Make with H2O.ai: Natural Language Processing with H2O Hydrogen Torch
- Predictive AI
- H2O Hydrogen Torch
Overview
This session introduces H2O Hydrogen Torch, H2O.ai's no-code deep learning engine for unstructured data such as text, images, video, and audio. The presenter walks through why the tool exists, to close the gap between the volume of unstructured data companies hold and the scarce deep learning talent needed to work with it, and demos several NLP problem types, including text regression, text classification, token classification, metric learning, sequence to sequence summarization, and question answering, using the platform's live experiment interface.
Key takeaways
- What H2O Hydrogen Torch is: a no-code deep learning engine that trains models on your own unstructured data, text, images, video, or audio, without writing PyTorch or TensorFlow code and without sharing data externally
- How to set up and run a text regression experiment end to end in the UI, including picking a pre-trained backbone model straight from Hugging Face
- How text classification, token classification (entity extraction), and metric learning (duplicate question detection) work as distinct NLP problem types in the platform
- How sequence to sequence models handle text summarization, and how span prediction handles question answering, both demoed live on sample datasets
- How to move from a novice experience with just a few exposed hyperparameters to a master view with full hyperparameter control, including multi-GPU training and grid search
- How to deploy a finished model, either as a REST API through H2O MLOps or as a portable scoring pipeline you can run anywhere with Python
Start owning your AI