Dr. Vishal Sikka, Founder and CEO of Vianai, took the stage for a keynote presentation during Oracle CloudWorld 2022.
Fittingly titled Human-Centered AI: Unleash the Breakthrough Potential of AI in Your Organization, Dr. Sikka’s presentation addressed the challenges companies face as they seek to drive real business outcomes through AI adoption, and deployment at scale. The problems Dr. Sikka sees include ML monitoring tools that are not user-friendly (and expensive), and the risks associated with machine-learning model drift, uncertainty and other issues, potentially becoming damaging for companies.
Vian H+AI Platform | Human-Centered AI
To help define human-centered AI, Dr. Sikka first outlined the Vian platform and its ability to monitor machine-learning models at a massive scale. Because the platform can handle such high volumes of data, the most important things for the platform to address are bias and fairness, and ultimately make the data actionable. The three main goals of an effective AI platform should be Causality, Monitoring, and Performance.
He showed the user experience in the Vian H+AI Platform to illustrate that technology doesn’t need to be complicated to be effective. Between easy-to-digest graphs and models, the platform is designed to support the people behind the models, empower them to make insightful decisions, and monitor models in real-time, to ensure they are still doing what they are supposed to.
Vianai also tested its capabilities on a semiconductor system with 10X inference acceleration, while reducing the size by 27X.
Shrinking ML and Acceleration
The Vian team has been hard at work ensuring their software can monitor models that handle thousands of events per second and successfully monitored AI models over 95 billion transactions in a quarter. The sheer scale of customized monitoring for potential risks and bias is incredibly powerful.
Another important dimension of Vianai’s platform work, is in accelerating and shrinking AI models. AI today has a voracious appetite for computing, on both performance and cost. Vianai has made great strides with acceleration on vision and tabular models, regardless of its running hardware. For example, Vianai used its technology on large vision models for autonomous driving and managed to create a 4x improvement in inference speed, while shrinking the size of the ML model to ⅕ of the original. Vianai also tested its capabilities on a large vision model used for semiconductor data, and achieved 10X inference acceleration, while reducing the size by 27X.
- 4X Autonomous vehicle inference speed
- 20X Semiconductor System Inference Speed
With models becoming increasingly larger (GPT has 175 billion parameters; Google’s PaLM released a 540 billion parameter model), performance acceleration and reducing footprints of models will become critical. Dr. Sikka showed examples where ML models on tabular data were accelerated by more than 100x, and the footprint shrunk by 300x.
Shrinking ML models is especially important for edge hardware. Edge machines such as computers sitting on oil rigs or hardware in a new car require smaller ML models, and Vianai is capable of shrinking the models to fit on these platforms while accelerating their predictive maintenance models in some cases by more than 100x.
As Dr. Sikka concluded his remarks, he made two big announcements:
- Limited availability technology preview enabling users to upload models and test performance acceleration for free (running exclusively on the Oracle cloud). https://www.vian.ai/ocw-trusted/
- New ML model monitoring program enabling data science and ML engineering teams to see how monitoring can plug into their existing landscapes, extend tools they already have and bring critical, new monitoring capabilities including in-depth monitoring plans for all models, and automatic retraining capabilities. https://www.vian.ai/ccp/