The emerging market for AI is dominated by tech giants that offer cloud-based AI solutions and APIs to users in exchange for little control over the usage of AI products and their own data. If this continues, it will only lead to monopolization of the market by these tech giants. This in turn will only result in reduced participation of other companies in AI innovation.
However, one of the AI predictions for 2020 is the emergence of AI democratization.
Machines Learning (ML) models need training data to be centralized on one machine or data center. A new approach is being explored as an alternative to data centralization – federated machine learning.
Federated machine learning or simple federated learning is also known as collaborative learning. It’s a type of machine learning technique where ML models are distributed over different devices for computation, instead of being computed on large, centralized machines. The result is a faster deployment and testing of smarter models, lower latency, and less power consumption, all the while ensuring privacy.
Image source: Google
Federated machine learning is a technique for training an ML model where devices are enabled to learn collaboratively from a shared model. This shared model is first trained on a server using proxy data. Once this is complete, each device then downloads the model and enhances it using decentralized data from the device.
The device training is done using locally available data. The changes that are made to. The models are summarized and synced with the cloud, though the training data and individual updates remain on the device. Now, in order to speed up the updates, the model is compressed using quantization. After the devices send their specific updated models to the cloud server, the models are averaged. This way a single combined model is created. This process occurs are multiple iterations until an optimized model is achieved.
There are plenty of use cases and applications of federated learning, particularly those with privacy issues. Currently, federated learning is implemented largely on smartphones. Here are some of the popular use cases.
This was one of the first large-scale deployment of federated learning. It was a part of Google’s keyboard application – Gboard. The goal of the project was to improve word suggestions without compromising user privacy. As the ML model is placed on the user devices, it is able to constantly learn from what the users are typing. The summarized key information is then sent back to Google servers. These summaries are then used to enhance Google’s predictive text feature, which is then tested and pushed out to users.
The complexity of data privacy and security and incredibly high in the healthcare industry. Organizations in the healthcare industry have large volumes of sensitive patient information. Due to the high volume of personal data, most countries have stringent laws on how healthcare data should be managed and used. Owing to the complex nature of the legalities surrounding the data, AI companies are exploring federated learning as a possible solution for AI innovation in healthcare.
Federated learning can be useful for self-driving vehicles to protect privacy or user data and potentially reduce latency. Traditional cloud-learning involves transfers of large volumes of data but slower learning capacity. Federated learning enables autonomous vehicles to act more quickly and accurately, thereby reducing accidents and improving passenger safe.
One of the biggest limitations of federated learning is the computation capacity of the device. When compared to traditional learning methods, federated learning requires more local device power and memory to train the ML model. However, new devices sport sufficient power to handle these functions. Since federated learning requires the transfer of only smaller amounts of data to central servers, it reduces data usage as well. Users don’t seem to mind this trade-off as long as the device is powerful enough to handle the federated learning of ML models.
Another technical limitation is bandwidth. Unlike traditional ML that occurs in data centers, federated learning occurs over WiFi or 4G. Hence the bandwidth rates are magnitudes lower. Bandwidth to devices hasn’t grown at the pace of their computational power. This causes latency and slows the learning process.
The third challenge is that devices can be dropped out or undergo some kind of disruption or the other during the training process. In such cases, the data of dropped out devices may not be usable resulting in inaccurate algorithmic model.
Federated learning is relatively a new and promising training model. It has quickly shown potential in various use cases with privacy complexities. This paradigm shift creates exciting opportunities for AI enthusiasts and organizations. It provides a new way of thinking about solving large-scale ML problems.
One of the first things that will change is the development of the model, its training, ad evaluation with no direct access to raw data. However, due to its shortcomings, federated machine learning might not replace traditional machine learning just yet. Since it is still in its early stages, the learning processes require more time and research to be developed. Another factor that will determine the pace at which this kind of machine learning will evolve is the level of commitment to privacy by tech companies. Though it is hard to say how rapidly progress will be made, we believe that the benefits of federated learning make tackling technical challenges worth the effort.
Get a free copy of our whitepaper.