HeyDonto DFT Labs launches for physics-based machine learning
HeyDonto created DFT Labs after a peer-reviewed paper, but early tests show a gap between synthetic and real-world data performance.
By Colin Brandt · Enterprise Reporter
· 3 min read
HeyDonto DFT Labs is the new research subsidiary of artificial intelligence startup HeyDonto AI Technology, formed to develop a physics-based approach to machine learning. The company did not disclose funding, headcount, revenue or customer figures for the unit, making the launch a research and positioning move rather than a reported financing or commercial expansion.
The subsidiary is built around Data Field Theory, a framework described in a peer-reviewed paper published in Frontiers in Big Data. The paper, titled “Data Field Theory: A Geometric Framework for Learning on Riemannian Manifolds with Synthetic Validation and Limitation Analysis,” was authored by Reza Nehzati, founder and chief scientific officer of DFT Labs.
HeyDonto says the framework treats learning as a continuous field on a Riemannian manifold, using mathematical concepts associated with curved geometry and physical fields. Those tools are common in physics work involving systems such as magnets and superconductors. In machine learning terms, the company is arguing that relationships among data points can be modeled as a structure rather than treated only as isolated samples.
What is HeyDonto DFT Labs?
DFT Labs is HeyDonto’s research arm for Data Field Theory, an attempt to apply geometry and field-based physics concepts to machine learning. The practical goal, according to the company, is to build methods that can improve AI systems by identifying useful relationships within data.
Rivers Morrell, HeyDonto’s chief executive, told SiliconANGLE that the company is focused on the relationships among data points and sees intelligence as emerging from those relationships. He compared the idea to studying the structure of a flock rather than examining individual birds one by one.
The early evidence is mixed. In the paper’s reported tests, the DFT mathematics reached 89.2% accuracy on data classified with a manifold model and beat the comparison methods used in that experiment. The same approach was far weaker on MNIST, the standard handwritten digit dataset used in image-recognition testing. After the digits were converted into data points, DFT classified 15.7% correctly, slightly above the 10% expected from random guessing. A nearest-neighbor method reached 51.7%.
That split is the main technical issue for DFT Labs. The results indicate that DFT performs better when the dataset matches its geometric assumptions, and much worse when used on real-world data where the underlying geometry is not known beforehand. The paper identifies a next step: DFT needs to discover or learn the appropriate geometry from the dataset itself.
Where HeyDonto says the research is already being used
HeyDonto says work from DFT Labs is already part of Axiomera, its semantic-intelligence platform for converting fragmented enterprise data into standardized, AI-ready information. The same platform supports Conduit, the company’s dental and medical interoperability app, and Quantara, a clinical intelligence application used in cancer research.
Quantara’s work includes organizing pathology, imaging and molecular data for clinical use, according to the company. HeyDonto executives also pointed to possible administrative and pharmaceutical uses, including inventory management, rebate analysis and data harmonization between research institutions.
Kristopher Fuhr, president of Quantara, told SiliconANGLE that faster and more accurate biomarker identification algorithms could help cancer patients receive appropriate treatment sooner. That remains a company claim rather than an independently reported clinical result in the announcement.
Morrell also said HeyDonto wants to apply DFT Labs’ research to large language model architecture and eventually build a foundational model that competes with major AI model developers, including Anthropic. The company said it expects to release benchmarks from industry and academic competition entries in about 12 months. Until those results are public, DFT Labs is best read as an ambitious research bet with limited real-world validation disclosed so far.
This story draws on original reporting from SiliconANGLE.
