The emerging AI market model is dominated by tech giants such as Google, Amazon and Microsoft that offer cloud-based AI solutions and APIs. This model offers users little control over the usage of AI products and their own data that is collected from their devices, locations, etc. In the long run, such a centralised model is not good for the society or the market, as it could lead to monopolisation of only a few strong players. Eventually, it would limit the participation of smaller companies or even larger enterprises in AI innovation, as well as result in lack of interoperability and interpretability of decisions driven by AI systems.
Luckily, as the spring of AI emerges in 2019, we are seeing the beginning of a decentralised AI market, born at the intersection of on-device AI, blockchain, and edge computing/IoT.
Standard machine learning models require centralising of the training data on one machine or in a data centre. For example, when an ecommerce startup wants to develop a model to understand its consumers’ propensity to purchase products, it runs the models on the data collected from its website or app. Such data may include the time spent on a particular product page, products bought together, products browsed but purchased, etc. Typically, up to thousands of data points are collected on every user over a period of time. Such data are parsed and sent over to a centralised data centre or machines for computation.
Recently, a new approach was considered for machine learning models trained from user interactions with mobile devices: it is called Federated Learning.
While the comparisons may be somewhat simplistic, the history of computing may be a decent proxy to what federated learning is all about. In the early days of information technology, we had large mainframes doing the heavy lifting of most of the computing. Eventually, we moved to a client-server framework, where the computes were distributed between central server(s) and multiple client computers.
The federated learning architecture deploys a similar model. Machine learning models, instead of being computed on large, centralised machines, are distributed over mobile devices for computation. This model of computing, while being theoretically possible, would not have been practical in the past, since computational abilities of mobile phones were very limited for running any ML model.
However, something changed in the mid-to-late 2018s. As billion-plus smartphones equipped with AI chips and significant computing power, starting with Samsung S9, or Apple X series, get shipped in the next three to five years, many of the ML models will be able to run locally on these mobile devices.
Functionally, a mobile device that is a part of a federated learning computing architecture downloads a model that is meant for running on mobile devices. It then runs the model locally on the phone and improves it by learning from data stored there. Subsequently, it summarises the changes as a small update, typically containing the model parameters and corresponding weights.
The update to the model is then sent to the cloud or central server using encrypted communication, for example, homomorphic encryption (HE). This update is then averaged with other user updates to improve the shared model. Most importantly, all the training data remains on a user’s device, and no individual update is identifiably stored in the cloud.
Federated learning allows for faster deployment and testing of smarter models, lower latency, and less power consumption, all while ensuring privacy. Also, in addition to providing an update to the shared model, the improved (local) model on your phone can also be used immediately, powering experiences personalised by the way you use your phone.
In the next few years, model building and computation on the edge, based on federated learning and secured with homomorphic encryption, will make significant progress. As one billion-plus smartphones equipped with AI chips and possessing significant computing power get into the market in the next three to five years, many of the ML models will be able to run locally on these mobile devices. Distributing the heavy-duty analytics and computations over smartphones “on the edge”, as opposed to central computing facilities, will drastically reduce time to develop data products such as hyper-personalised recommendation engines, e-commerce pricing engines etc. Enterprises will embrace a distributed machine learning model building framework for taking advantage of faster model deployment and to provide quicker response to fast-changing consumer behaviour, besides a vastly reduced cost.
For machine learning practitioners and enthusiasts, this paradigm shift provides an exciting opportunity to democratise AI. It also opens up new avenues for adopting new tools, and most importantly, a new way of thinking about solving large-scale ML problems.
Model development, training and evaluation with no direct access to or labelling of raw data will be challenging at first. However, in emerging markets such as India, where hyper-personalisation and highly contextual recommendation engines will be key for driving, say, app adoption, or e-commerce purchase, the bet is that federated learning will play a key role in the future. The user benefits of federated learning make tackling the technical challenges worthwhile.