Machine learning operations startup Abacus.AI Inc. is in the news today after raising $50 million in a new round of funding and launching platform updates that include support for hybrid deep learning models and a new computer vision-as-a-service offering.

The Series C funding round was co-led by Tiger Global, Coatue, Index Ventures and Alkeon Ventures and brings Abacus.AI’s total amount raised to $90.3 million.

Abacus.AI has been able to attract those millions thanks to its enterprise-scale MLOps platform that companies of any size can use to stream real-time data from online purchases, social media interactions, “internet of things” sensors and more. Then they can process that data, transform it, and use it to create deep learning-based artificial intelligence models capable of generating contextual predictions in real-time.

Until Abacus.AI arrived on the scene, only a few large enterprises with hundreds of thousands of dollars available to invest in such endeavors were ever able to build such complex models.

So the Abacus platform effectively brings deep learning to the masses. It enables an MLOps discipline that aims to put workflows into operation by fostering more collaboration and communication between data scientists and developers.

It provides all of the components needed to build complex AI systems, with easy setup of data pipelines, tools for cleaning and transforming data, model training and hosting, monitoring and explainability capabilities and more. Users can train models using Abacus.AI’s own neural network architecture or alternatively they can use an open-source framework such as PyTorch or TensorFlow.

The company has effectively created an “end-to-end AI service” that has the potential to become a one-stop shop for all enterprise AI needs, Andy Thurai, an analyst with Constellation Research Inc., told SiliconANGLE.

It’s a powerful platform that’s now getting even better. With today’s release, Abacus.AI is adding support for “hybrid models” that can generate predictions from a combination of language, vision, and tabular data. This, the company said, is a novel technique that relies on neural net architecture search combined with pre-trained embedding for image and language data to extract intelligence.

It’s a promising development because existing AI and machine learning tools typically only work with structured tabular data, meaning unstructured data such as photos, videos and call transcripts cannot be used to train models. With Abacus.AI’s hybrid models, that unstructured data will no longer go to waste. Instead, it can be used to create more complex and capable models than ever before, by combining tabular, text and vision-based data.

Abacus AI co-founder and Chief Executive Bindu Reddy (pictured, center) told SiliconANGLE that hybrid AI models are a logical next step, because it’s actually quite easy to use  unstructured data to train AI using modern deep learning techniques. The issue is that traditionally, most organizations use classical detection tree-based models that can only work with structured data.

Abacus AI has created new deep learning techniques that are able to extract signals from language-based and image-based data with relative ease. As a result, companies will now be able to build models that significantly outperform classical tree models, she said.

“Data science teams can now build complex models and extract powerful insights within hours by putting to use all the data they have, be it tabular, text, or vision,” Reddy explained. “For example, you can predict the closing price of homes based on unstructured data like listing description and house photos along with structured tabular data including the number of bedrooms, bathrooms, et cetera. This is done by combining all this data and using the Abacus.AI predictive workflow to generate a hybrid predictive model that combines all the data types. It’s a very powerful way to extract intelligence from all the available data.”

In addition, Abacus.AI is making it simple for companies of all sizes to implement computer vision use-cases such as image detection, classification and segmentation. The new computer vision-as-a-service offering adds to Abacus.AI’s existing support for predictive analytics, personalization, anomalies, forecast and language. Now users will be able to develop bespoke deep learning-based computer vision models too, Reddy said.

Looking ahead, Reddy said Abacus.AI pans to add support for additional common enterprise use cases in Vision AI and continue to invest in new neural network techniques for verticals such as personalization, forecasting, anomaly detection and language hybrid models.

“We are also expanding our areas of research to reinforcement and meta-learning,” she said. “Finally, we continue to add depth to our ML and DLOps modules that include a real-time feature store, model drift, observability and a vector matching engine.”

Photo: Abacus.AI