doc.ai releases open source technology that unlocks model building on mobile devices

Tensor/IO brings machine learning framework to ecosystem of smart devices

Our lives, like our health, are continuous functions. Our health is not just what’s assessed at a point in time during a doctor’s office visit, but all the data points that impact our wellness in-between. We can now capture and monitor these real world data points with smartphones, wearables and a growing number of at-home tests to evaluate your nutrition, microbiome or the impact of genetics on health. 

The amount of health data an individual can generate is growing, and collectively, could hold the secrets to better treatments and health outcomes. In fact, projections indicate that there could be as much as 2,314 exabytes of new data generated in 2020. Yet, having all of this data, without the proper tools to make sense of it, does not get us closer to a more personalized, 360 degree view of health.

We need AI and machine learning to help analyze this wave of health data, much of which is coming from smartphones. To facilitate our twin goals of respecting privacy while learning and creating models from health data, doc.ai has committed to researching and developing cutting edge privacy-preserving technologies such as edge AI and federated learning. 

If you’ve followed the doc.ai journey, you may have seen our previous presentations on these technologies at conferences like ICML and NeurIPS. Today, we’re excited to announce a new milestone in the step towards making federated learning ready for training on mobile devices.

doc.ai brings on-device training to Android with TensorFlow and Tensor/IO

Today, we released Tensor/IO with full support for Android. Tensor/IO is our machine learning framework for on-device inference and training with TensorFlow. For those unfamiliar with TensorFlow, it’s Google’s end-to-end open source machine learning platform.

While we had made this framework available for iOS devices in 2019, bringing Tensor/IO TensorFlow to Android brings Tensor/IO’s declarative interface and deployment system for full TensorFlow models to potentially hundreds of millions more additional devices. 

On top of Tensor/IO we have also developed Tensor/IO Federated, doc.ai’s proprietary federated learning system that involves client-server communication and federated task coordination. This means with our Tensor/IO advancements, doc.ai can now offer a fully cross-platform solution for prediction, training, and federated learning on any mobile phone. 

For example, with Tensor/IO you can deploy a model to a user’s device that distinguishes the foreground from the background in images, train it on images on the phone without ever needing to send those images back to your server side systems, and finally either redeploy the improved model to production immediately or collect the updated model from many devices for later aggregation in a federated training round.

Tensor/IO enables the TensorFlow community to:

  • Package and deploy full TensorFlow models for iOS and Android devices with a declarative interface and built-in data transformations.
  • Perform inference with full TensorFlow models on iOS and Android devices.
  • Perform training with full TensorFlow models on iOS and Android devices.

Before offering full TensorFlow support with Tensor/IO, we previously supported TensorFlow Lite. TensorFlow Lite is a capable machine learning library but only supports lightweight models and inference. With full TensorFlow support we can run much more complex models on mobile devices and perform both inference and training with them.

Tensor/IO for Android is open source under an Apache 2 license and is part of the broader Tensor/IO project, which is also open source under the same license. Learn more about this project in our technical blog and put the code to use yourself at https://github.com/doc-ai/tensorio

Understanding privacy-preserving algorithmic learning

Machine learning promises to revolutionize many industries, including healthcare, however its application is restricted to areas where there is enough data to train useful models. Often, the barriers to data acquisition are not technological but issues such as privacy, trust, regulatory compliance, and intellectual property. This is especially the case in healthcare, where patients and consumers expect privacy with respect to personal information and where organizations want to protect the value of their data and are also required to follow regulatory laws such as HIPAA in the United States and the GDPR in the EU.

Privacy ultimately starts at the technology layer of any healthcare product or experience. Our framework for decentralized machine learning is composed of multiple parts:

Edge AI is the idea that you can perform AI and specifically machine learning tasks on user devices rather than on server-side systems, for example, on your phone instead of in the cloud. 

Federated learning will take edge AI a step further and enable personalized health algorithms to actually learn from data on user devices without the data ever needing to leave the phone. With federated learning, we can perform the second fundamental machine learning task, training, on the device as well, rather than returning that data to the server. 

Zero Trust is a new security paradigm that believes the underlying network is not to be trusted, and therefore needs verification before any interaction takes place. Services must cryptographically establish trust with each other before communication is allowed.

Together these technologies, edge AI and federated learning in a zero trust environment, constitute the foundation for privacy-preserving algorithmic learning. Our next challenge is to apply differential privacy and multi-party encryption, specifically homomorphic encryption, to begin training algorithms securely via mobile devices.

With the release of Tensor/IO for Android doc.ai is taking another step in its commitment to data privacy and to keeping personal data ownership where it belongs: in the individual’s hands. 

Upcoming webinar on Tensor/IO 

On Thursday, Jan. 21 at 9am PT, Philip Dow, Federated Learning engineer at doc.ai will be holding a webinar to discuss and demonstrate the capabilities of Tensor/IO. Register today at doc.ai/federated learning webinar.

Baseline version

 

Leave a Comment