与人类的视野相比，基于卷积神经网络（CNN）的计算机视觉更容易受到对抗性的噪音。这种差异可能归因于眼睛如何样本视觉输入以及大脑如何通过其背侧和腹侧视觉途径处理视网膜样品，这些途径尚未探索计算机视觉。受到大脑的启发，我们设计了复发性神经网络，包括模拟人类视网膜的输入采样器，它是一个指导下一步位置的背面网络，以及代表视网膜样品的腹网络。组合这些模块，这些模型学会了多一眼图像，每一眼就注意一个明显的部分，并随着时间的推移积累表示形式以识别图像。我们测试了此类模型的稳健性，并在不同水平的对抗噪声上测试，特别关注不同输入采样策略的效果。我们的发现表明，视网膜凹和采样使模型更加可靠，并且在给予更长的时间以更多地看一眼图像时，该模型可能会从攻击中纠正自身。总之，强大的视觉识别可以从三种受脑启发的机制的综合使用中受益：视网膜转化，注意力引导的眼动运动和经常性处理，而不是仅喂食的CNN。

在自然语言处理中，大多数模型都尝试仅仅从文本学习语义表示。学习的表示编码了分布语义，但未能连接到物理世界的任何知识。相比之下，人类通过在感知和行动中接地概念来学习语言，并且大脑编码接地语义进行认知。灵感来自这一概念和最近的愿景 - 语言学习的工作，我们设计了一个用于愿景中的接地语言学习的两流模型。该模型包括基于VGG的视觉流和基于BERT的语言流。这两条流合并到联合代表空间中。通过跨模型对比学习，该模型首先学会与MS Coco DataSet对齐视觉和语言表示。该模型进一步学习通过跨模型注意模块检索具有语言查询的视觉对象，并通过与视觉基因组数据集推断通过双线性操作员通过双线性运算符之间的视觉关系。在培训之后，该模型的语言流是一种独立语言模型，能够在视觉上接地的语义空间中嵌入概念。这种语义空间表现出主要尺寸可与人类直觉和神经生物学知识达到典型。这个语义空间中的单词嵌入是预测人类定义的语义特征规范，并且被隔离成感知的独特簇。此外，视觉接地的语言模型还通过基于图像，文本或其组合的查询来实现基于视觉知识和多模式图像搜索的组成语言理解。

Model quantization is a widely used technique to compress and accelerate deep neural network (DNN) inference. Emergent DNN hardware accelerators begin to support mixed precision (1-8 bits) to further improve the computation efficiency, which raises a great challenge to find the optimal bitwidth for each layer: it requires domain experts to explore the vast design space trading off among accuracy, latency, energy, and model size, which is both timeconsuming and sub-optimal. There are plenty of specialized hardware for neural networks, but little research has been done for specialized neural network optimization for a particular hardware architecture. Conventional quantization algorithm ignores the different hardware architectures and quantizes all the layers in a uniform way. In this paper, we introduce the Hardware-Aware Automated Quantization (HAQ) framework which leverages the reinforcement learning to automatically determine the quantization policy, and we take the hardware accelerator's feedback in the design loop. Rather than relying on proxy signals such as FLOPs and model size, we employ a hardware simulator to generate direct feedback signals (latency and energy) to the RL agent. Compared with conventional methods, our framework is fully automated and can specialize the quantization policy for different neural network architectures and hardware architectures. Our framework effectively reduced the latency by 1.4-1.95× and the energy consumption by 1.9× with negligible loss of accuracy compared with the fixed bitwidth (8 bits) quantization. Our framework reveals that the optimal policies on different hardware architectures (i.e., edge and cloud architectures) under different resource constraints (i.e., latency, energy and model size) are drastically different. We interpreted the implication of different quantization policies, which offer insights for both neural network architecture design and hardware architecture design. * indicates equal contributions. 68 69 70 71 72 73 25 44 63 82 101 120 MobileNets (fixed 8-bit quantization) MobileNets (our flexible-bit quantization) Latency (ms) Top-1 Accuracy (%) 1MB 2MB 3MB Model Size:Figure 1: We need mixed precision for different layers. We quantize MobileNets [12] to different number of bits (both weights and activations), and it lies on a better pareto curve (yellow) than fixed bit quantization (blue). The reason is that different layers have different redundancy and have different arithmetic intensity (OPs/byte) on the hardware, which advocates for using mixed precision for different layers.

Masked image modeling (MIM) performs strongly in pre-training large vision Transformers (ViTs). However, small models that are critical for real-world applications cannot or only marginally benefit from this pre-training approach. In this paper, we explore distillation techniques to transfer the success of large MIM-based pre-trained models to smaller ones. We systematically study different options in the distillation framework, including distilling targets, losses, input, network regularization, sequential distillation, etc, revealing that: 1) Distilling token relations is more effective than CLS token- and feature-based distillation; 2) An intermediate layer of the teacher network as target perform better than that using the last layer when the depth of the student mismatches that of the teacher; 3) Weak regularization is preferred; etc. With these findings, we achieve significant fine-tuning accuracy improvements over the scratch MIM pre-training on ImageNet-1K classification, using all the ViT-Tiny, ViT-Small, and ViT-base models, with +4.2%/+2.4%/+1.4% gains, respectively. Our TinyMIM model of base size achieves 52.2 mIoU in AE20K semantic segmentation, which is +4.1 higher than the MAE baseline. Our TinyMIM model of tiny size achieves 79.6% top-1 accuracy on ImageNet-1K image classification, which sets a new record for small vision models of the same size and computation budget. This strong performance suggests an alternative way for developing small vision Transformer models, that is, by exploring better training methods rather than introducing inductive biases into architectures as in most previous works. Code is available at https://github.com/OliverRensu/TinyMIM.

Few Shot Instance Segmentation (FSIS) requires models to detect and segment novel classes with limited several support examples. In this work, we explore a simple yet unified solution for FSIS as well as its incremental variants, and introduce a new framework named Reference Twice (RefT) to fully explore the relationship between support/query features based on a Transformer-like framework. Our key insights are two folds: Firstly, with the aid of support masks, we can generate dynamic class centers more appropriately to re-weight query features. Secondly, we find that support object queries have already encoded key factors after base training. In this way, the query features can be enhanced twice from two aspects, i.e., feature-level and instance-level. In particular, we firstly design a mask-based dynamic weighting module to enhance support features and then propose to link object queries for better calibration via cross-attention. After the above steps, the novel classes can be improved significantly over our strong baseline. Additionally, our new framework can be easily extended to incremental FSIS with minor modification. When benchmarking results on the COCO dataset for FSIS, gFSIS, and iFSIS settings, our method achieves a competitive performance compared to existing approaches across different shots, e.g., we boost nAP by noticeable +8.2/+9.4 over the current state-of-the-art FSIS method for 10/30-shot. We further demonstrate the superiority of our approach on Few Shot Object Detection. Code and model will be available.

We present Muse, a text-to-image Transformer model that achieves state-of-the-art image generation performance while being significantly more efficient than diffusion or autoregressive models. Muse is trained on a masked modeling task in discrete token space: given the text embedding extracted from a pre-trained large language model (LLM), Muse is trained to predict randomly masked image tokens. Compared to pixel-space diffusion models, such as Imagen and DALL-E 2, Muse is significantly more efficient due to the use of discrete tokens and requiring fewer sampling iterations; compared to autoregressive models, such as Parti, Muse is more efficient due to the use of parallel decoding. The use of a pre-trained LLM enables fine-grained language understanding, translating to high-fidelity image generation and the understanding of visual concepts such as objects, their spatial relationships, pose, cardinality etc. Our 900M parameter model achieves a new SOTA on CC3M, with an FID score of 6.06. The Muse 3B parameter model achieves an FID of 7.88 on zero-shot COCO evaluation, along with a CLIP score of 0.32. Muse also directly enables a number of image editing applications without the need to fine-tune or invert the model: inpainting, outpainting, and mask-free editing. More results are available at https://muse-model.github.io

Learning the underlying distribution of molecular graphs and generating high-fidelity samples is a fundamental research problem in drug discovery and material science. However, accurately modeling distribution and rapidly generating novel molecular graphs remain crucial and challenging goals. To accomplish these goals, we propose a novel Conditional Diffusion model based on discrete Graph Structures (CDGS) for molecular graph generation. Specifically, we construct a forward graph diffusion process on both graph structures and inherent features through stochastic differential equations (SDE) and derive discrete graph structures as the condition for reverse generative processes. We present a specialized hybrid graph noise prediction model that extracts the global context and the local node-edge dependency from intermediate graph states. We further utilize ordinary differential equation (ODE) solvers for efficient graph sampling, based on the semi-linear structure of the probability flow ODE. Experiments on diverse datasets validate the effectiveness of our framework. Particularly, the proposed method still generates high-quality molecular graphs in a limited number of steps.

Deep neural networks are vulnerable to adversarial attacks. In this paper, we take the role of investigators who want to trace the attack and identify the source, that is, the particular model which the adversarial examples are generated from. Techniques derived would aid forensic investigation of attack incidents and serve as deterrence to potential attacks. We consider the buyers-seller setting where a machine learning model is to be distributed to various buyers and each buyer receives a slightly different copy with same functionality. A malicious buyer generates adversarial examples from a particular copy $\mathcal{M}_i$ and uses them to attack other copies. From these adversarial examples, the investigator wants to identify the source $\mathcal{M}_i$. To address this problem, we propose a two-stage separate-and-trace framework. The model separation stage generates multiple copies of a model for a same classification task. This process injects unique characteristics into each copy so that adversarial examples generated have distinct and traceable features. We give a parallel structure which embeds a ``tracer'' in each copy, and a noise-sensitive training loss to achieve this goal. The tracing stage takes in adversarial examples and a few candidate models, and identifies the likely source. Based on the unique features induced by the noise-sensitive loss function, we could effectively trace the potential adversarial copy by considering the output logits from each tracer. Empirical results show that it is possible to trace the origin of the adversarial example and the mechanism can be applied to a wide range of architectures and datasets.

This paper presents a novel framework for planning in unknown and occluded urban spaces. We specifically focus on turns and intersections where occlusions significantly impact navigability. Our approach uses an inpainting model to fill in a sparse, occluded, semantic lidar point cloud and plans dynamically feasible paths for a vehicle to traverse through the open and inpainted spaces. We demonstrate our approach using a car's lidar data with real-time occlusions, and show that by inpainting occluded areas, we can plan longer paths, with more turn options compared to without inpainting; in addition, our approach more closely follows paths derived from a planner with no occlusions (called the ground truth) compared to other state of the art approaches.

Video representation learning has been successful in video-text pre-training for zero-shot transfer, where each sentence is trained to be close to the paired video clips in a common feature space. For long videos, given a paragraph of description where the sentences describe different segments of the video, by matching all sentence-clip pairs, the paragraph and the full video are aligned implicitly. However, such unit-level similarity measure may ignore the global temporal context over a long time span, which inevitably limits the generalization ability. In this paper, we propose a contrastive learning framework TempCLR to compare the full video and the paragraph explicitly. As the video/paragraph is formulated as a sequence of clips/sentences, under the constraint of their temporal order, we use dynamic time warping to compute the minimum cumulative cost over sentence-clip pairs as the sequence-level distance. To explore the temporal dynamics, we break the consistency of temporal order by shuffling the video clips or sentences according to the temporal granularity. In this way, we obtain the representations for clips/sentences, which perceive the temporal information and thus facilitate the sequence alignment. In addition to pre-training on the video and paragraph, our approach can also generalize on the matching between different video instances. We evaluate our approach on video retrieval, action step localization, and few-shot action recognition, and achieve consistent performance gain over all three tasks. Detailed ablation studies are provided to justify the approach design.