Uncovering big data trends

A new type of artificial intelligence eliminates human bias, allowing the technology to “see” previously unobserved information.

By Katharine Coggeshall & Gayle Geis-O'Dowd | July 26, 2021

Big Data Opt

Remember the Magic Eye books? Hidden within swirling patterns and splashes of color were eye-catching 3D images. Of course, the image was always there, but it took a special way of looking at the page to see what was “hidden.” Patterns in big data sets are no different. These patterns are buried in massive amounts of noise, less-important information, and intricate webs of interconnected variables. No human eye can see through the mess to identify a pattern, and human-guided software, such as traditional artificial intelligence (AI), is blind to the patterns as well.

As the world becomes awash with data and our scientific equipment records more detailed and complex information, many modern data sets sit unexplored. Demand exists for a tool that can see valuable patterns, such as identifying cancerous mutations in the human genome. That’s where Los Alamos National Laboratory researchers Boian Alexandrov and Velimir (Monty) Vesselinov come in. They co-created the first unsupervised AI tool, which operates without human bias.

Demand exists for a tool that can see valuable patterns, such as identifying cancerous mutations in the human genome. 

In traditional AI, humans teach the software what to look for and how to interpret the data. “Traditional AI relies on human influences, such as data labeling, subject-matter-expert opinions, and physics assumptions,” Vesselinov says. This is both the foundation of AI and its biggest limitation.

The unsupervised—or free from human bias—tool that Alexandrov and Vesselinov created is named SmartTensors because it harnesses the power of mathematical tensors (algebraic expressions that describe the relationship between sets of objects) to transform millions upon millions of data bytes into bite-sized, manageable cubes of information.

Not only does SmartTensors compress the data, which is ideal for storage, but it hones in on hidden data features, stripping out the noise. Imagine an audio recording of a crowded party. You’ll hear a jumble of voices, music, and glasses clinking. Previously, extracting the individual voices from the cacophony would have been close to impossible. But with SmartTensors, individual “voices” can be extracted easily from the collected data.

Since its inception in 2014, SmartTensors has evolved into a suite of unsupervised AI tools that operate on multiple platforms (from personal laptops to supercomputers to quantum computers). Alexandrov and Vesselinov have applied SmartTensors to more than 30 fields already, and that number is growing as members of the public try the technology (just go to github.com and search for “tensordecompositions”). What will you find in your data?