FedML Raises $11.5 To Combine MLOps Tool With Decentralized AI Computing Network

Image credit: v_Alex/Getty Images

Interest in AI continues to grow among businesses, with one recent survey finding that nearly two-thirds of businesses plan to increase or maintain spending on AI and machine learning this year. . However, these companies often encounter obstacles in putting various forms of AI into production.

A 2020 poll from Rexer Analytics found that only 11% of AI models are constantly deployed. And one Gartner analyst estimates that nearly 85% of his big data projects fail.

Inspired by a challenge, Salman Avestimehr, the first director of the USC-Amazon Center on Trustworthy Machine Learning, co-founded a startup that enables companies to train, deploy, monitor, and improve AI models in the cloud or at the edge. bottom. The company, called FedML, is led by Camford Capital, with participation from Road Capital and Finality Capital, and has raised $11.5 million in seed funding at a valuation of $56.5 million.

“Many companies are keen to train and fine-tune custom AI models based on company-specific and industry data to use AI to meet different business needs,” says Avestimehr. told TechCrunch in an email interview. “Unfortunately, custom AI models are prohibitively expensive to build and maintain due to high data, cloud infrastructure, and engineering costs. They are often expensive, regulated or siled.”

FedML overcomes these barriers by providing a “collaborative” AI platform that enables companies and developers to work together on AI tasks by sharing data, models and computing resources, Ave said. Stimer argues.

FedML can run any number of custom AI models or models from the open source community. The platform allows customers to create groups of collaborators and automatically sync her AI applications across devices (e.g. PCs). Collaborators can add devices they use to train AI models, such as servers and mobile his devices, and track training progress in real time.

Recently, FedML published FedLLM, a training pipeline for building “domain-specific” large-scale language models (LLMs) based on your own data, similar to OpenAI’s GPT-4. FedLLM is compatible with popular his LLM libraries, such as Hugging Face and Microsoft’s DeepSpeed, and is designed to speed up custom AI development while maintaining security and privacy, Abestimer said. says Mr. (Don’t get me wrong, the jury hasn’t ruled on whether that’s exactly achievable.)

As such, FedML is not much different than any other MLOps platform out there. “MLOps” refers to tools that streamline the process of deploying, maintaining, and monitoring AI models into production. There are (to name a few) Galileo and Arize, but also Seldon, Qwak, Comet, and more. Incumbents such as AWS, Microsoft, and Google Cloud also provide some form of his MLOps tooling (see SageMaker, Azure Machine Learning, etc.).

But FedML has ambitions beyond developing AI and machine learning model tools.

In Avestimehr’s words, the goal is to build a “community” of CPU and GPU resources to host and serve deployment-ready models. Details are yet to be finalized, but FedML intends to encourage users to contribute computing to the platform through tokens and other types of rewards.

Decentralized distributed computing for serving AI models is not a new idea. Gensys, Run.AI, and Petals are among the companies that have tried and are still trying to do this. Nonetheless, Avestimehr believes that combining this paradigm of computing with his MLOps suite will allow FedML to achieve greater reach and success.

“FedML enables custom AI models by allowing developers and enterprises to build their own private LLMs at low cost and at scale,” said Avestimehr. “The uniqueness of FedML is that anywhere he can train, deploy, monitor and improve ML models, and collaborate on a combination of data, models and compute, significantly reducing costs and time to market.”

He noted that the 17-employee FedML has about 10 paying customers, including a “tier 1” auto supplier, and has a total of $13.5 million in funds under custody, including new funding. It is said that there is Avestimehr claims that the platform is used by more than 3,000 users worldwide and that he runs more than 8,500 training his jobs on more than 10,000 devices.

“For data or technical decision makers, FedML makes custom, affordable AI models and large-scale language models a reality,” Avetimer said confidently. “And thanks to a foundation of federated learning technologies, MLOps platforms, and collaborative AI tools that help developers train, serve, and observe custom models, building custom replacement models is an accessible best practice.”