Zurich, Switzerland – Aug 22, 2023 – LatticeFlow, the leading Artificial Intelligence (AI) platform for improving model performance and safety, announced a groundbreaking approach to building and deploying the next generation of highly resilient AI systems in collaboration with the U.S. army.
Stemming from a three-year long strategic engagement between the two organizations, this initiative aims to elevate security measures and curtail safety risks to prevent unintended consequences. “LatticeFlow’s approach is instrumental in enabling military commanders and decision-makers to confidently deploy mission-critical AI systems.” – said Dr. Igor Linkov, Senior Technical Manager at the US Army Corps of Engineers. He added: “A distinctive feature of their technology is the ability to create resilient AI models, which ensures that mission critical objectives are met despite potential disruptions and mis-predictions during deployment. Dr. Linkov unveiled this novel approach at the international conference on Cyber Conflict hosted by the NATO Cooperative Cyber Defense Center of Excellence. With a formidable assembly of over 600 attendees, including government officials, elite military personnel, and leading tech luminaries from powerhouses such as the US Army, NATO, and armasuisse, the event emphasized the pivotal role of AI in shaping the future of modern military warfare. The cornerstone of this approach resides in the integration of LatticeFlow’s robust framework and toolset, designed to systematically diagnose AI models and reveal hidden blind spots to ensure utmost accuracy, reliability, and safety.
For example, an analysis done by LatticeFlow revealed that the performance of a state-of-the-art AI model for detecting airplanes drops to 75% when operating in urban environments – thus underscoring the reliability and trustworthiness of the AI model’s predictions in such environments (Fig. 2). Beyond enhancing the performance and safety of individual AI models in isolation, this approach enables decision-makers to assess the resilience of the entire AI system as a whole. Achieving this milestone involves modeling the interaction among multiple AI models within the AI system, as well as specifying mission-critical objectives.
Following this approach, LatticeFlow in collaboration with the U.S. Army showcased the ability to build resilient autonomous agents (Fig. 3) at the NATO conference . The results demonstrate that while individual models can appear to be reliable with only minor performance degradation during environmental disruptions, the entire system could undergo a dramatic degradation by 150x.
Dr. Petar Tsankov, CEO & Co-founder of LatticeFlow, commented on the result: “Our approach to introducing a systematic methodology and toolset for improving performance and safety of AI systems lays the foundation for building next generation AI systems we can all trust. The importance of this is truly pivotal, especially in the manufacturing, medical, and insurance sectors, where AI-driven predictions affect critical business operations and human lives, and any disruptions can lead to profound consequences.”
Earlier this year, LatticeFlow also launched the industry’s first AI assessment service following the latest ISO standards for data quality and model robustness. The new service allows clients to obtain a comprehensive and unbiased evaluation of their AI data and models, to manage risks by identifying and eliminating gaps in data quality and model performance.
If you want to understand how LatticeFlow can help your ML teams deploy production AI models faster with greater security and reliability, check out our model diagnostics product.
