Tan Yu

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Tan Yu, Zhipeng Jin, Jie Liu et al. · baidu

Existing advertisements click-through rate (CTR) prediction models are mainly dependent on behavior ID features, which are learned based on the historical user-ad interactions. Nevertheless, behavior ID features relying on historical user behaviors are not feasible to describe new ads without previous interactions with users. To overcome the limitations of behavior ID features in modeling new ads, we exploit the visual content in ads to boost the performance of CTR prediction models. Specifically, we map each ad into a set of visual IDs based on its visual content. These visual IDs are further used for generating the visual embedding for enhancing CTR prediction models. We formulate the learning of visual IDs into a supervised quantization problem. Due to a lack of class labels for commercial images in advertisements, we exploit image textual descriptions as the supervision to optimize the image extractor for generating effective visual IDs. Meanwhile, since the hard quantization is non-differentiable, we soften the quantization operation to make it support the end-to-end network training. After mapping each image into visual IDs, we learn the embedding for each visual ID based on the historical user-ad interactions accumulated in the past. Since the visual ID embedding depends only on the visual content, it generalizes well to new ads. Meanwhile, the visual ID embedding complements the ad behavior ID embedding. Thus, it can considerably boost the performance of the CTR prediction models previously relying on behavior ID features for both new ads and ads that have accumulated rich user behaviors. After incorporating the visual ID embedding in the CTR prediction model of Baidu online advertising, the average CTR of ads improves by 1.46%, and the total charge increases by 1.10%.

Tan Yu, Jie Liu, Yi Yang et al. · baidu

The advancement of the communication technology and the popularity of the smart phones foster the booming of video ads. Baidu, as one of the leading search engine companies in the world, receives billions of search queries per day. How to pair the video ads with the user search is the core task of Baidu video advertising. Due to the modality gap, the query-to-video retrieval is much more challenging than traditional query-to-document retrieval and image-to-image search. Traditionally, the query-to-video retrieval is tackled by the query-to-title retrieval, which is not reliable when the quality of tiles are not high. With the rapid progress achieved in computer vision and natural language processing in recent years, content-based search methods becomes promising for the query-to-video retrieval. Benefited from pretraining on large-scale datasets, some visionBERT methods based on cross-modal attention have achieved excellent performance in many vision-language tasks not only in academia but also in industry. Nevertheless, the expensive computation cost of cross-modal attention makes it impractical for large-scale search in industrial applications. In this work, we present a tree-based combo-attention network (TCAN) which has been recently launched in Baidu's dynamic video advertising platform. It provides a practical solution to deploy the heavy cross-modal attention for the large-scale query-to-video search. After launching tree-based combo-attention network, click-through rate gets improved by 2.29\% and conversion rate get improved by 2.63\%.

Yue Zhang, Hongliang Fei, Dingcheng Li et al.

Prompt learning is a new learning paradigm which reformulates downstream tasks as similar pretraining tasks on pretrained models by leveraging textual prompts. Recent works have demonstrated that prompt learning is particularly useful for few-shot learning, where there is limited training data. Depending on the granularity of prompts, those methods can be roughly divided into task-level prompting and instance-level prompting. Task-level prompting methods learn one universal prompt for all input samples, which is efficient but ineffective to capture subtle differences among different classes. Instance-level prompting methods learn a specific prompt for each input, though effective but inefficient. In this work, we develop a novel prototype-based prompt learning method to overcome the above limitations. In particular, we focus on few-shot image recognition tasks on pretrained vision-language models (PVLMs) and develop a method of prompting through prototype (PTP), where we define $K$ image prototypes and $K$ prompt prototypes. In PTP, the image prototype represents a centroid of a certain image cluster in the latent space and a prompt prototype is defined as a soft prompt in the continuous space. The similarity between a query image and an image prototype determines how much this prediction relies on the corresponding prompt prototype. Hence, in PTP, similar images will utilize similar prompting ways. Through extensive experiments on seven real-world benchmarks, we show that PTP is an effective method to leverage the latent knowledge and adaptive to various PVLMs. Moreover, through detailed analysis, we discuss pros and cons for prompt learning and parameter-efficient fine-tuning under the context of few-shot learning.

Rama Akkiraju, Anbang Xu, Deepak Bora et al.

Enterprise chatbots, powered by generative AI, are emerging as key applications to enhance employee productivity. Retrieval Augmented Generation (RAG), Large Language Models (LLMs), and orchestration frameworks like Langchain and Llamaindex are crucial for building these chatbots. However, creating effective enterprise chatbots is challenging and requires meticulous RAG pipeline engineering. This includes fine-tuning embeddings and LLMs, extracting documents from vector databases, rephrasing queries, reranking results, designing prompts, honoring document access controls, providing concise responses, including references, safeguarding personal information, and building orchestration agents. We present a framework for building RAG-based chatbots based on our experience with three NVIDIA chatbots: for IT/HR benefits, financial earnings, and general content. Our contributions are three-fold: introducing the FACTS framework (Freshness, Architectures, Cost, Testing, Security), presenting fifteen RAG pipeline control points, and providing empirical results on accuracy-latency tradeoffs between large and small LLMs. To the best of our knowledge, this is the first paper of its kind that provides a holistic view of the factors as well as solutions for building secure enterprise-grade chatbots."

Tan Yu, Ping Li

The formidable accomplishment of Transformers in natural language processing has motivated the researchers in the computer vision community to build Vision Transformers. Compared with the Convolution Neural Networks (CNN), a Vision Transformer has a larger receptive field which is capable of characterizing the long-range dependencies. Nevertheless, the large receptive field of Vision Transformer is accompanied by the huge computational cost. To boost efficiency, the window-based Vision Transformers emerge. They crop an image into several local windows, and the self-attention is conducted within each window. To bring back the global receptive field, window-based Vision Transformers have devoted a lot of efforts to achieving cross-window communications by developing several sophisticated operations. In this work, we check the necessity of the key design element of Swin Transformer, the shifted window partitioning. We discover that a simple depthwise convolution is sufficient for achieving effective cross-window communications. Specifically, with the existence of the depthwise convolution, the shifted window configuration in Swin Transformer cannot lead to an additional performance improvement. Thus, we degenerate the Swin Transformer to a plain Window-based (Win) Transformer by discarding sophisticated shifted window partitioning. The proposed Win Transformer is conceptually simpler and easier for implementation than Swin Transformer. Meanwhile, our Win Transformer achieves consistently superior performance than Swin Transformer on multiple computer vision tasks, including image recognition, semantic segmentation, and object detection.

Shuo Chen, Tan Yu, Ping Li

Recently, vision architectures based exclusively on multi-layer perceptrons (MLPs) have gained much attention in the computer vision community. MLP-like models achieve competitive performance on a single 2D image classification with less inductive bias without hand-crafted convolution layers. In this work, we explore the effectiveness of MLP-based architecture for the view-based 3D object recognition task. We present an MLP-based architecture termed as Round-Roll MLP (R$^2$-MLP). It extends the spatial-shift MLP backbone by considering the communications between patches from different views. R$^2$-MLP rolls part of the channels along the view dimension and promotes information exchange between neighboring views. We benchmark MLP results on ModelNet10 and ModelNet40 datasets with ablations in various aspects. The experimental results show that, with a conceptually simple structure, our R$^2$-MLP achieves competitive performance compared with existing state-of-the-art methods.

Tan Yu, Anbang Xu, Rama Akkiraju

Overcoming the limited context limitations in early-generation LLMs, retrieval-augmented generation (RAG) has been a reliable solution for context-based answer generation in the past. Recently, the emergence of long-context LLMs allows the models to incorporate much longer text sequences, making RAG less attractive. Recent studies show that long-context LLMs significantly outperform RAG in long-context applications. Unlike the existing works favoring the long-context LLM over RAG, we argue that the extremely long context in LLMs suffers from a diminished focus on relevant information and leads to potential degradation in answer quality. This paper revisits the RAG in long-context answer generation. We propose an order-preserve retrieval-augmented generation (OP-RAG) mechanism, which significantly improves the performance of RAG for long-context question-answer applications. With OP-RAG, as the number of retrieved chunks increases, the answer quality initially rises, and then declines, forming an inverted U-shaped curve. There exist sweet points where OP-RAG could achieve higher answer quality with much less tokens than long-context LLM taking the whole context as input. Extensive experiments on public benchmark demonstrate the superiority of our OP-RAG.

Shu Zhao, Nilesh Ahuja, Tan Yu et al.

Pre-trained vision language models have shown remarkable performance on visual recognition tasks, but they typically assume the availability of complete multimodal inputs during both training and inference. In real-world scenarios, however, modalities may be missing due to privacy constraints, collection difficulties, or resource limitations. While previous approaches have addressed this challenge using prompt learning techniques, they fail to capture the cross-modal relationships necessary for effective multimodal visual recognition and suffer from inevitable computational overhead. In this paper, we introduce MoRA, a parameter-efficient fine-tuning method that explicitly models cross-modal interactions while maintaining modality-specific adaptations. MoRA introduces modality-common parameters between text and vision encoders, enabling bidirectional knowledge transfer. Additionally, combined with the modality-specific parameters, MoRA allows the backbone model to maintain inter-modality interaction and enable intra-modality flexibility. Extensive experiments on standard benchmarks demonstrate that MoRA achieves an average performance improvement in missing-modality scenarios by 5.24% and uses only 25.90% of the inference time compared to the SOTA method while requiring only 0.11% of trainable parameters compared to full fine-tuning.

Shu Zhao, Xiaohan Zou, Tan Yu et al.

Pre-trained large multi-modal models (LMMs) exploit fine-tuning to adapt diverse user applications. Nevertheless, fine-tuning may face challenges due to deactivated sensors (e.g., cameras turned off for privacy or technical issues), yielding modality-incomplete data and leading to inconsistency in training data and the data for inference. Additionally, continuous training leads to catastrophic forgetting, diluting the knowledge in pre-trained LMMs. To overcome these challenges, we introduce a novel task, Continual Missing Modality Learning (CMML), to investigate how models can generalize when data of certain modalities is missing during continual fine-tuning. Our preliminary benchmarks reveal that existing methods suffer from a significant performance drop in CMML, even with the aid of advanced continual learning techniques. Therefore, we devise a framework termed Reconstruct before Query (RebQ). It decomposes prompts into modality-specific ones and breaks them into components stored in pools accessible via a key-query mechanism, which facilitates ParameterEfficient Fine-Tuning and enhances knowledge transferability for subsequent tasks. Meanwhile, our RebQ leverages extensive multi-modal knowledge from pre-trained LMMs to reconstruct the data of missing modality. Comprehensive experiments demonstrate that RebQ effectively reconstructs the missing modality information and retains pre-trained knowledge. Specifically, compared with the baseline, RebQ improves average precision from 20.00 to 50.92 and decreases average forgetting from 75.95 to 8.56. Code and datasets are available on https://github.com/Tree-Shu-Zhao/RebQ.pytorch

Aaditya Shukla, Sidney Knowles, Meenakshi Madugula et al.

Enterprise AI agents must continuously adapt to maintain accuracy, reduce latency, and remain aligned with user needs. We present a practical implementation of a data flywheel in NVInfo AI, NVIDIA's Mixture-of-Experts (MoE) Knowledge Assistant serving over 30,000 employees. By operationalizing a MAPE-driven data flywheel, we built a closed-loop system that systematically addresses failures in retrieval-augmented generation (RAG) pipelines and enables continuous learning. Over a 3-month post-deployment period, we monitored feedback and collected 495 negative samples. Analysis revealed two major failure modes: routing errors (5.25\%) and query rephrasal errors (3.2\%). Using NVIDIA NeMo microservices, we implemented targeted improvements through fine-tuning. For routing, we replaced a Llama 3.1 70B model with a fine-tuned 8B variant, achieving 96\% accuracy, a 10x reduction in model size, and 70\% latency improvement. For query rephrasal, fine-tuning yielded a 3.7\% gain in accuracy and a 40\% latency reduction. Our approach demonstrates how human-in-the-loop (HITL) feedback, when structured within a data flywheel, transforms enterprise AI agents into self-improving systems. Key learnings include approaches to ensure agent robustness despite limited user feedback, navigating privacy constraints, and executing staged rollouts in production. This work offers a repeatable blueprint for building robust, adaptive enterprise AI agents capable of learning from real-world usage at scale.

Le Shen, Qian Qiao, Tan Yu et al.

Deploying massive diffusion models for real-time, infinite-duration, audio-driven avatar generation presents a significant engineering challenge, primarily due to the conflict between computational load and strict latency constraints. Existing approaches often compromise visual fidelity by enforcing strictly unidirectional attention mechanisms or reducing model capacity. To address this problem, we introduce \textbf{SoulX-FlashTalk}, a 14B-parameter framework optimized for high-fidelity real-time streaming. Diverging from conventional unidirectional paradigms, we use a \textbf{Self-correcting Bidirectional Distillation} strategy that retains bidirectional attention within video chunks. This design preserves critical spatiotemporal correlations, significantly enhancing motion coherence and visual detail. To ensure stability during infinite generation, we incorporate a \textbf{Multi-step Retrospective Self-Correction Mechanism}, enabling the model to autonomously recover from accumulated errors and preventing collapse. Furthermore, we engineered a full-stack inference acceleration suite incorporating hybrid sequence parallelism, Parallel VAE, and kernel-level optimizations. Extensive evaluations confirm that SoulX-FlashTalk is the first 14B-scale system to achieve a \textbf{sub-second start-up latency (0.87s)} while reaching a real-time throughput of \textbf{32 FPS}, setting a new standard for high-fidelity interactive digital human synthesis.

Shu Zhao, Tan Yu, Anbang Xu et al.

Reasoning-augmented search agents such as Search-R1, trained via reinforcement learning with verifiable rewards (RLVR), demonstrate remarkable capabilities in multi-step information retrieval from external knowledge sources. These agents address the limitations of their parametric memory by dynamically gathering relevant facts to address complex reasoning tasks. However, existing approaches suffer from a fundamental architectural limitation: they process search queries strictly sequentially, even when handling inherently parallelizable and logically independent comparisons. This sequential bottleneck significantly constrains computational efficiency, particularly for queries that require multiple entity comparisons. To address this critical limitation, we propose ParallelSearch, a novel reinforcement learning framework that empowers large language models (LLMs) to recognize parallelizable query structures and execute multiple search operations concurrently. Our approach introduces dedicated reward functions that incentivize the identification of independent query components while preserving answer accuracy through jointly considering correctness, query decomposition quality, and parallel execution benefits. Comprehensive experiments demonstrate that ParallelSearch outperforms state-of-the-art baselines by an average performance gain of 2.9% across seven question-answering benchmarks. Notably, on parallelizable questions, our method achieves a 12.7% performance improvement while requiring only 69.6% of the LLM calls compared to sequential approaches.

Guoshenghui Zhao, Weijie Zhao, Tan Yu

Diffusion large language models generate text through multi-step denoising, where hallucination signals may emerge throughout the trajectory rather than only in the final output. Existing detectors mainly rely on output uncertainty or coarse trace statistics, which often fail to capture the richer hidden dynamics of D-LLMs. We propose HIVE, a hidden-evidence verification framework that extracts compressed hidden evidence from denoising trajectories, selects informative step-layer evidence, and conditions a verifier language model on the selected evidence through prefix embeddings. HIVE produces both a continuous hallucination score from verifier decision logits and structured verification outputs, including hallucination types, evidence pairs, and short rationales. Across two D-LLMs and three QA benchmarks, HIVE consistently outperforms eight strong baselines and achieves up to 0.9236 AUROC and 0.9537 AUPRC. Ablation studies further confirm the importance of hidden-evidence conditioning, learned evidence selection, two-stream evidence representation, and step-layer embeddings. These results suggest that selected hidden evidence from denoising trajectories provides a stronger and more usable hallucination signal than output-only uncertainty or coarse trace statistics.

Huawei Lin, Yingjie Lao, Tong Geng et al.

Large Language Models (LLMs) are vulnerable to attacks like prompt injection, backdoor attacks, and adversarial attacks, which manipulate prompts or models to generate harmful outputs. In this paper, departing from traditional deep learning attack paradigms, we explore their intrinsic relationship and collectively term them Prompt Trigger Attacks (PTA). This raises a key question: Can we determine if a prompt is benign or poisoned? To address this, we propose UniGuardian, the first unified defense mechanism designed to detect prompt injection, backdoor attacks, and adversarial attacks in LLMs. Additionally, we introduce a single-forward strategy to optimize the detection pipeline, enabling simultaneous attack detection and text generation within a single forward pass. Our experiments confirm that UniGuardian accurately and efficiently identifies malicious prompts in LLMs.

Shu Zhao, Tan Yu, Anbang Xu

Reasoning-augmented search agents, such as Search-R1, are trained to reason, search, and generate the final answer iteratively. Nevertheless, due to their limited capabilities in reasoning and search, their performance on multi-hop QA benchmarks remains far from satisfactory. To handle complex or compound queries, we train an LLM-based search agent with the native capability of query expansion through reinforcement learning. In each turn, our search agent proposes several query variants, which are searched simultaneously to cover more relevant information. Meanwhile, given limited post-training data and computing resources, it is very challenging for a search agent to master multiple tasks, including query generation, retrieved information understanding, and answer generation. Therefore, we propose incorporating a pre-trained squeezer model that helps the search agent understand the retrieved documents, allowing the search agent to focus on query generation for high retrieval recall. With the assistance of the squeezer model, we discover that even a small-scale 3B LLM can demonstrate a strong capability of query expansion and achieve state-of-the-art accuracy on the multi-hop QA benchmarks. To be specific, our experiments across seven question-answering benchmarks demonstrate that our method, named ExpandSearch, achieves an average improvement of 4.4% compared to state-of-the-art baselines, with strong gains on multi-hop reasoning tasks requiring diverse evidence aggregation.

Shu Zhao, Tan Yu, Xiaoshuai Hao et al.

Deep hashing has been widely used for large-scale approximate nearest neighbor search due to its storage and search efficiency. However, existing deep hashing methods predominantly rely on abundant training data, leaving the more challenging scenario of low-resource adaptation for deep hashing relatively underexplored. This setting involves adapting pre-trained models to downstream tasks with only an extremely small number of training samples available. Our preliminary benchmarks reveal that current methods suffer significant performance degradation due to the distribution shift caused by limited training samples. To address these challenges, we introduce Class-Calibration LoRA (CLoRA), a novel plug-and-play approach that dynamically constructs low-rank adaptation matrices by leveraging class-level textual knowledge embeddings. CLoRA effectively incorporates prior class knowledge as anchors, enabling parameter-efficient fine-tuning while maintaining the original data distribution. Furthermore, we propose Knowledge-Guided Discrete Optimization (KIDDO), a framework to utilize class knowledge to compensate for the scarcity of visual information and enhance the discriminability of hash codes. Extensive experiments demonstrate that our proposed method, Knowledge- Anchored Low-Resource Adaptation Hashing (KALAHash), significantly boosts retrieval performance and achieves a 4x data efficiency in low-resource scenarios.

Fangyu Du, Taiqing Li, Ziwei Zhang et al.

Audio-driven portrait animation aims to synthesize realistic and natural talking head videos from an input audio signal and a single reference image. While existing methods achieve high-quality results by leveraging high-dimensional intermediate representations and explicitly modeling motion dynamics, their computational complexity renders them unsuitable for real-time deployment. Real-time inference imposes stringent latency and memory constraints, often necessitating the use of highly compressed latent representations. However, operating in such compact spaces hinders the preservation of fine-grained spatiotemporal details, thereby complicating audio-visual synchronization RAP (Real-time Audio-driven Portrait animation), a unified framework for generating high-quality talking portraits under real-time constraints. Specifically, RAP introduces a hybrid attention mechanism for fine-grained audio control, and a static-dynamic training-inference paradigm that avoids explicit motion supervision. Through these techniques, RAP achieves precise audio-driven control, mitigates long-term temporal drift, and maintains high visual fidelity. Extensive experiments demonstrate that RAP achieves state-of-the-art performance while operating under real-time constraints.

Dingcheng Zhen, Qian Qiao, Xu Zheng et al.

We introduce TransDiff, the first image generation model that marries Autoregressive (AR) Transformer with diffusion models. In this joint modeling framework, TransDiff encodes labels and images into high-level semantic features and employs a diffusion model to estimate the distribution of image samples. On the ImageNet 256x256 benchmark, TransDiff significantly outperforms other image generation models based on standalone AR Transformer or diffusion models. Specifically, TransDiff achieves a Frechet Inception Distance (FID) of 1.61 and an Inception Score (IS) of 293.4, and further provides x2 faster inference latency compared to state-of-the-art methods based on AR Transformer and x112 faster inference compared to diffusion-only models. Furthermore, building on the TransDiff model, we introduce a novel image generation paradigm called Multi-Reference Autoregression (MRAR), which performs autoregressive generation by predicting the next image. MRAR enables the model to reference multiple previously generated images, thereby facilitating the learning of more diverse representations and improving the quality of generated images in subsequent iterations. By applying MRAR, the performance of TransDiff is improved, with the FID reduced from 1.61 to 1.42. We expect TransDiff to open up a new frontier in the field of image generation.

Tan Yu, Gangming Zhao, Ping Li et al.

Vision Transformers achieved outstanding performance in many computer vision tasks. Early Vision Transformers such as ViT and DeiT adopt global self-attention, which is computationally expensive when the number of patches is large. To improve efficiency, recent Vision Transformers adopt local self-attention mechanisms, where self-attention is computed within local windows. Despite the fact that window-based local self-attention significantly boosts efficiency, it fails to capture the relationships between distant but similar patches in the image plane. To overcome this limitation of image-space local attention, in this paper, we further exploit the locality of patches in the feature space. We group the patches into multiple clusters using their features, and self-attention is computed within every cluster. Such feature-space local attention effectively captures the connections between patches across different local windows but still relevant. We propose a Bilateral lOcal Attention vision Transformer (BOAT), which integrates feature-space local attention with image-space local attention. We further integrate BOAT with both Swin and CSWin models, and extensive experiments on several benchmark datasets demonstrate that our BOAT-CSWin model clearly and consistently outperforms existing state-of-the-art CNN models and vision Transformers.

Shuo Chen, Tan Yu, Ping Li

Inspired by the great success achieved by CNN in image recognition, view-based methods applied CNNs to model the projected views for 3D object understanding and achieved excellent performance. Nevertheless, multi-view CNN models cannot model the communications between patches from different views, limiting its effectiveness in 3D object recognition. Inspired by the recent success gained by vision Transformer in image recognition, we propose a Multi-view Vision Transformer (MVT) for 3D object recognition. Since each patch feature in a Transformer block has a global reception field, it naturally achieves communications between patches from different views. Meanwhile, it takes much less inductive bias compared with its CNN counterparts. Considering both effectiveness and efficiency, we develop a global-local structure for our MVT. Our experiments on two public benchmarks, ModelNet40 and ModelNet10, demonstrate the competitive performance of our MVT.

Tan Yu, Xu Li, Yunfeng Cai et al.

Recently, MLP-based vision backbones emerge. MLP-based vision architectures with less inductive bias achieve competitive performance in image recognition compared with CNNs and vision Transformers. Among them, spatial-shift MLP (S$^2$-MLP), adopting the straightforward spatial-shift operation, achieves better performance than the pioneering works including MLP-mixer and ResMLP. More recently, using smaller patches with a pyramid structure, Vision Permutator (ViP) and Global Filter Network (GFNet) achieve better performance than S$^2$-MLP. In this paper, we improve the S$^2$-MLP vision backbone. We expand the feature map along the channel dimension and split the expanded feature map into several parts. We conduct different spatial-shift operations on split parts. Meanwhile, we exploit the split-attention operation to fuse these split parts. Moreover, like the counterparts, we adopt smaller-scale patches and use a pyramid structure for boosting the image recognition accuracy. We term the improved spatial-shift MLP vision backbone as S$^2$-MLPv2. Using 55M parameters, our medium-scale model, S$^2$-MLPv2-Medium achieves an $83.6\%$ top-1 accuracy on the ImageNet-1K benchmark using $224\times 224$ images without self-attention and external training data.

Tan Yu, Xu Li, Yunfeng Cai et al.

In the past decade, we have witnessed rapid progress in the machine vision backbone. By introducing the inductive bias from the image processing, convolution neural network (CNN) has achieved excellent performance in numerous computer vision tasks and has been established as \emph{de facto} backbone. In recent years, inspired by the great success achieved by Transformer in NLP tasks, vision Transformer models emerge. Using much less inductive bias, they have achieved promising performance in computer vision tasks compared with their CNN counterparts. More recently, researchers investigate using the pure-MLP architecture to build the vision backbone to further reduce the inductive bias, achieving good performance. The pure-MLP backbone is built upon channel-mixing MLPs to fuse the channels and token-mixing MLPs for communications between patches. In this paper, we re-think the design of the token-mixing MLP. We discover that token-mixing MLPs in existing MLP-based backbones are spatial-specific, and thus it is sensitive to spatial translation. Meanwhile, the channel-agnostic property of the existing token-mixing MLPs limits their capability in mixing tokens. To overcome those limitations, we propose an improved structure termed as Circulant Channel-Specific (CCS) token-mixing MLP, which is spatial-invariant and channel-specific. It takes fewer parameters but achieves higher classification accuracy on ImageNet1K benchmark.

Tan Yu, Xu Li, Yunfeng Cai et al.

Recently, visual Transformer (ViT) and its following works abandon the convolution and exploit the self-attention operation, attaining a comparable or even higher accuracy than CNNs. More recently, MLP-Mixer abandons both the convolution and the self-attention operation, proposing an architecture containing only MLP layers. To achieve cross-patch communications, it devises an additional token-mixing MLP besides the channel-mixing MLP. It achieves promising results when training on an extremely large-scale dataset. But it cannot achieve as outstanding performance as its CNN and ViT counterparts when training on medium-scale datasets such as ImageNet1K and ImageNet21K. The performance drop of MLP-Mixer motivates us to rethink the token-mixing MLP. We discover that the token-mixing MLP is a variant of the depthwise convolution with a global reception field and spatial-specific configuration. But the global reception field and the spatial-specific property make token-mixing MLP prone to over-fitting. In this paper, we propose a novel pure MLP architecture, spatial-shift MLP (S$^2$-MLP). Different from MLP-Mixer, our S$^2$-MLP only contains channel-mixing MLP. We utilize a spatial-shift operation for communications between patches. It has a local reception field and is spatial-agnostic. It is parameter-free and efficient for computation. The proposed S$^2$-MLP attains higher recognition accuracy than MLP-Mixer when training on ImageNet-1K dataset. Meanwhile, S$^2$-MLP accomplishes as excellent performance as ViT on ImageNet-1K dataset with considerably simpler architecture and fewer FLOPs and parameters.