Unifying AI and Geometry

We are closing the gap between groundbreaking mathematical techniques and algorithm design, to enable the deployment of robust AI in production

The Company


We believe that AI algorithms are missing out on groundbreaking mathematical techniques that are too complex to be used out of the box when deploying AI in production. Our team of AI researchers, software engineers, and seasoned entrepreneurs are driven by the mission to make a great product that removes the roadblocks to adopting the latest and greatest technology in industrial scenarios.


Technology


Under the hood, our product leverages the power of geometric computational engines to enable clients to design and deploy machine-learning systems that are “geometry aware” -- they take into account the intrinsic geometry in the data in order to deliver higher accuracy and robustness to the application.


bunny

Geometric computation is an ensemble of mathematical techniques that allows for efficient parameterization of geometric transformations. In the past decade, the Ogarantia team has been a key contributor to the advance of geometric computation for machine learning, with several papers published in the subject (for example, this and this) and multiple citations, becoming the catalist of a new AI research field. This is a great advantage that allows us to directly incorporate and test the latest ideas within our products


Product


Our software sits on the interface between computational infrastructure and ML algorithms, helping AI designers to achieve the required level of performance without the need for months of R&D.

Ogarantia software is compatible with the standard tools used in the development of AI systems, e.g., Tensorflow. It can be easily installed in any computing system that supports Docker containers -- For example, Windows, Linux, and Mac. This enables us to supply the large majority of AI designers out there.


Case Study (for embedded devices)


To the right of the flipping point (point 1), i.e., when the number of parameters in the neural network model is greater than 500K, Ogarantia consistently delivers higher accuracy performance. Modern neural networks operate in this region. At point 2, we have the sweet spot where Ogarantia achieves the same performance as standard Tensorflow with 90% less parameters. If one needs the absolutely highest performance, at Point 3, Ogarantia delivers extra 4.5 log-evidence, well above what is achieved by Tensorflow.

Our computing engine offers a new way, besides quantization and pruning, to obtain compact neural networks without sacrificing performance. This example highlights the benefits of using Ogarantia when targeting deployment in embedded devices, a task that requires minimizing the neural network footprint (size and power consumption). Click on the image for detailed information.