AI vendor H2O.ai introduces new deep learning tool
H2O Hydrogen Torch can be used in healthcare, HR and customer experience settings. Users can make models for image, video and natural language processing applications.
The vendor said the tool will enable enterprises to create machine learning models from unstructured data such as images, videos and product reviews for applications in healthcare, manufacturing, insurance and other areas.
"H2O[.ai] has a lot of experience in operationalizing machine learning models with enterprises," said Kashyap Kompella, an analyst at RPA2AI Research. "This is an extension of that capability in the deep learning context."
H2O Driverless AI is an automated machine learning platform for difficult data science and machine learning workflows such as feature engineering and model validation and tuning.
H2O[.ai] has a lot of experience in operationalizing machine learning models with enterprises. This is an extension of that capability in the deep learning context.
Kashyap KompellaAnalyst, RPA2AI Research
However, the system is designed mainly for applications that apply traditional machine learning methods not deep learning.
Hydrogen Torch's focus on unstructured data will enhance H2O.ai's portfolio in the deep learning area, Kompella said.
"This also reflects the adoption of deep learning by enterprises," he said, noting that H2O.ai is positioning itself as a provider of commercial support and add-on products to go on top of core open source offerings.
AI and healthcare
But H2O.ai may be at risk of overpromising with the various potential applications of Hydrogen Torch for deep learning, said Dan Miller, an analyst at Opus Research.
H2O.ai said for images and videos, Hydrogen Torch can classify, regress and detect objects. The tool could also be trained to analyze medical X-ray images for abnormalities with a "human in the loop," for example, according to the vendor.
One use case of Hydrogen Torch is training models to identify abnormalities in the medical field.
Using AI in medical settings, though, may be more complicated than H2O.ai thinks, Miller said.
"Medical applications of AI is a quagmire that was exposed when IBM Watson tried to make a showcase of human-in-the-loop applications for diagnostics," Miller continued, referring to IBM's Watson Health AI system, most of whose tech assets the tech giant sold off recently.
H2O.ai's attempt to reach into healthcare is reminiscent of how IBM tried to use Watson to work with hospitals to find patterns and anomalies in X-rays, Miller said.
While IBM Watson had some successes, ultimately healthcare providers and the vendor realized that accurate diagnoses and treatment recommendations must come from physicians and not mainly from AI models.
While IBM Watson and Hydrogen Torch do compete as AI vendors, a better comparison would be between hype versus reality in AI, argued Sri Ambati, CEO and cofounder of H2O.ai.
"Overpromis[ing] and underdeliver[ing] has been a problem for our space, but that doesn't take away the fact that there's a lot of good innovation that has come to bear," Ambati said.
And while healthcare is complex and it is easy to get bogged down by the complexity of the industry, it is important to work with human domain experts, he said.
"The problem is difficult, and deep and multidimensional," he said.
Because diseases and health conditions affect individual patients differently, among other reasons, it is difficult to apply AI to situations that vary so widely, Ambati said.
To ensure AI algorithms and machine learning models are responsive to such complex real-world scenarios, H2O.ai is working closely with major U.S. hospitals and healthcare organizations to bring what Ambati called a "multidimensional approach" to healthcare, he said.
"I think the early successes or pronounced failures of Watson and others were because they were kind of the dinosaurs of our space," he continued. "The market has changed dramatically since those failures."
Creating a no-code model
For NLP, Hydrogen Torch can be used to create models for text classification, regression, token classification and sequence-to-sequence analysis. H2O.ai said the tool can predict how satisfied a customer is from a transcribed phone call or by summarizing and analyzing a large portion of text such as a medical transcript.
However, "the description of using H2O Hydrogen Torch for natural language processing doesn't feel accessible to the sorts of folks they are appealing to," Opus Research's Miller said.
Professionals who are looking for a no-code AI tool may not necessarily understand the features of Hydrogen Torch, he said.
"If you're going to ingest a lot of unstructured data ... and then detect patterns, then there ought to be a sort of a natural language way to understand the patterns that you saw," Miller continued. "I think they're saying that they do that, but I just can't tell."
According to Ambati, new data scientists or business users do not need to use code to create or train models in Hydrogen Torch. However, the low-code side of the tool comes into play when taking the models created with Hydrogen Torch and embedding them in applications or deploying them into API-based services to make them easier to consume.
H2O.ai said Hydrogen Torch was built by Kaggle Grandmasters, expert members of the Google-owned online community of data science and machine learning practitioners who host competitions and a cloud-based workbench, and also publish data sets.
The tool is currently available on H2O's AI cloud platform.
Customers can try the deep learning tool for free. Prices for customers that want to purchase the service depends on the application, Ambati said.
