7 Papers & Radios | 推理速度比Stable Diffusion快2倍;视觉Transformer统一图像文本
本周论文包括 MIT 造出薄如纸的太阳能电池板;推理速度比 Stable Diffusion 快 2 倍的 Muse 模型等研究。
目录:
- One Model to Edit Them All: Free-Form Text-Driven Image Manipulation with Semantic Modulations
- Printed Organic Photovoltaic Modules on Transferable Ultra-thin Substrates as Additive Power Sources
- A Survey of Deep Learning for Mathematical Reasoning
- Muse: Text-To-Image Generation via Masked Generative Transformers
- Positive-Incentive Noise
- ABPN: Adaptive Blend Pyramid Network for Real-Time Local Retouching of Ultra High-Resolution Photo
- Image-and-Language Understanding from Pixels Only
- ArXiv Weekly Radiostation:NLP、CV、ML 更多精选论文(附音频)
论文 1:One Model to Edit Them All: Free-Form Text-Driven Image Manipulation with Semantic Modulations
- 作者:Yiming Zhu 、 Hongyu Liu 等
- 论文地址:https://arxiv.org/pdf/2210.07883.pdf
不同于经典的 StyleCLIP 模型,我们的模型无需对每个文本单独训练一个模型,一个模型就可以响应多个文本从而对图像做有效的编辑,所以我们的模型成为 FFCLIP-Free Form Text-Driven Image Manipulation。同时我们的模型在经典的教堂,人脸以及汽车数据集上都取得了非常不错的效果。
论文 2:Printed Organic Photovoltaic Modules on Transferable Ultra-thin Substrates as Additive Power Sources
- 作者:Mayuran Saravanapavanantham、Jeremiah Mwaura 等
- 论文地址:https://onlinelibrary.wiley.com/doi/10.1002/smtd.202200940
MIT 的研究人员制造出了比人类头发还薄的太阳能电池板,该电池板每公斤提供的能量是目前玻璃和硅基太阳能电池板的 18 倍。这些太阳能电池板的重量只有传统光电电池的百分之一。
这种超薄太阳能板也可以安装到船帆、无人机机翼和帐篷上。它们在偏远地区和救灾行动中尤其有用。
论文 3:A Survey of Deep Learning for Mathematical Reasoning
- 作者:Pan Lu、 Liang Qiu 等
- 论文地址:https://arxiv.org/pdf/2212.10535.pdf
具体而言,本文讨论了各种任务和数据集(第 2 节),并研究了神经网络(第 3 节)和预训练语言模型(第 4 节)在数学领域的进展。此外还探讨了大型语言模型的上下文学习在数学推理中的快速发展(第 5 节)。文章进一步分析了现有的基准,发现对多模态和低资源环境的关注较少(第 6.1 节)。基于循证的研究表明,目前的计算能力表征是不充分的,深度学习方法在数学推理方面也是不一致的(第 6.2 节)。随后,作者建议在概括性和鲁棒性、可信推理、从反馈中学习和多模态数学推理方面改进目前的工作(第 7 节)。
论文 4:Muse: Text-To-Image Generation via Masked Generative Transformers
- 作者:Huiwen Chang 、 Han Zhang 等
- 论文地址:https://arxiv.org/pdf/2301.00704v1.pdf
与建立在级联像素空间(pixel-space)扩散模型上的 Imagen (Saharia et al., 2022) 或 Dall-E2 (Ramesh et al., 2022) 相比,Muse 由于使用了离散 token,效率显著提升。与 SOTA 自回归模型 Parti (Yu et al., 2022) 相比,Muse 因使用并行解码而效率更高。
基于在 TPU-v4 上的实验结果,研究者估计 Muse 在推理速度上比 Imagen-3B 或 Parti-3B 模型快 10 倍以上,比 Stable Diffusion v1.4 (Rombach et al., 2022) 快 2 倍。研究者认为:Muse 比 Stable Diffusion 推理速度更快是因为 Stable Diffusion v1.4 中使用了扩散模型,在推理时明显需要更多次迭代。
论文 5:Positive-Incentive Noise
- 作者:李学龙
- 论文地址:https://ieeexplore.ieee.org/document/10003114
恰如图 1 所示,在一个图像智能分类系统中,对图像加入适量的噪声后再训练,识别准确率反而上升了。这给我们带来一点启发:图像中加入一些噪声,而不是去除,再执行图像分类任务,可能效果会更好。只要噪声对目标的影响远小于噪声对背景的影响,产生 “伤敌(背景噪声)一千,自(目标信号)损八百” 的效果就有意义,因为任务追求的是高信噪比。从本质上来说,面对传统分类问题,在特征后随机加上适度的噪声,相当于升高了特征维度,某种意义上说,类似是给特征增加了一个核函数,实际上完成了一种低维空间到高维空间的映射,使数据更可分,从而提高了分类效果。
论文 6:ABPN: Adaptive Blend Pyramid Network for Real-Time Local Retouching of Ultra High-Resolution Photo
- 作者:Biwen Lei 、 Xiefan Guo 等
- 论文地址:https://openaccess.thecvf.com/content/CVPR2022/papers/Lei_ABPN_Adaptive_Blend_Pyramid_Network_for_Real-Time_Local_Retouching_of_CVPR_2022_paper.pdf
如上图所示,网络结构主要由两个部分组成:上下文感知的局部修饰层(LRL)和自适应混合金字塔层(BPL)。其中 LRL 的目的是对降采样后的低分辨率图像进行局部修饰,生成低分辨率的修饰结果图,充分考虑全局的上下文信息以及局部的纹理信息。进一步,BPL 用于将 LRL 中生成的低分辨率结果逐步向上拓展到高分辨率结果。其中,我们设计了一个自适应混合模块(ABM)及其逆向模块(R-ABM),利用中间混合图层 Bi,可实现原图与结果图之间的自适应转换以及向上拓展,展现了强大的可拓展性和细节保真能力。我们在脸部修饰及服饰修饰两个数据集中进行了大量实验,结果表明我们的方法在效果和效率上都大幅度地领先了现有方法。值得一提的是,我们的模型在单卡 P100 上实现了 4K 超高分辨率图像的实时推理。
论文 7:Image-and-Language Understanding from Pixels Only
- 作者:Michael Tschannen、Basil Mustafa 等
- 论文地址:https://arxiv.org/pdf/2212.08045.pdf
本文将对使用纯基于像素的模型进行文本和图像的多模态学习进行探索。该模型是一个单独的视觉 Transformer,它处理视觉输入或文本,或两者一起,所有都呈现为 RGB 图像。所有模态都使用相同的模型参数,包括低级特征处理;也就是说,不存在特定于模态的初始卷积、tokenization 算法或输入嵌入表。该模型仅用一个任务训练:对比学习,正如 CLIP 和 ALIGN 所推广的那样。因此模型被称作 CLIP-Pixels Only(CLIPPO)。
1. Rethinking with Retrieval: Faithful Large Language Model Inference. (from Hongming Zhang, Dan Roth) 2. Understanding Political Polarisation using Language Models: A dataset and method. (from Bhiksha Raj) 3. Towards Table-to-Text Generation with Pretrained Language Model: A Table Structure Understanding and Text Deliberating Approach. (from Hui Xiong) 4. Examining Political Rhetoric with Epistemic Stance Detection. (from Brendan O'Connor) 5. Towards Knowledge-Intensive Text-to-SQL Semantic Parsing with Formulaic Knowledge. (from Min-Yen Kan) 6. Leveraging World Knowledge in Implicit Hate Speech Detection. (from Jessica Lin) 7. Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers. (from Furu Wei) 8. EZInterviewer: To Improve Job Interview Performance with Mock Interview Generator. (from Tao Zhang) 9. Memory Augmented Lookup Dictionary based Language Modeling for Automatic Speech Recognition. (from Yuxuan Wang) 10. Parameter-Efficient Fine-Tuning Design Spaces. (from Diyi Yang)
1. CA$^2$T-Net: Category-Agnostic 3D Articulation Transfer from Single Image. (from Jitendra Malik) 2. Mapping smallholder cashew plantations to inform sustainable tree crop expansion in Benin. (from Vipin Kumar) 3. Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning. (from Trevor Darrell) 4. STEPs: Self-Supervised Key Step Extraction from Unlabeled Procedural Videos. (from Rama Chellappa) 5. Muse: Text-To-Image Generation via Masked Generative Transformers. (from Ming-Hsuan Yang, Kevin Murphy, William T. Freeman) 6. Understanding Imbalanced Semantic Segmentation Through Neural Collapse. (from Xiangyu Zhang, Jiaya Jia) 7. Cross Modal Transformer via Coordinates Encoding for 3D Object Dectection. (from Xiangyu Zhang) 8. Learning Road Scene-level Representations via Semantic Region Prediction. (from Alan Yuille) 9. Learning by Sorting: Self-supervised Learning with Group Ordering Constraints. (from Bernt Schiele) 10. AttEntropy: Segmenting Unknown Objects in Complex Scenes using the Spatial Attention Entropy of Semantic Segmentation Transformers. (from Pascal Fua)
本周 10 篇 ML 精选论文是:
1. Self-organization Preserved Graph Structure Learning with Principle of Relevant Information. (from Philip S. Yu) 2. Modified Query Expansion Through Generative Adversarial Networks for Information Extraction in E-Commerce. (from Altan Cakir) 3. Disentangled Explanations of Neural Network Predictions by Finding Relevant Subspaces. (from Klaus-Robert Müller) 4. L-HYDRA: Multi-Head Physics-Informed Neural Networks. (from George Em Karniadakis) 5. On Transforming Reinforcement Learning by Transformer: The Development Trajectory. (from Dacheng Tao) 6. Boosting Neural Networks to Decompile Optimized Binaries. (from Kai Chen) 7. NeuroExplainer: Fine-Grained Attention Decoding to Uncover Cortical Development Patterns of Preterm Infants. (from Dinggang Shen) 8. A Theory of Human-Like Few-Shot Learning. (from Ming Li) 9. Temporal Difference Learning with Compressed Updates: Error-Feedback meets Reinforcement Learning. (from George J. Pappas) 10. Estimating Latent Population Flows from Aggregated Data via Inversing Multi-Marginal Optimal Transport. (from Hongyuan Zha) © THE END 转载请联系本公众号获得授权 投稿或寻求报道:[email protected]