Emerging AI Business Model Promotes Distributed ML
Pausing at the end of 2019 to take stock of where enterprise AI stands, a new business model floated over the last year encourages distributed model training through the sharing of data.
Depending on the project, the critical question raised by the emerging AI model dubbed “federated learning” is whether corporate rivals would ever be willing to share data? On the flip side, a distributed machine learning framework could create an opening for startups, ultimately improving AI applications.
AI consultant Alexandre Gonfalonieri makes the case that federated learning—that, is a distributed machine learning framework used to build a “collective model” shared among different data owners—represents a viable AI business model.
Some tech vendors have already embraced the federated approach for medical applications that bring with them difficult data privacy issues. For example, Nvidia (NASDAQ: NVDA) this week unveiled a federated learning tool as part of its Clara AI platform designed to promote collaboration while protecting patient privacy.
Aimed at the medical imaging community, Nvidia said its federated approach would enable distributed clients to train deep learning models locally via edge servers, then collaborate to develop accurate “global models” that can be used by clinicians.
Indeed, Gonfalonieri argues that even in heavily regulated markets like medical science, a distributed machine learning approach could yield improved models and compelling AI applications. “In the medical field, [federated learning] could be a synonym for better treatment and faster drug discovery,” he argued.
Given recent setbacks such as IBM’s (NYSE: IBM) recent withdrawal of its Watson AI drug discovery tools, critics assert that the retreat highlights AI’s shortcomings. Still, some observers note that machine learning frameworks show promise in areas like diagnostic imaging where “confirmation bias” and other prejudices remain an issue.
“Centralized [machine learning] is far from being perfect,” Gonfalonieri concluded. “Indeed, training the models requires companies to amass mountains of relevant data to central servers or datacenters. In some projects, that means collecting a user’s sensitive data.”
It is the sharing of data—either competitive or regulated—that appears to be the largest stumbling block to advancing distributed machine learning as a future AI business model. In some respects, the federated approach is akin to established open source software and hardware development that have been widely embraced by hyper-scalers.
Others have proposed federated learning approaches, including frameworks in which model development is distributed among millions of edge devices. Proponents argue they can ensure data privacy by restricting access to or labeling raw user data.
“We are seeing the beginning of a decentralized AI market, born at the intersection of on-device AI, blockchain, and edge computing” and the Internet of Things, wrote Santanu Bhattacharya, an investor in the federated learning startup S20.ai.
Even AI leaders like Google (NASDAQ: GOOGL) have promoted federated learning as a way of building smarter models while maintaining data privacy. Google has tested the approach in scenarios where mobile phones are able to collaboratively learn a shared prediction model while all training data is stored locally.
Reactions to the distributed machine learning framework have been positive, albeit cautious. Responding to AI consultant Gonfalonieri’s blog post, one reader warned that federated learning would encourage data storage on devices, perhaps creating a new cottage industry for hackers selling tools to access local data.
“The security [and] privacy threat models are quite real,” noted one reader. “It’s important to solve these” along with other data privacy issues.