LeCun连续转发,新作VISReg攻克JEPA世界模型「表征坍塌」核心难题

marsbitPublished on 2026-07-28Last updated on 2026-07-28

Abstract

近日,图灵奖得主Yann LeCun连续转发并高度认可了一项名为VISReg的自监督学习新工作。该工作旨在攻克JEPA世界模型中长期存在的“表征坍塌”核心难题。表征坍塌是指模型倾向于将不同输入映射到相同或极少数向量上,导致学到的表征缺乏判别力。 传统方法多依赖复杂的启发式技巧来抑制坍塌,但训练脆弱且可解释性差。VISReg的解决思路是将正则化目标解耦为独立的“尺度”和“形状”两部分。尺度部分通过方差项防止表征幅值塌缩,并在塌缩时能保持稳定的梯度;形状部分则通过带停止梯度的归一化和切片Wasserstein距离,将表征的分布形状对齐到标准高斯分布,从而完整地约束高阶统计特性。 这种方法实现了尺度和形状的优化互不干扰,有效解决了前代方法SIGReg在塌缩时梯度消失、以及尺度与形状耦合的问题。VISReg实现简洁,计算复杂度低,且易于扩展到大规模训练。 实验表明,在未使用任何启发式技巧的情况下,VISReg在15个数据集(涵盖域内、分布外、医疗、天文等多领域)的综合评估中,表现优于或媲美DINO、MAE等7种主流方法。特别是在分布外泛化任务上,仅使用约十分之一的数据量,其性能便追平了使用十倍数据训练的DINOv2。此外,在迁移微调、低质量数据鲁棒性以及生成引导等任务上也显示出优势。 VISReg为自监督学习提供了一种更稳定、高效且泛化能力强的正则化方案,为构建更强大的JEPA世界模型奠定了基础。

JEPA世界模型的底层是Yann LeCun自2017年起持续倡导的自监督学习(Self-Supervised Learning, SSL)。

SSL 无需人工标注即可从海量数据中学习通用表征,但普遍面临一个核心难题——表征坍塌(representation collapse):模型倾向于把不同输入映射到相同或极少数几个向量上,看似完成了训练,实则未学到有判别力的表征。

为抑制坍塌,主流方法大多依赖一系列启发式技巧(EMA、教师-学生网络、停止梯度、冻结层等)。这些技巧使训练变得脆弱、难以调参,也削弱了方法的可解释性与可扩展性。

另一条路线是通过正则项直接约束表征分布。

LeCun团队提出的VICReg将学习目标拆为方差、不变性、协方差三项,用协方差约束各维度之间的相关性;但协方差仅刻画二阶统计量,无法区分「均值、方差相同,而分布形状迥异」的两种表征。

其后提出的SIGReg基于Cramér–Wold定理,用sketching技术将整个嵌入分布对齐到标准高斯,从而约束完整的分布形状。

然而SIGReg仍存在两个关键缺陷:

  • 坍塌时梯度消失:当表征开始坍塌时,SIGReg的梯度随之衰减——坍塌越严重、修正信号越弱,模型难以自行恢复;
  • 尺度与形状耦合:未将「幅度大小(尺度)」与「分布形态(形状)」两个独立属性分离,二者在优化中相互干扰,导致在长尾、低质量、低秩数据上适配性较差。

也就是说,在模型最需要梯度信号来逃离坍塌状态时,SIGReg的梯度恰恰趋近于消失。

这正是VISReg要解决的核心问题。

近日,自监督学习新工作VISReg(Variance-Invariance-Sketching Regularization)获图灵奖得主Yann LeCun连续转发并给予高度认可——他在转发时评价道「VICReg begat SIGReg which begat VISReg」(VICReg孕育了SIGReg,SIGReg又孕育了VISReg),一句话点明了这条正则化路线的技术传承。

能获得LeCun如此认可,VISReg究竟强在哪里?

答案在于:它精准命中了LeCun长期押注的JEPA世界模型的核心难题——表征坍塌(representation collapse)。

论文链接:https://arxiv.org/abs/2606.02572

代码 / 预训练权重:https://github.com/HaiyuWu/visreg

项目主页:https://haiyuwu.github.io/visreg/

VISReg将防止坍塌的正则项解耦为「尺度」与「形状」两个独立目标,在不依赖任何启发式训练技巧、也不依赖海量数据的前提下,于15个数据集上综合表现超过7种主流自监督学习方法;其中仅用约1/10的训练数据,即在分布外(OOD)基准上追平DINOv2。

图 2:不同正则方法在表征坍缩各阶段的梯度幅值‖∇L‖模拟。VISReg在坍缩状态下仍能保持强梯度,而SIGReg的梯度几近消失。

核心方法

VISReg对VICReg与SIGReg取长补短:保留VICReg的方差项来控制尺度,同时用基于切片Wasserstein距离(Sliced Wasserstein Distance, SWD)的 sketching 目标替代协方差项来控制形状,并通过停止梯度将二者彻底解耦。整个正则目标由三部分组成。

1.尺度正则(Scale Regularization)

第一部分约束每一维的方差,防止幅值坍缩:

其关键性质在于:当模型坍缩时,该项的梯度趋近于一个常数,从而保证模型能够稳定地恢复——这恰好弥补了SIGReg梯度消失的缺陷。

2.形状正则(Shape Regularization)

第二部分先归一化以消除尺度影响,再单独约束形状。关键的一步是带「停止梯度」(stop-gradient, sg)的归一化:

这里对标准差 σσ 施加停止梯度,使得形状损失的优化不会反过来改变尺度——这正是「尺度」与「形状」两个目标真正解耦、互不干扰的机理所在。

归一化之后,再用切片Wasserstein距离将分布的几何形状对齐到各向同性高斯:

其中

为标准高斯分位数,

为随机投影方向(即「切片 / sketching」)。

其理论依据是Cramér–Wold定理(论文Lemma 3.1):两个分布相等,当且仅当它们沿单位球面上所有方向的一维投影都相等。因此,只要把高维表征沿足够多的随机一维方向切片后逐一对齐到高斯,就等价于在高维空间对齐了整个分布——这使得可以用廉价的一维排序操作刻画完整的分布形状,而非仅仅二阶统计量。

3.合并目标

第三部分是一个将 batch 均值μ拉向原点的中心化损失

三个正则项按权重组合:

预测损失沿用 JEPA / LeJEPA 的不变性目标——让各视角(global + local,共V个)的嵌入

都向全局视角的均值

对齐:

最后用单一超参λ在预测与正则之间平衡,得到完整目标:

与VICReg的对比:VICReg同样将正则解耦为方差 + 协方差,但协方差只刻画二阶统计量;VISReg用基于切片Wasserstein的sketching目标完整刻画了分布形状,同时保留方差项做尺度控制——既保留了VICReg的灵活性,又获得了分布层面的严格性。

仅需约15行PyTorch代码

该正则目标在实现上非常轻量,核心逻辑只需约15行:

计算复杂度与扩展性

在计算与扩展性上,VISReg同样具备优势。其正则部分的计算复杂度为

(N为batch、D为维度、K为切片数),对所有扩展因子都是线性的;相比之下,VICReg的协方差项为

,随维度平方增长

在同等batch规模下,VISReg 在单块 H100 GPU 上的运行速度与显存占用均优于 SIGReg。

更重要的是,K个随机切片可以分摊到多块GPU上:在 M块 GPU 上每块各生成 K/M个切片,效果等价于单卡生成全部K个。

实验中,当单卡切片数不足时,改用8卡、每卡128个切片(合计1024),即可把与「单卡 1024 切片」之间的精度差距从约2.4%缩小到0.22%。这意味着扩大训练规模时 KK 可保持常数,几乎不增加单卡负担。

图:在固定K与D时,增加GPU数量带来的线性探测精度变化。当K不足(K = 1⁄4D)时,用8倍的GPU数量即可把精度补齐到K = 2D的水平——这使得在大规模训练中保持常数K成为可能。

实验结果

回到标题的问题——VISReg 到底强在哪里?研究团队在15个数据集(8个域内 + 6个分布外 + ADE20K稠密预测)上,将VISReg与MoCoV3、DINO、iBOT、I-JEPA、MAE、data2vec等7种主流自监督方法进行了对比,场景涵盖天文、医疗、遥感、纹理、花卉等。答案体现在从识别到分割、生成的多个维度上。

1.域内(In-Domain)线性探测

为保证比较公平,实验按是否使用启发式技巧分为两组。在不使用任何启发式技巧的一组中,VISReg领先:ViT-B/16的域内线性探测精度达75.7%,高于MAE(75.1%);ViT-L/14 进一步提升至77.0%,高于LeJEPA(75.6%)。与使用启发式技巧的iBOT、DINO 相比,VISReg 在常规数据集上仅略低,但在纹理数据集DTD上反超全部方法——这表明其跨域泛化能力源于方法本身,而非人工技巧的堆叠。

2.分布外(OOD)泛化:全面最优

分布外泛化是比域内精度更严格的检验:依赖启发式的方法常在 ImageNet 域内被充分调优,却未必能迁移到差异较大的新分布。研究团队在覆盖医疗(ChestXRay、RetinaMNIST、OrganAMNIST)、天文(Galaxy10)、遥感(AID)、纹理(DTD)的 6 个 OOD 数据集上评测,这些数据集与 ImageNet 训练域完全无关。结果显示,VISReg 在所有方法、所有骨干规模上都取得了最佳的平均 OOD 精度,甚至超过部分使用启发式技巧、且骨干更大的方法。

图4:平均 OOD 线性探测精度。VISReg 全面优于 iBOT、DINO、MoCoV3、I-JEPA、MAE、data2vec 等方法。

如图4所示,ViT-B/16 的 VISReg 平均 OOD 精度为 70.19%,ViT-L/14 为 70.63%,明显高于 MAE(67.85%),并优于 MoCoV3(69.46%)、DINO(69.56%)、I-JEPA(68.55%)等方法。

3.数据效率:以 1/10 数据比肩 DINOv2

将 VISReg(ViT-L/14)在 ImageNet-22K(约 1400 万张图像)上预训练后,其 6 个 OOD 数据集的平均精度达72.94%,与在10 倍规模的 LVD-142M(1.42 亿张图像)上训练的DINOv2(72.93%)基本持平。也就是说,VISReg 以约1/10 的数据达到了同等水平。(作为对照,同为 ViT-L/14、但仅用 ImageNet-1K 预训练的 VISReg 平均精度为 70.63%。)这说明其学到的表征具有很强的通用性。

图5:在 ImageNet-22K 上预训练的 VISReg,在 OOD 基准上比肩用 10 倍数据(LVD-142M)训练的 DINOv2。

4.迁移微调:全面超过 DINO

尽管 VISReg 在部分域内数据集上的线性探测精度略低于 DINO,但经过微调后,它在 CIFAR-10、CIFAR-100、Flowers、ImageNet-1K、Galaxy10 全部五个数据集上均超过 DINO 与有监督预训练——这表明其表征分布更均匀、冗余更低、可迁移性更强。

图:迁移学习对比。微调后,VISReg 在所有测试数据集(CIFAR-10、CIFAR-100、Flowers、ImageNet-1K、Galaxy10)上都优于 DINO 与有监督预训练。

5.稠密预测与生成引导

VISReg 的优势不局限于分类。在ADE20K 线性语义分割上(ViT-B/16),其 mIoU 为30.16,高于 DINO(29.40)与 MAE(23.60),仅次于 MoCoV3(31.69);在不使用任何启发式技巧的前提下,这一结果具有竞争力。论文亦坦言,稠密预测与最佳方法仍有差距,是后续优化的重点。

图 7:ADE20K 上的线性语义分割。在不使用任何启发式技巧的情况下,VISReg 取得了具有竞争力的 mIoU,仅次于 MoCoV3。

生成引导上(SiT-B/2,iREPA 框架,10 万步训练),由 VISReg 特征引导的生成在四项指标中的三项优于 DINO:gFID40.36(DINO 41.15)、Precision51.38(DINO 50.51)、Recall61.26(DINO 60.70),IS 基本持平(33.48 vs 33.47)。这说明 VISReg 学到的表征作为生成引导信号同样更优。

图8:使用 SiT-B/2、分别由 VISReg 与 DINO 特征引导的图像生成。VISReg 在多数指标上都提供了更好的引导(更低的 gFID、更高的 Precision 与 Recall)。

6.低质量数据上的鲁棒性

长尾分布(ImageNet-LT)与低秩(Galaxy10)等低质量数据集上,VISReg 能稳定地防止坍塌并学到有意义的表征,而 DINO 在缺乏精细调参时直接失败。

表 1:ImageNet-LT上的线性探测精度

表 2:Galaxy10 上的域内线性探测精度

结论

VISReg 表明:将表征正则解耦为「尺度」与「形状」两个独立组件,可以得到一种比现有方法更稳定、更高效、泛化性更强的自监督学习方法。

在不使用任何训练启发式技巧的前提下,它在图像识别、分割与生成引导等多个维度上取得了领先或接近领先的结果,并以约 1/10 的数据达到了 DINOv2 的 OOD 水平。这为 JEPA 世界模型长期存在的表征坍塌问题提供了一种新的正则化解法。

参考资料:

https://arxiv.org/abs/2606.02572

本文来自微信公众号“新智元”,作者:LRST

Trending Cryptos

Related Questions

QVISReg方法主要解决了自监督学习中的什么核心难题?

AVISReg主要解决了自监督学习(SSL)中‘表征坍塌’的核心难题。表征坍塌是指模型倾向于将不同的输入映射到相同或极少数几个向量上,导致学习到的表征缺乏判别力。

Q与VICReg和SIGReg相比,VISReg的核心创新点是什么?

AVISReg的核心创新点在于:它继承了VICReg的方差项来控制表征的‘尺度’,同时引入了基于切片Wasserstein距离(SWD)的sketching目标来约束表征的‘形状’,并通过‘停止梯度’技术将这两个目标彻底解耦,使其在优化中互不干扰。这解决了SIGReg在表征坍塌时梯度消失、以及尺度与形状耦合的问题。

Q在实验结果中,VISReg在哪些关键性能指标上表现突出?请列举至少三个。

A1. 分布外(OOD)泛化:在6个与训练域无关的OOD数据集上,VISReg的平均精度全面优于包括DINO、MAE在内的多种主流方法。 2. 数据效率:使用约1/10的训练数据(ImageNet-22K),其在OOD基准上的表现追平了使用10倍数据(LVD-142M)训练的DINOv2。 3. 迁移微调:经过微调后,在CIFAR-10、CIFAR-100等多个数据集上的分类准确率均超过DINO和有监督预训练模型。 4. 低质量数据鲁棒性:在ImageNet-LT(长尾分布)和Galaxy10(低秩数据)等低质量数据集上能稳定学习,而DINO等方法在缺乏精细调参时可能失败。

QVISReg的计算复杂度与VICReg相比有何优势?

AVISReg的正则项计算复杂度为O(N*D*K),其中N为批大小,D为特征维度,K为随机切片数量,对所有扩展因子(如D和K)是线性的。而VICReg的协方差项复杂度为O(N*D²),随特征维度D呈平方增长。因此,VISReg在计算上更高效,尤其在高维场景下。此外,其随机切片操作易于在多GPU间并行,扩展性更好。

QLeCun对VISReg的评价是什么?这反映了该工作在技术路线上的何种地位?

ALeCun在转发VISReg工作时评价道:“VICReg begat SIGReg which begat VISReg”(VICReg孕育了SIGReg,SIGReg又孕育了VISReg)。这句话简洁地指明了VISReg在技术谱系中的传承关系:它是在VICReg和SIGReg基础上的重要发展,标志着其在解决JEPA世界模型‘表征坍塌’这一核心难题的正则化路线上,是一个关键的迭代和进步。

Related Reads

Weekly Editor's Picks (0725-0731)

Weekly Editor's Picks (0725-0731) provides a curated selection of deep analysis, filtering out market noise. Key themes from this week include: **Macro & Policy:** The Federal Reserve's upcoming meeting is marked by high uncertainty, balancing cooling inflation data against persistent price pressures. Meanwhile, the U.S. crypto regulatory Clarity Act faces critical political hurdles, with its 2026 passage probability seen as low. **Investing & Crypto:** Analysis suggests long-term crypto success depends on conviction through volatile cycles, focusing on assets like Bitcoin and core smart contract platforms. A trend noted is the increasing similarity between global equity markets (especially tech) and crypto, driven by narrative and leverage. Several major crypto protocols show strong revenue growth, but this isn't always translating to token price appreciation due to sell pressure and structural factors. **AI & Semiconductors:** Nvidia's rising credit default swap rates signal market concern over AI infrastructure financing risks. The storage sector experienced volatility as markets began pricing in potential 2027 oversupply. Despite a record profitable quarter, SK Hynix's results were deemed "below expectations," reflecting heightened investor demands for future growth visibility. **Markets & DeFi:** TradeXYZ demonstrated remarkable accuracy in pre-market pricing for a major A股 listing. The token ONDO saw gains, linked to its growing role in the on-chain tokenized stock ecosystem. **Ethereum:** Post-Pectra upgrade, a major structural shift is underway as Lido begins migrating millions of ETH to new validator architectures designed for capital efficiency. **Also Highlighted:** Butian's bullish stock market move; OpenAI's Altman promising major advances; Samsung and SK Hynix securing large AI chip deals; Apple reaching a $5T market cap; and ongoing discussions around exchange security following Poolin's bankruptcy case.

marsbit22m ago

Weekly Editor's Picks (0725-0731)

marsbit22m ago

Low Investment Isn't Apple's Immunity Pass

While Meta and Google face investor scrutiny over ballooning AI capital expenditures, Apple's minimal AI investment has paradoxically become a strength. Its market cap recently reclaimed the global top spot, surpassing $5 trillion. The irony is deep: Apple's own AI efforts have lagged, with "Apple Intelligence" delayed and core talent lost, forcing reliance on partners like Google Gemini and Alibaba's Qianwen. Its Q3 FY2026 (Q2 CY) earnings initially seemed stellar. Revenue hit $109.4B (up 16% YoY), with iPhone and Mac sales, growing 22% and 29% respectively, driving most of the growth. However, the stock fell over 8% post-earnings. The primary concern was a weaker Q4 revenue growth forecast of 9-11%, below expectations, due to looming supply chain constraints. Apple is feeling the indirect cost of the AI boom. Soaring memory and chip prices, fueled by massive data center investments from Microsoft, Amazon, and others, are forcing Apple to raise Mac and iPad prices significantly. The upcoming iPhone launch is also expected to see substantial price hikes. Despite avoiding heavy AI infrastructure spending—its capital expenditures are actually down 28%—Apple cannot escape the industry-wide supply and cost pressures. While Apple's operating cash flow remains robust, its substantial R&D spending (up 32% YoY) has yet to yield major AI breakthroughs. As Tim Cook prepares to step down as CEO, Apple faces a challenging transition: balancing its premium hardware success against the strategic and cost pressures of the AI era it has so far cautiously navigated.

marsbit1h ago

Low Investment Isn't Apple's Immunity Pass

marsbit1h ago

PA Graphics Explanation | One Chart to Understand the Major Web3 Events in August 2026

**PANews Crypto Calendar: Key Web3 Events in August 2026** PANews introduces its revamped crypto calendar, featuring comprehensive coverage, flexible filtering, and easy export options. The market in August will be shaped by multiple key events across macroeconomics, regulation, tokenomics, and project developments: * **Macro & Policy:** Key US economic data releases (July Non-Farm Payrolls, CPI), the Federal Reserve meeting minutes, and the Jackson Hole Economic Symposium will be in focus. On the regulatory front, the US Senate plans to release a new draft of the *CLARITY Act*, while the EU's expanded crypto ban against Belarus comes into effect. * **Token Unlocks:** Significant token unlocks are scheduled for assets including ENA, AVAX, CONX, ZRO, and KAITO, which may influence market volatility. * **Project Updates & Shutdowns:** Several services, including Exchange Art, Ctrl Wallet, Zapper, NFTfi, and Summer.fi, are set to cease operations or undergo major adjustments. Users are advised to manage their assets accordingly. * **Corporate Activity:** Q2 earnings reports from companies like SpaceX, Circle, and Nvidia are due. Unitree Robotics will initiate its IPO subscription on the STAR Market, and Moonshot AI plans to begin a Pre-IPO financing round. * **Industry Events:** Major conferences such as Bitcoin Asia 2026 and the 2026 Digital Expo will take place. The overarching market narrative for August will revolve around macroeconomic expectations, regulatory developments, token unlock schedules, and ongoing industry consolidation.

marsbit1h ago

PA Graphics Explanation | One Chart to Understand the Major Web3 Events in August 2026

marsbit1h ago

Wall Street's Most Famous 'Cassandra' Now Has His Sights Set on Nvidia

Michael Burry, the famed "Big Short" investor, has once again captured Wall Street's attention with a series of short positions against major tech and semiconductor stocks, most notably Nvidia. In late June and July, through his "Cassandra Unchained" newsletter, Burry disclosed short bets against Nvidia, Tesla, Applied Materials, Caterpillar, the SOXX semiconductor ETF, and later, Micron Technology. His core thesis revolves around potential distortions in the AI infrastructure boom, specifically questioning whether extended depreciation schedules (e.g., 6 years vs. a realistic 2-3 years for AI chips) by cloud giants like Microsoft and Google artificially inflate profits. He also raises concerns about possible "off-balance-sheet circular financing," where chip demand might be propped up by vendor-backed funding to clients. Nvidia's stock experienced volatility following these disclosures, briefly dipping but largely holding near Burry's reported entry points, leaving his positions roughly flat or slightly underwater as of late July. This move is part of a pattern for Burry, whose track record since his legendary 2008 bet is mixed. He has faced notable losses, such as on Tesla in 2021, while scoring on broader market turns like the 2020 pandemic crash. His methodology focuses intensely on free cash flow and scrutinizing original financial documents to spot overvaluation and structural risks, but it often struggles with timing the market. The article contrasts Burry's stance with other prominent investors. Steve Eisman, another "Big Short" figure, is not shorting Nvidia, citing strong fundamentals but expressing nervousness about sustainability. Jim Chanos agrees with the broad "accounting mismatch" concern—comparing it to the dot-com bubble—but targets financial leverage in private equity firms rather than the chip stocks themselves. While Nvidia's short interest remains relatively low at 1.3-1.4% of float, the massive stock size means absolute short losses have been significant, exceeding $5 billion earlier this year. The piece concludes that for ordinary investors, the key takeaway is not replicating specific short bets but learning from the critical frameworks these investors use: questioning rosy accounting, identifying structural vulnerabilities, and maintaining skepticism during market euphoria, even if pinpointing the exact catalyst for a downturn remains elusive.

marsbit1h ago

Wall Street's Most Famous 'Cassandra' Now Has His Sights Set on Nvidia

marsbit1h ago

Trading

Spot

Hot Articles

How to Buy CORE

Welcome to HTX.com! We've made purchasing CORE (CORE) simple and convenient. Follow our step-by-step guide to embark on your crypto journey.Step 1: Create Your HTX AccountUse your email or phone number to sign up for a free account on HTX. Experience a hassle-free registration journey and unlock all features.Get My AccountStep 2: Go to Buy Crypto and Choose Your Payment MethodCredit/Debit Card: Use your Visa or Mastercard to buy CORE (CORE) instantly.Balance: Use funds from your HTX account balance to trade seamlessly.Third Parties: We've added popular payment methods such as Google Pay and Apple Pay to enhance convenience.P2P: Trade directly with other users on HTX.Over-the-Counter (OTC): We offer tailor-made services and competitive exchange rates for traders.Step 3: Store Your CORE (CORE)After purchasing your CORE (CORE), store it in your HTX account. Alternatively, you can send it elsewhere via blockchain transfer or use it to trade other cryptocurrencies.Step 4: Trade CORE (CORE)Easily trade CORE (CORE) on HTX's spot market. Simply access your account, select your trading pair, execute your trades, and monitor in real-time. We offer a user-friendly experience for both beginners and seasoned traders.

6.0k Total ViewsPublished 2024.03.29Updated 2026.06.02

How to Buy CORE

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of CORE (CORE) are presented below.

活动图片