今天分享的是：2025航空行业报告：360亿方智能航空AI白皮书 报告共计：34页 深度学习与大语言模型专利格局全景扫描：中美领跑，产业整合加速 近日，国际知名知识产权机构Questel发布《2024深度学习专利全景分析》报告，全面揭示了全球人工智能领域，特别是深度学 习与大语言模型技术的专利布局趋势。报告指出，深度学习已成为推动多行业创新的基础性技术，而大语言模型作为其重要分 支，正以前所未有的速度重塑技术生态与市场竞争格局。 深度学习专利持续增长，中美主导全球格局 自2011年深度学习技术初步成型以来，该领域已累计产生超过31万个专利族。尽管增速从早期的近50%放缓至2019–2023年的复 合年增长率16%，但增长势头依然强劲，显示出该技术作为创新基础设施的长期价值。 从国家层面看，中国在专利数量上占据绝对优势，2023年贡献了全球80%的深度学习专利申请。然而，在更具战略意义的国际 专利族指标上，美国表现出更强的全球布局意识，其IPF占比高达35%，仅次于中国的40%，凸显其在国际市场上的商业雄心与 技术影响力。韩国迅速崛起，欧洲与日本则保持稳定参与。 企业竞争格局固化，百度、谷歌、微软领跑 在深度学 ...

【导读】硅谷最新「神仙打架」来了：一个是预言40年全准的AI教父，一个是嘴炮满级、逮谁怼谁的深度学习头号黑粉。也许LeCun离开Meta仅仅是未 来硅谷AI风起云涌的一次预演。 如果要评选近期AI圈的「超级地震」，Yann LeCun被曝将离开Meta绝对算得上大地震。 作为图灵奖得主，LLM的公开反对者，世界模型的苦行僧，开源模型的守护者，X的全职博主。。。 Yann LeCun身上有太多的标签。 最近外媒WSJ给LeCun写了一篇有点「歌颂功德」味道的文章，声称「他一直正确了40年」。 作为一名和Hinton齐名，乃至共事过的AI老兵，LeCun的地位也当得起这些夸赞。 但这篇文章，和LeCun本人却遭到了马库斯的极度不认可。 马库斯的观点是我们都被他骗了十年，不要将他神化！甚至表示： Yann LeCun赖以成名的CNN卷积神经网络成果晚于1988年张伟等人发表的文章。 虽然LeCun在卷积神经网络的发展过程中发挥了作用。但他既非其发明者，也非首个将反向传播算法应用于训练网络权重的研究者（尽管许多 人误以为是如此）。 如果说Yann LeCun是深度学习阵营内部的「反对派」，他反对LLM，但坚持深度学习 ...

Core Insights - ByteDance has recently integrated its technology teams for commercialization, e-commerce, and life services, establishing a new department called China Transaction and Advertising, led by Wang Fengkun, the head of Douyin's life service technology [1] - The restructuring aims to enhance the research and development efficiency of advertising and transaction businesses, specifically e-commerce and life services [1] - The integration only affects the engineering technology teams, contrary to earlier reports suggesting a broader scope of "technical team integration," which has been deemed inaccurate by internal sources [1] Summary by Categories - **Company Structure** - ByteDance has formed a new department focused on transaction and advertising to streamline its operations in e-commerce and life services [1] - The new department will leverage personalized recommendations, deep learning, and large model technologies across various products like Douyin, Toutiao, Xigua Video, and Tomato Novel [1] - **Operational Focus** - The integration is designed to build algorithm strategies and engineering architecture for core revenue-generating businesses, including Douyin e-commerce, life services, and advertising marketing [1]

撰文丨王聪 编辑丨王多鱼 排版丨水成文 在日常生活中，智能手环、智能手表等 可穿戴设备 已经成为了我们健康监测、运动追踪的得力助手。然而，当这些设 备遇到跑步、振动或者姿势变化等运动干扰时，其识别的准确率往往会大打折扣。 2025 年 11 月 17 日，加州大学圣地亚哥分校 徐升 团队、 Joseph Wang 团队合作，在 Nature 新子刊 Nature Sensors 上发表了题为： A noise-tolerant human–machine interface based on deep learning-enhanced wearable sensors 的研究论文，这也是这本新期刊上线的首篇论文。 该研究开发出了一款 基于深度学习的抗运动干扰人机交互界面 ，即使在剧烈运动环境下也能实现精准的手势识别和机 械臂控制。 运动干扰：手势识别的"天敌" 传统的可穿戴惯性测量单元 （IMU） 在虚拟现实、非语言交流、运动康复和机器人控制等领域展现出巨大潜力。 然 而，在真实世界应用中，手势信号常常受到运动 伪影 的干扰。 这些干扰可能来自行走、跑步或乘坐交通工具时的环境活动，也可能源于因重力矢量导致的 ...

相关概述：文本转语音技术，即TTS技术，是一种将文字内容转换为语音输出的技术。它通过计算机程序和算法，将文本信息转化为自然流畅的语音信号， 让用户能够听到文字内容，而无需手动阅读。 产业链：从文本转语音技术行业产业链来看，上游环节为整个行业提供核心硬件、算法框架与基础设施支持，是文本转语音技术实现的前提。中游是文本转 语音技术的核心环节。下游是指应用领域，包括教育、金融、医疗、媒体等领域。 产业链下游：网络视听行业，作为新媒体领域的重要组成部分，是以互联网为传播渠道，专注于提供视频、音频等丰富媒体内容服务的产业链。这一行业涵 盖了内容制作、发行、传播及终端观看等多个关键环节。其中在内容制作方面，越来越多的创作者选择利用文本转语音技术生成语音解说，再通过音频提取 技术将其与视频内容结合，制作出高质量的多媒体内容。近年来，随着互联网的普及，我国网络视听用户规模呈现上升趋势，2024年中国网络视听用户规模 达到10.91亿人，同比上涨1.58%。这意味着网络视听行业具有强大的市场活力和广阔的发展空间。随着网络视听行业的发展，文本转语音技术行业应用将不 断深化。 发展现状：随着人工智能和深度学习技术的快速发展，文本转 ...

Core Insights - The article highlights the inspiring journey of Soumith Chintala, who faced numerous rejections but ultimately created PyTorch, a significant tool in the AI landscape [1][10][22] Group 1: Background and Challenges - Soumith Chintala had a humble beginning, born in Hyderabad, India, and attended a second-tier university [2] - He faced significant challenges, including poor math skills and being rejected by 12 U.S. universities despite scoring 1420 on the GRE [4] - After obtaining a J-1 visa, he struggled to find direction and funding for further studies, leading to a series of rejections from graduate programs [4][5] Group 2: Career Development - Initially, Soumith worked as a test engineer at Amazon before joining Facebook AI Research (FAIR) [4][5] - He started as a low-level engineer but gained recognition after identifying and fixing a critical bug in an ImageNet task [5][6] - Despite initial skepticism about his project, he and his team decided to revamp Torch7, leading to the creation of PyTorch [8][9] Group 3: PyTorch's Impact - PyTorch was officially open-sourced in 2017 and quickly gained traction among top research labs, becoming a mainstream tool for deep learning [10][19] - The framework's flexibility and intuitive design allowed researchers to experiment more freely, leading to a rapid increase in its adoption [17][19] - By 2021, PyTorch's search volume surpassed that of TensorFlow, indicating its growing popularity in the AI community [17][21] Group 4: Community and Legacy - PyTorch has evolved from a niche framework to a foundational tool in AI, with a vast community of developers contributing to its ecosystem [21][26] - Soumith's journey from being rejected multiple times to becoming a respected figure in AI exemplifies resilience and dedication [22][27] - The framework is now integral to many leading AI models, including OpenAI's GPT series and Stability's generative models [26][30]

Group 1 - Samsung announced a total investment of 450 trillion KRW in South Korea over the next five years, focusing on R&D and expanding semiconductor investments, including the construction of a fifth factory in Pyeongtaek, expected to be operational by 2028 [2] - Chipone Integrated Circuit released a new silicon carbide G2.0 technology platform aimed at high efficiency, high power density, and high reliability, applicable in electric vehicles and AI data center power supplies [2] - An international research team led by Aalto University in Finland developed a method to perform complex tensor calculations using single light propagation, marking a significant step towards general AI hardware development [2] Group 2 - China's first high-energy direct geometric non-elastic neutron scattering time-of-flight spectrometer has been successfully accepted, filling a gap in non-elastic neutron scattering above 100 meV in the country [2]

整理 | 褚杏娟 据《金融时报》报道，全球社交媒体巨头 Meta 的首席人工智能科学家、图灵奖得主 Yann LeCun 计划离开公司，并着手创办一家 AI 初创企业。 知情人士透露，这位被誉为"深度学习先驱"的法裔美籍科学家已在内部沟通中表示，将在未来几个月内离职。LeCun 已经开始与潜在投资者进行早 期接触。 无论是 Meta 还是 LeCun 本人，目前都未对此消息发表评论。这位图灵奖得主尚未透露离职时间表，其新公司的具体方向也暂无公开信息。他在 纽约大学的教授职位将保持不变。 LeCun 的离职发生在扎克伯格全力重塑 Meta 人工智能战略的关键时期。扎克伯格已逐步将重心从 LeCun 自 2013 年起领导的 Meta 基础 AI 研究 实验室（FAIR）那种长期研究型工作，转向更快速地推出模型和 AI 产品。他认为 Meta 在竞争中已经落后。 不意外的离开 LeCun 的离开并不令人意外。过去几个月，他对 Meta 内部的一些变化愈发感到不满。据报道，他尤其反感公司新出台的内部研究发表规定，即研 究成果在对外发布前必须经过更严格的内部审查。多名团队成员认为，这一政策限制了学术自由。 Meta ...

Core Insights - The report emphasizes the construction of an AI-based low-frequency index enhancement strategy using deep learning frameworks to build alpha and risk models [3] Market Index Performance - As of November 14, 2025, the Shanghai Composite Index decreased by 0.18%, the Shenzhen Component Index fell by 1.40%, and the CSI 300 Index dropped by 1.08%, indicating a turbulent market with most indices declining [5][8] - Year-to-date performance shows the CSI 300 Index has risen by 17.6%, while the CSI 300 enhanced portfolio has increased by 28.5%, yielding an excess return of 10.9% [20] - The CSI 500 Index has increased by 26.4% year-to-date, with its enhanced portfolio up by 35.0%, resulting in an excess return of 8.6% [25] - The CSI 1000 Index has risen by 25.9% this year, while its enhanced portfolio has surged by 41.7%, achieving an excess return of 15.8% [31] Index Enhancement Fund Performance - For the week ending November 14, 2025, the CSI 300 enhanced fund had an excess return ranging from -1.98% to 1.21%, with a median of 0.24% [12][13] - The CSI 500 enhanced fund's excess return ranged from -0.59% to 2.09%, with a median of 0.32% [12][13] - The CSI 1000 enhanced fund showed an excess return between -0.92% and 1.86%, with a median of 0.03% [12][13] Tracking Portfolio Performance - The report outlines the construction of enhanced portfolios for the CSI 300, CSI 500, and CSI 1000 indices using deep learning frameworks, with weekly rebalancing and a maximum turnover rate of 10% [16] - The alpha signals are derived from a multi-source feature set and stacked multi-model strategies, while risk signals are identified using neural networks [16] CSI 300 Enhanced Portfolio Performance - As of November 14, 2025, the CSI 300 enhanced portfolio has achieved a year-to-date return of 28.5%, compared to the CSI 300's 17.6%, resulting in an excess return of 10.9% [20][21] CSI 500 Enhanced Portfolio Performance - The CSI 500 enhanced portfolio has recorded a year-to-date return of 35.0%, outperforming the CSI 500's 26.4% return, leading to an excess return of 8.6% [25][26] CSI 1000 Enhanced Portfolio Performance - The CSI 1000 enhanced portfolio has increased by 41.7% year-to-date, significantly surpassing the CSI 1000's 25.9% return, resulting in an excess return of 15.8% [31][32]

Core Viewpoint - The upcoming IPO of Easy Vision (Hangzhou) Technology Co., Ltd. on the Sci-Tech Innovation Board reflects the expectations for a new star in the capital market and highlights the progress of China's high-end manufacturing industry in terms of independent innovation and high-quality development [1][8] Group 1: Market Position and Demand - The Chinese automotive manufacturing industry has a high demand for precise and efficient machine vision systems, traditionally dominated by international giants [2] - Easy Vision has emerged as a leader in the machine vision sector for automotive manufacturing, achieving a market share of 13.7% in the automotive manufacturing sector and 22.5% in the complete vehicle manufacturing sector in 2024, surpassing international competitors [2] Group 2: Technological Foundation - Easy Vision has a strong research and development foundation, with a team of 251 members, accounting for 45.89% of total employees, and nearly 90% holding bachelor's degrees or higher [3] - The company has invested over 300 million yuan in R&D, resulting in 22 core technology modules and holding 119 software copyrights and 387 domestic and international patents [3] Group 3: Industry Capabilities - Easy Vision integrates various disciplines such as imaging, algorithms, software, and sensors, positioning itself as a preferred supplier for numerous automotive manufacturers [4] - The company has achieved systematic application of its products across six major automotive manufacturing processes, including stamping and assembly [4] Group 4: International Expansion and New Markets - Easy Vision's products have been exported to global factories of companies like Volvo and Rivian, and it has entered the overseas production bases of domestic leaders like BYD and Chery [5] - The company is expanding into new markets such as rail transit and aviation, with successful product implementations in various urban transit systems [5] Group 5: Financial Performance - Easy Vision has experienced significant revenue growth, with annual revenues of 223 million yuan, 355 million yuan, and 392 million yuan over the past three years, reflecting a compound annual growth rate of 32.59% [6] - The net profit has also shown remarkable growth, with figures of 5.39 million yuan, 57.75 million yuan, and 84.43 million yuan, achieving a compound annual growth rate of 295.66% [6] Group 6: Policy Environment and Future Outlook - Recent government policies have favored strategic emerging industries and digital transformation, providing a supportive environment for machine vision as a core component of smart manufacturing [7] - Easy Vision is positioned to lead the transition from domestic substitution to industry leadership, enhancing the competitiveness of China's automotive supply chain [7] - The upcoming IPO is expected to inject significant capital for innovation, further strengthening the company's international competitiveness and technological independence [7][8]