Federated learning is a machine learning technique that allows multiple parties to train a model without sharing their data. It is being used across several industries, from mobile device keyboards to health care to autonomous vehicles to oil rigs. It is particularly useful in situations where data sharing is limited by regulation, or is sensitive or proprietary, as it allows organizations to collaborate on machine learning projects without sacrificing data privacy. It is also helpful in situations where data sizes are prohibitively large, making data centralization slow and costly.
One of the main obstacles in machine learning is the need for large amounts of data. This can be a challenge for organizations that do not have access to large datasets, or for those that are working with sensitive data that cannot be shared. Federated learning allows these organizations to contribute to a shared model without having to share their data.
Federated learning can also help to overcome the issue of data homogeneity. In many cases, models are trained on data from a small set of sources that do not represent the general population. Models trained on narrow datasets don’t generalize well and thus underperform when deployed more broadly. Federated learning allows training models on a larger and more diverse set of data sources without requiring the data from all of these data sources to be centralized, thus leading to more robust models with better performance.
Additionally, the cost of cloud compute resources can be an obstacle in machine learning. Training machine learning models can be computationally intensive, requiring expensive hardware like Graphical Processing Units (GPUs). Using cloud instances for training can become expensive very quickly. Federated learning allows organizations to share the load of model training and use under-utilized compute resources or servers that they already have in their data centers. This can lead to a significant cost savings in large compute-intensive training processes.
Many organizations are also concerned about creating redundant copies of large data sets. This can rack up high storage costs, as well as costs to cloud providers for transferring the data between on-prem data centers and cloud accounts, or between different cloud accounts. Federated learning allows organizations to maintain a single copy of their data and doesn’t require moving it to a different location or cloud account in order to train models with the data.
Another challenge that can limit the use of machine learning is privacy and regulatory constraints. The data used to train models may contain sensitive information such as Personally Identifiable Information (PII) or Personal Health Information (PHI). Federated learning allows organizations to train models without having to share their data, which can help to mitigate these privacy and regulatory concerns.
Federated learning is already being used across several industries in order to unlock the power of larger and more diverse datasets without data sharing. For example, in 2021 a COVID decision support algorithm was trained with data from 20 hospitals around the world using federated learning (full disclosure: this project was led by our co-founder and CEO), and in 2022 a brain cancer margin detection algorithm was trained with data from 71 hospitals around the world using. Google has been using federated learning to predict the next word typed on Google Android keyboards since 2018 (full disclosure: before co-founding my company, I worked at Google and was involved in projects utilizing federated learning).
In summary, federated learning is helping to overcome a number of obstacles in machine learning, including the need for large amounts of data, the cost of compute resources and data storage and transfer, the challenge of data homogeneity, and privacy and regulatory concerns. It allows organizations to collaborate on machine learning projects without sacrificing data privacy, democratizing the use of machine learning and access to large diverse training data, yielding more robust and better-performing models.