OpenAI’s recent release of GPT-3 API set the Twitterverse on fire with developers using its massive language model to generate everything from text to code:

OpenAI最近发布的GPT-3 API使得Twitterverse受到开发人员的欢迎，他们使用其庞大的语言模型来生成从文本到代码的所有内容：

While the hype and the attention towards GPT-3 were justified, another exciting development in AI went largely unnoticed. On July 14th, Snorkel AI, which spun out of Stanford AI Lab in 2019, emerged out of stealth with $15 million in funding from Greylock, GV, and In-Q-Tel. The launch showcased Snorkel Flow, a machine learning platform to programmatically label and prepare the training data to accelerate the build and deploy process for ML models. Snorkel’s early users include Google, Intel, Apple, and Stanford Medicine. Although the project is still in its infancy, Snorkel’s approach to AI may be an important breakthrough for enterprise AI and the broad adoption of AI/ML across various verticals.

尽管对GPT-3的大肆宣传和关注是有道理的，但AI的另一个激动人心的发展在很大程度上并未引起注意。 7月14日， Snorkel AI于2019年从斯坦福大学AI实验室中分离出来，并从Greylock，GV和In-Q-Tel获得了1500万美元的资金 ，脱颖而出。 此次发布会展示了Snorkel Flow，这是一个机器学习平台，可通过编程方式标记和准备训练数据，从而加快ML模型的构建和部署过程。 Snorkel的早期用户包括Google ， Intel ， Apple和Stanford Medicine 。 尽管该项目仍处于起步阶段，但Snorkel的AI方法可能是企业AI的重要突破，以及AI / ML在各个垂直领域的广泛采用。

手贴训练数据瓶颈 (Hand Labeling Training Data Bottleneck)

Despite massive improvements to machine learning frameworks (e.g. Tensorflow, PyTorch, Keras), hardware (e.g. GPUs, TPUs), and research across the board (e.g. AlphaGo, GPT-3), preparing training data remains largely a manual process. From drawing bounding boxes on images to annotating audio files, data scientists have to either label massive files by hand or crowdsource the task to an army of contract workers. The problem is exacerbated by the massive amounts of data needed to train and build deep learning models. This bottleneck is becoming more apparent as deep learning is more accessible than ever thanks to various open-source tools and cloud-hosted workloads.

尽管机器学习框架大量的改进(如Tensorflow ， PyTorch ， Keras )，硬件(例如图形处理器，热塑性聚氨酯)，以及全线研究(如AlphaGo，GPT-3)，准备训练数据在很大程度上仍然是手动过程。 从在图像上绘制边界框到注释音频文件，数据科学家必须手动标记大量文件或将任务众包给合同工大军。 训练和建立深度学习模型所需的大量数据加剧了这个问题。 由于各种开放源代码工具和云托管工作负载，深度学习比以往任何时候都更容易访问，因此这一瓶颈变得越来越明显。

Compare the manual labeling process to significant improvements in other stages of machine learning. We now have AutoML and other programmatic methods to expedite feature engineering and hyperparameter tuning. Infrastructure provisioning tools and cloud architectures allow the deployment of those models easier than ever. On the other hand, hand labeling data does not scale (at least not cost-efficiently, especially when expertise and privacy may be involved — e.g. medical images, financial statements) and is extremely error-prone.

将手动标记过程与机器学习其他阶段的重大改进进行比较。 现在，我们有了AutoML和其他编程方法来加速特征工程和超参数调整。 基础架构配置工具和云体系结构使这些模型的部署比以往更加容易。 另一方面，手贴标签数据无法扩展(至少不具有成本效益，特别是在涉及专业知识和隐私的情况下，例如医学图像，财务报表)，并且容易出错。

The team at Snorkel argues that the process of manually labeling training data is fundamentally broken. Behind today’s massive success of ML milestones belie a hidden cost. For example, it took Dr. Fei Fei Li and her team at Stanford AI Lab two years to create ImageNet, the foundational dataset that led to incredible research at Google, Clarifai, DeepMind, Baidu, and Huawei. In order for data to truly become the new oil in the digital economy, turning dataset labeling into an iterative development process is paramount.

Snorkel的团队认为，手动标记训练数据的过程从根本上被破坏了。 在当今ML里程碑取得巨大成功的背后，隐藏着代价。 例如， 费飞利博士及其团队在斯坦福大学AI实验室花了两年时间创建了ImageNet ，该基础数据集导致Google，Clarifai，DeepMind，百度和华为进行了惊人的研究。 为了使数据真正成为数字经济中的新油，将数据集标签转变为迭代开发过程至关重要。

“Despite spending billions of dollars on AI, few organizations have been able to use it as widely and effectively as they want to. This is because available solutions either ignore the most important part of AI today — the labeled training data that fuels modern approaches — or rely on armies of human labelers to produce it.” — Alex Ratner, CEO of Snorkel AI

“尽管在人工智能上花费了数十亿美元，但很少有组织能够像他们想要的那样广泛而有效地使用它。 这是因为可用的解决方案要么忽略了当今AI最重要的部分-标记了现代方法的培训数据，要么依赖人类标签制造商的队伍来生产它。” — Snorkel AI首席执行官Alex Ratner

程序标记数据 (Programmatic Labeling of Data)

Snorkel AI’s founding team — Alex Ratner (Assistant Professor at the University of Washington) and Chris Ré (Associate Professor at Stanford and 2015 MacArthur Genius Fellow) — sought to implement rule-based systems to programmatic label and build quality training data. The idea is to allow the programmer or the domain expert to define labeling functions to generate labeled data. The example used on the Snorkel blog takes legal documents where a legal analyst creates a function to tag documents as “Employment contracts” if “employment” is in the title of the document.

Snorkel AI的创始团队-华盛顿大学助理教授亚历克斯·拉特纳(Alex Ratner)和斯坦福大学副教授兼克里斯·雷( ChrisRé )以及2015年麦克阿瑟·天才研究员(MacArthur Genius Fellow )–试图实现基于规则的系统来程序化标签和建立高质量的培训数据。 这个想法是允许程序员或领域专家定义标记函数以生成标记数据。 Snorkel博客上使用的示例获取法律文档，其中，如果文档标题中包含“ employment”，则法律分析师会创建一个将文档标记为“ Employment contract”的功能。

While rule-based systems have the advantage of being simple and direct, it is also brittle in that it lacks robustness to deal with inputs outside the defined schema (e.g. what happens if “contract” is used instead of “employment”). So Snorkel Flow takes the dataset created by labeling functions, treats it as a generative model filled with noise, and uses weak supervision ML models to “denoise” or generalize that algorithm. In essence, it combines human-defined rules and machine learning algorithms to create a more robust labeled dataset.

尽管基于规则的系统具有简单直接的优势，但它也很脆弱，因为它缺乏处理已定义模式之外的输入的鲁棒性(例如，如果使用“合同”而非“雇佣”会发生什么)。 因此，Snorkel Flow提取了通过标注函数创建的数据集，将其视为充满噪声的生成模型，并使用弱监督ML模型来“降噪”或推广该算法。 本质上，它结合了人类定义的规则和机器学习算法，以创建更强大的标记数据集。

呼吸管流量 (Snorkel Flow)

While the original, open-source Snorkel project focused on programmatic approaches to training data, Snorkel Flow goes further to tackle building an end-to-end ML solution. This includes data augmentation (e.g. adding rotate or blurred images), data slicing (i.e. taking subsets of data), and operational components needed for a production-grade ML application (e.g. monitoring, alerting, IaaC).

最初的开源Snorkel项目专注于训练数据的编程方法，而Snorkel Flow则进一步致力于构建端到端的ML解决方案。 这包括数据扩充(例如，添加旋转或模糊的图像)，数据切片(即，获取数据的子集)以及生产级ML应用程序所需的操作组件(例如，监视，警报，IaaC)。

Access to Snorkel Flow is currently limited (need to request a demo), but the fundamental idea behind Snorkel seems like the missing link to democratizing machine learning beyond the tech behemoths. Keep an eye on this startup and see how it reshapes the ML industry in years to come.

目前对Snorkel Flow的访问受到限制( 需要请求演示 )，但Snorkel背后的基本思想似乎是使技术巨头之外的机器学习民主化的缺失环节。 密切关注这家初创公司，看看它如何在未来几年内重塑机器学习行业。