How Federated Learning Moves AI Closer To The Edge

Forbes

16-07-2025

Vikram Gupta is Chief Product Officer, SVP & GM of the IoT Processor Business Division at Synaptics, a leading EdgeAI semiconductor company.
In my previous articles, I explored how the rapid growth of edge AI is calling for a new class of AI-native compute platforms and how multimodal sensing—including vision and audio—is enabling more intuitive, context-aware user experiences.
These trends mark a decisive shift from centralized cloud processing to intelligent, personalized and privacy-conscious computing at the edge. Building on that foundation, the next frontier is not just how AI runs at the edge, but how it learns and evolves there.
This is where federated machine learning (FML) enters the picture.
Despite the ubiquity of AI in our everyday lives, the vast majority of AI model development and processing still happens in the cloud, often far away from where we interact with it. The approach has served us well, with centralized and powerful compute engines doing the heavy lifting involved in collecting data and training sophisticated learning models.
As AI proliferates and billions of connected devices generate data at the edge, traditional centralized model training is becoming increasingly impractical, constrained by privacy concerns, regulatory pressures and latency limitations.
At the same time, the push for more personalized, context-aware experiences is accelerating AI processing toward a fragmented landscape of "far edge" devices, such as smartwatches, wearables and industrial sensors, that rely on real-time, local understanding of their environments. This transformation is not only enabling today's intelligent experiences but also laying the groundwork for an entirely new class of cloud-agnostic, AI-driven applications—many of which we have yet to imagine—that will operate independently and become seamlessly woven into the fabric of everyday life.
This shift has given rise to the new paradigm known as federated machine learning. By enabling localized intelligence directly on the devices where data is created, FML introduces a wider range of more private, personalized and responsive alternatives to cloud-centric models. Realizing this vision means evolving the AI ecosystem from system architecture and silicon design to software tooling.
It also extends into how data is collected, used and protected. The need for centralized computing resources won't necessarily go away, but this broader range of synergistic processing approaches is being driven by a more federated future. One size does not fit all in the world of edge AI.
Meeting the need for broader and more diverse deployment of AI, the rise of FML allows systems to become progressively more intelligent and autonomous by using on-device data and sharing only encrypted model updates.
Devices that can benefit from FML are increasingly present in our everyday lives, such as smart home assistants learning speech patterns locally, wearables monitoring health metrics without cloud sync and industrial machines predicting failure based on unique deployment environments. Recent announcements from companies like Google and OpenAI point to a future where AI is moving beyond phones into a new generation of devices.
The evolution of devices such as extended reality (XR) wearables raises questions: Do these devices need a cloud connection? A phone tether? Or can they operate independently, or even coordinate locally through a hub?
FML introduces the idea of processing zones, which could range from on-device to near-edge aggregation hubs or the centralized cloud. This transition to the future of edge AI depends on flexible, multitiered intelligence.
Ecosystem Complexity At The Edge: Fragmentation, Tooling And Hardware Diversity
The adoption of AI at the edge is not without some unique challenges. Unlike the relatively structured centralized processing model of the current data center-centric approach, the edge is messy. It features different operating systems (such as RTOS, Linux and Android variants with proprietary firmware), chip architectures (such as Arm, RISC-V and x86) and AI toolkits. On top of that, many devices lack the optimal processing, memory or power for robust on-device inference—let alone training.
Tooling for deploying and updating models is fragmented, particularly at scale. FML doesn't scale unless tools and hardware converge around modularity, efficiency and openness.
The Chip Supplier's New Role: A Scalable, Neutral Enabler
As this federated future of AI unfolds, success will hinge on delivering flexible, scalable solutions that span silicon, software and tools capable of adapting to diverse devices and dynamic ecosystems. By embracing openness, efficiency and intelligent decentralization, companies can unlock the full potential of edge AI.
The shift toward distributed intelligence is redefining how we interact with technology, making it more private, responsive and relevant to real-world environments. Real progress in edge AI depends on open-source tools, accessible frameworks and broad collaboration across the ecosystem. By focusing on practical solutions and inclusive innovation, this transformation can bring smarter experiences closer to where they matter most.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?