Jan. Update: Misconceptions of AI, democratizing weak supervision, and more!
Happy new year from the Snorkel team!
With over 175K online articles, artificial intelligence was the most discussed technology in 2021. We believe this is going to be the case in 2022 as well.
Despite so much interest, for the general public, who mostly learn about AI through the internet or movies and books, one of the primary misconceptions is that they often think of AI as artificial general intelligence (AGI). They tend to correlate AI to something that is even sentient or like a whole agent with whole intelligence, and they are not aware that most practical AI these days is applied to specific tasks. Dive deeper into the misconceptions of AI with Abigail See, an NLP Research Scientist at Stanford University.
Weak supervision was widely adopted in 2021 as a popular and practical way to apply AI/ML by bypassing hand-labeling large datasets for training data-hungry models. Snorkel AI Researcher Fred Sala and team are pushing the state-of-the-art forward by proposing new techniques to universalize weak supervision, enabling it to work for any label type while still offering desirable properties, including practical flexibility, computational efficiency, and theoretical guarantees. We expect to see more cutting-edge research on weak supervision and its application in the real world in 2022.
In 2021, data-centric AI development practice also went mainstream. Data science and ML teams can unlock higher model accuracy and faster development times than using model-centric approaches alone by creating, managing, and iterating on training data programmatically. Join our upcoming webinar on February 1, 2022, at 11:00 AM Pacific Time, where Snorkel AI engineering leader Roshni Malani and machine learning engineer Priyal Aggarwal will show how to build data-centric AI applications. Register now>>
Until our next edition, anchors aweigh!
BLOG
Epoxy: using semi-supervised learning to augment weak supervision
Humza Iqbal, ML Research Engineer at Snorkel AI, discusses the paper “Train and You’ll Miss It: Interactive Model Iteration with Weak Supervision and Pre-Trained Embeddings.” Watch this ML Whiteboard session to learn how to use Epoxy to train models at programmatically-interactive speeds while retaining the performance of training deep networks.
Watch now»
EVENTS
Live webinar: building data-centric applications
Join our upcoming webinar on February 1, 2022, at 11:00 AM Pacific Time to learn how modern AI application development is shifting to a data-centric approach. During the live event, you’ll see how you can build high-quality applications using the Snorkel framework, techniques for iteration, adaptation, and collaboration, and how to empower domain experts. Plus there will be plenty of time to ask questions during the live Q&A session.
Register now»
How financial services can automate data labeling and accelerate AI application development
Watch the recording of a recent webinar to learn how Snorkel AI is helping top US banks and global financial organizations cut down AI application development time by 10–100x with programmatic labeling.
Watch now»
ML whiteboard live
Join us for an upcoming live, virtual ML whiteboard session. Where data scientists, machine learning engineers, developers, and Snorkel AI team members discuss the latest research and new techniques for machine learning, deep learning, NLP, and more.
Register now»
We’re hiring!
Snorkel AI is growing rapidly, and we’re hiring passionate problem solvers who want to shape the future of AI.
See open roles
PODCASTS
Removing the AI bottleneck to roll out the Model T
In this episode of “The Next Wave,” Snorkel Co-Founder & CEO Alex Ratner discusses how we are just approaching the late innings of the first game in AI. Listen in to hear how companies are accelerating AI application development with Snorkel’s programmatic approach to labeling.
Listen in»
RESEARCH
Universalizing Weak Supervision
Snorkel AI Researcher Fred Sala and team propose a technique to universalize weak supervision, enabling it to work for any label type while still offering desirable properties, including practical flexibility, computational efficiency, and theoretical guarantees.
Read more»
