H2O Hydrogen Torch promises to make deep learning model development easier

Source: H2O.ai. Screen from H2O Hydrogen Torch.

H2O.ai has announced H2O Hydrogen Torch, a deep learning training engine that makes it easy for companies of any size in any industry to make state-of-the-art image, video and natural language processing (NLP) models without coding.

Creating deep learning models typically requires extensive data science knowledge and time. H2O Hydrogen Torch uses a no-code user interface that allows data scientists and developers to rapidly build models for image, video and NLP processing use cases such as identifying or classifying objects, analysing sentiment or finding relevant information in text.

According to multiple analyst estimates, 80% to 90% of data is unstructured information, yet only a small percentage of organisations are able to derive value from unstructured data. Deep learning models could be used in industries like healthcare for computer-aided disease detection or diagnosis through the analysis of medical images; in insurance for the automation of claims and damage analysis from reports and images; and in manufacturing for predictive maintenance through analysing images, video and other sensor data).

H2O Hydrogen Torch user Aura.ceo offers a talent screening platform that uses public data to evaluate the array of roles, skills and experience inside a company of any size and see how it compares to competitors.

Said Stelios Anagnostopoulos, CTO at Aura.ceo, “H2O Hydrogen Torch has been a key enabler in helping us operationalise machine learning for shifting data. We can get from a new dataset to a deployed model and updated tables in our data warehouse in a couple of days instead of weeks.”

H2O Hydrogen Torch can be trained for classification, regression, object detection, semantic segmentation and metric learning* of images and videos. In a medical setting, for example, H2O Hydrogen Torch could analyse X-ray images for abnormalities with a “human in the loop” to make the final decision. Other image-based use cases include object detection in a manufacturing facility to determine whether a part is missing or metric learning that alerts an online retailer to duplicate images on a website.

For text-based or NLP use cases, H2O Hydrogen Torch can be trained for text classification and regression, token classification**, span prediction**, sequence-to-sequence analysis*** and metric learning. NLP use cases include predicting customer satisfaction from transcribed phone calls to sequence-to-sequence analysis to summarise a large portion of text, such as from medical transcripts, in a few sentences.

These models then can be packaged automatically for easy deployment to external Python environments or in a consumable format directly to H2O MLOps for production. MLOps refers to methodology that brings machine learning and operations together.

“Accelerated by COVID-19, video streams, speech, audio podcasts, email and natural language text have become the fastest growing data for our customers in every industry. Transforming and finetuning pre-built deep learning models to deliver high accuracy requires a no-code AI engine to democratise AI for these use cases,” said Sri Ambati, CEO and founder, H2O.ai. AI stands for artificial intelligence.

“H2O Hydrogen Torch does exactly that by bringing best practices from Grandmasters**** to tackle problems ranging from improving in-store customer experiences, identifying fashion trends, and discovering vaccines, to saving lives with video-enabled drones fighting fires with AI on the edge. With H2O Hydrogen Torch as a core AI engine of the H2O AI Cloud, our customers can train models in deep learning and better serve their customers and challenge tech giants.”

H2O Hydrogen Torch is part of H2O.ai’s broad and rapidly expanding set of H2O AI Cloud products, including the recently-announced H2O AI Feature Store and H2O Document AI.

Explore:

Get a free trial of H2O Hydrogen Torch.

*Metric learning refers to a way of deciding whether two items are similar to each other.

**Token classification is the identification of a single item in a document as a token, then categorising it. Span prediction refers to predictions made on a cluster of tokens within a document, or a larger segment of a document.

***Sequence-to-sequence analysis refers to being able to transfer expertise from one domain into another. A classic example would be to port a sentence from one language into another language.

****Grandmasters are data scientists who are top-ranked in Kaggle rankings.