Advancing Snorkel from Research to Production



Snorkel Research Project

The Snorkel AI founding team started the Snorkel Research Project at Stanford AI Lab in 2015, where we set out to explore a higher-level interface to machine learning through training data. This project was sponsored by Google, Intel, DARPA, and several other leading organizations and the research was represented in over 40 academic conferences such as ACL, NeurIPS, Nature and more. 

Snorkel Open Source Research Library was primarily developed from 2015 to 2017 as a prototyping tool. It is a Python library that contains a legacy base class for defining code-based Labeling Functions (LFs) and some early algorithms for combining LF votes, rather than a comprehensive platform supporting the AI development lifecycle.


Image

Snorkel Research Project

Snorkel Flow

Snorkel Flow Platform was built by the original creators of the Snorkel Research Project, incorporating years of experience from applying weak supervision and programmatic labeling concepts to real-world ML problems.

In Snorkel Flow, users can label and manage data using code, train models and identify model error modes to iteratively improve them in a rapid, data-centric workflow, using both SDK and no-code interfaces. This shortens the development cycle and improves application quality significantly while being able to manage bias or adapt to any deterioration in production performance.

Snorkel Flow is used by some of the world’s most advanced organizations in banking, insurance, biotech, telecommunications and several government agencies.


Image
Image

Snorkel Flow Platform Interfaces

Snorkel Research Project



Feature Evolution

Snorkel Flow is an enterprise-grade platform that uses the core concepts of Snorkel Research Project. With Snorkel Flow, enterprises are able to build and deploy accurate and adaptable AI applications rapidly.


Snorkel Flow Platform
Image
FEATURE
PROJECT
SNORKEL
RESEARCH PROJECT
SNORKEL FLOW
FLOW

PROGRAMMATIC LABELING

Programmers write Labeling Functions (LF) in Python code
Image
Image
Programmers and SME users create LFs in a no-code, push-button UI
Image
UI-based analysis, feedback, and suggesting to guide iterative LF development
Image
Auto-suggest and auto-tuning of LFs
Image
Built-in interactive data visualization with support for building LFs by drawing directly on data plots
Image
Automated management and versioning of LFs
Image


TRAINING DATASET MANAGEMENT

Basic algorithms for denoising and combining LF outputs
Image
Image
Advanced algorithms and automated tuning for denoising and combining LF outputs, including correlation analysis
Image
One-click to execute LFs with automated parallelization and label model optimization
Image
Automated management and versioning of training datasets
Image


MODEL TRAINING AND ANALYSIS

Train custom models using Python
Image
Image
One-click to train and tune pre-configured, state-of-the-art models via the built-in model zoo
Image
Auto-generated UI-based model analysis with suggestions for model and LF improvement
Image
Automated management and versioning of models
Image


APPLICATION AND MODEL SERVING

Support for complex application graphs combining multiple models, pre-/post-processors, and custom business logic
Image
One-click endpoint creation for model/ application serving
Image
One-click model/application export for serving at-scale
Image


DEPLOYMENT AND SECURITY

REST API, monitoring services, and managed workers for job execution
Image
Snorkel AI-hosted and managed hybrid cloud (AWS) deployment
Image
Support for distributed deployment via Kubernetes
Image

Encryption (in-transit and at-rest), authentication, and role-based access control (RBAC)

Image
Managed SSO integration with SAML 2.0 support
Image


TRAINING AND SUPPORT

Enterprise training and support from Snorkel AI engineers
Image

Accelerate your AI application development today


Accelerate your ML application development today

Request demo
ImageImage