chen, Saining Xie et al.

This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pre-training and shows promising scaling behavior.

Kristen Grauman, Andrew Westbury, Eugene Byrne et al.

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,670 hours of daily-life activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 931 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception. Project page: https://ego4d-data.org/

Yanghao Li, Tushar Nagarajan, Bo Xiong et al.

We introduce an approach for pre-training egocentric video models using large-scale third-person video datasets. Learning from purely egocentric data is limited by low dataset scale and diversity, while using purely exocentric (third-person) data introduces a large domain mismatch. Our idea is to discover latent signals in third-person video that are predictive of key egocentric-specific properties. Incorporating these signals as knowledge distillation losses during pre-training results in models that benefit from both the scale and diversity of third-person video data, as well as representations that capture salient egocentric properties. Our experiments show that our Ego-Exo framework can be seamlessly integrated into standard video models; it outperforms all baselines when fine-tuned for egocentric activity recognition, achieving state-of-the-art results on Charades-Ego and EPIC-Kitchens-100.

Yanghao Li, Bichuan Guo, Jiangtao Wen et al.

Denoisers trained with synthetic data often fail to cope with the diversity of unknown noises, giving way to methods that can adapt to existing noise without knowing its ground truth. Previous image-based method leads to noise overfitting if directly applied to video denoisers, and has inadequate temporal information management especially in terms of occlusion and lighting variation, which considerably hinders its denoising performance. In this paper, we propose a general framework for video denoising networks that successfully addresses these challenges. A novel twin sampler assembles training data by decoupling inputs from targets without altering semantics, which not only effectively solves the noise overfitting problem, but also generates better occlusion masks efficiently by checking optical flow consistency. An online denoising scheme and a warping loss regularizer are employed for better temporal alignment. Lighting variation is quantified based on the local similarity of aligned frames. Our method consistently outperforms the prior art by 0.6-3.2dB PSNR on multiple noises, datasets and network architectures. State-of-the-art results on reducing model-blind video noises are achieved. Extensive ablation studies are conducted to demonstrate the significance of each technical components.

Sijie Song, Jiaying Liu, Yanghao Li et al.

With the prevalence of RGB-D cameras, multi-modal video data have become more available for human action recognition. One main challenge for this task lies in how to effectively leverage their complementary information. In this work, we propose a Modality Compensation Network (MCN) to explore the relationships of different modalities, and boost the representations for human action recognition. We regard RGB/optical flow videos as source modalities, skeletons as auxiliary modality. Our goal is to extract more discriminative features from source modalities, with the help of auxiliary modality. Built on deep Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM) networks, our model bridges data from source and auxiliary modalities by a modality adaptation block to achieve adaptive representation learning, that the network learns to compensate for the loss of skeletons at test time and even at training time. We explore multiple adaptation schemes to narrow the distance between source and auxiliary modal distributions from different levels, according to the alignment of source and auxiliary data in training. In addition, skeletons are only required in the training phase. Our model is able to improve the recognition performance with source data when testing. Experimental results reveal that MCN outperforms state-of-the-art approaches on four widely-used action recognition benchmarks.

Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer et al.

First-person video naturally brings the use of a physical environment to the forefront, since it shows the camera wearer interacting fluidly in a space based on his intentions. However, current methods largely separate the observed actions from the persistent space itself. We introduce a model for environment affordances that is learned directly from egocentric video. The main idea is to gain a human-centric model of a physical space (such as a kitchen) that captures (1) the primary spatial zones of interaction and (2) the likely activities they support. Our approach decomposes a space into a topological map derived from first-person activity, organizing an ego-video into a series of visits to the different zones. Further, we show how to link zones across multiple related environments (e.g., from videos of multiple kitchens) to obtain a consolidated representation of environment functionality. On EPIC-Kitchens and EGTEA+, we demonstrate our approach for learning scene affordances and anticipating future actions in long-form video.

Yanghao Li, Yuntao Chen, Naiyan Wang et al.

Scale variation is one of the key challenges in object detection. In this work, we first present a controlled experiment to investigate the effect of receptive fields for scale variation in object detection. Based on the findings from the exploration experiments, we propose a novel Trident Network (TridentNet) aiming to generate scale-specific feature maps with a uniform representational power. We construct a parallel multi-branch architecture in which each branch shares the same transformation parameters but with different receptive fields. Then, we adopt a scale-aware training scheme to specialize each branch by sampling object instances of proper scales for training. As a bonus, a fast approximation version of TridentNet could achieve significant improvements without any additional parameters and computational cost compared with the vanilla detector. On the COCO dataset, our TridentNet with ResNet-101 backbone achieves state-of-the-art single-model results of 48.4 mAP. Codes are available at https://git.io/fj5vR.

Yanghao Li, Sijie Song, Yuqi Li et al.

Temporal modeling in videos is a fundamental yet challenging problem in computer vision. In this paper, we propose a novel Temporal Bilinear (TB) model to capture the temporal pairwise feature interactions between adjacent frames. Compared with some existing temporal methods which are limited in linear transformations, our TB model considers explicit quadratic bilinear transformations in the temporal domain for motion evolution and sequential relation modeling. We further leverage the factorized bilinear model in linear complexity and a bottleneck network design to build our TB blocks, which also constrains the parameters and computation cost. We consider two schemes in terms of the incorporation of TB blocks and the original 2D spatial convolutions, namely wide and deep Temporal Bilinear Networks (TBN). Finally, we perform experiments on several widely adopted datasets including Kinetics, UCF101 and HMDB51. The effectiveness of our TBNs is validated by comprehensive ablation analyses and comparisons with various state-of-the-art methods.

Chunhui Liu, Yueyu Hu, Yanghao Li et al.

Despite the fact that many 3D human activity benchmarks being proposed, most existing action datasets focus on the action recognition tasks for the segmented videos. There is a lack of standard large-scale benchmarks, especially for current popular data-hungry deep learning based methods. In this paper, we introduce a new large scale benchmark (PKU-MMD) for continuous multi-modality 3D human action understanding and cover a wide range of complex human activities with well annotated information. PKU-MMD contains 1076 long video sequences in 51 action categories, performed by 66 subjects in three camera views. It contains almost 20,000 action instances and 5.4 million frames in total. Our dataset also provides multi-modality data sources, including RGB, depth, Infrared Radiation and Skeleton. With different modalities, we conduct extensive experiments on our dataset in terms of two scenarios and evaluate different methods by various metrics, including a new proposed evaluation protocol 2D-AP. We believe this large-scale dataset will benefit future researches on action detection for the community.

Yanghao Li, Naiyan Wang, Jiaying Liu et al.

Neural Style Transfer has recently demonstrated very exciting results which catches eyes in both academia and industry. Despite the amazing results, the principle of neural style transfer, especially why the Gram matrices could represent style remains unclear. In this paper, we propose a novel interpretation of neural style transfer by treating it as a domain adaptation problem. Specifically, we theoretically show that matching the Gram matrices of feature maps is equivalent to minimize the Maximum Mean Discrepancy (MMD) with the second order polynomial kernel. Thus, we argue that the essence of neural style transfer is to match the feature distributions between the style images and the generated images. To further support our standpoint, we experiment with several other distribution alignment methods, and achieve appealing results. We believe this novel interpretation connects these two important research fields, and could enlighten future researches.

Yanghao Li, Naiyan Wang, Jiaying Liu et al.

Although Deep Convolutional Neural Networks (CNNs) have liberated their power in various computer vision tasks, the most important components of CNN, convolutional layers and fully connected layers, are still limited to linear transformations. In this paper, we propose a novel Factorized Bilinear (FB) layer to model the pairwise feature interactions by considering the quadratic terms in the transformations. Compared with existing methods that tried to incorporate complex non-linearity structures into CNNs, the factorized parameterization makes our FB layer only require a linear increase of parameters and affordable computational cost. To further reduce the risk of overfitting of the FB layer, a specific remedy called DropFactor is devised during the training process. We also analyze the connection between FB layer and some existing models, and show FB layer is a generalization to them. Finally, we validate the effectiveness of FB layer on several widely adopted datasets including CIFAR-10, CIFAR-100 and ImageNet, and demonstrate superior results compared with various state-of-the-art deep models.

Yanghao Li, Cuiling Lan, Junliang Xing et al.

Human action recognition from well-segmented 3D skeleton data has been intensively studied and has been attracting an increasing attention. Online action detection goes one step further and is more challenging, which identifies the action type and localizes the action positions on the fly from the untrimmed stream data. In this paper, we study the problem of online action detection from streaming skeleton data. We propose a multi-task end-to-end Joint Classification-Regression Recurrent Neural Network to better explore the action type and temporal localization information. By employing a joint classification and regression optimization objective, this network is capable of automatically localizing the start and end points of actions more accurately. Specifically, by leveraging the merits of the deep Long Short-Term Memory (LSTM) subnetwork, the proposed model automatically captures the complex long-range temporal dynamics, which naturally avoids the typical sliding window design and thus ensures high computational efficiency. Furthermore, the subtask of regression optimization provides the ability to forecast the action prior to its occurrence. To evaluate our proposed model, we build a large streaming video dataset with annotations. Experimental results on our dataset and the public G3D dataset both demonstrate very promising performance of our scheme.

Wentao Zhu, Cuiling Lan, Junliang Xing et al.

Skeleton based action recognition distinguishes human actions using the trajectories of skeleton joints, which provide a very good representation for describing actions. Considering that recurrent neural networks (RNNs) with Long Short-Term Memory (LSTM) can learn feature representations and model long-term temporal dependencies automatically, we propose an end-to-end fully connected deep LSTM network for skeleton based action recognition. Inspired by the observation that the co-occurrences of the joints intrinsically characterize human actions, we take the skeleton as the input at each time slot and introduce a novel regularization scheme to learn the co-occurrence features of skeleton joints. To train the deep LSTM network effectively, we propose a new dropout algorithm which simultaneously operates on the gates, cells, and output responses of the LSTM neurons. Experimental results on three human action recognition datasets consistently demonstrate the effectiveness of the proposed model.

Yanghao Li, Naiyan Wang, Jianping Shi et al.

Deep neural networks (DNN) have shown unprecedented success in various computer vision applications such as image classification and object detection. However, it is still a common annoyance during the training phase, that one has to prepare at least thousands of labeled images to fine-tune a network to a specific domain. Recent study (Tommasi et al. 2015) shows that a DNN has strong dependency towards the training dataset, and the learned features cannot be easily transferred to a different but relevant task without fine-tuning. In this paper, we propose a simple yet powerful remedy, called Adaptive Batch Normalization (AdaBN) to increase the generalization ability of a DNN. By modulating the statistics in all Batch Normalization layers across the network, our approach achieves deep adaptation effect for domain adaptation tasks. In contrary to other deep learning domain adaptation methods, our method does not require additional components, and is parameter-free. It archives state-of-the-art performance despite its surprising simplicity. Furthermore, we demonstrate that our method is complementary with other existing methods. Combining AdaBN with existing domain adaptation treatments may further improve model performance.