Federated Learning- The Perfect Amalgamation of IoT and AI!
One regular Tuesday morning I was looking for new papers on edge Internet of Things networks, like we all do, and came across a new term called Federated Learning and immediately went on a quest to understand this technique.
What is Federated Learning?
Federated Learning(FL) in its simplest form is a union between the vivacious worlds of Machine Learning and the Internet of Things. It is an ML-based solution that improves the functionality of edge devices in IoT networks.
Edge devices in an Internet of Things network are the mobile devices that collect data and help in achieving the purpose of the network remotely instead of the traditionally connected nodes in a network. Our smartphones are the most common example of edge devices.
According to the Google AI Blog from 2017- “Federated Learning enables mobile phones to collaboratively learn a shared prediction model while keeping all the training data on the device, decoupling the ability to do machine learning from the need to store the data in the cloud.”
But Why Do We Need FL?
Machine Learning’s biggest disadvantage in the age of edge devices is the necessity for large storage space and computational power to train humungous datasets. It is not feasible for edge devices in a widely spread IoT network to house that amount of data and to train them all individually. On the other hand, collecting data from a number of such edge devices and training it in one place defeats the purpose of decentralisation.
Thus Federated Learning exists as a perfect amalgamation of the best of both worlds- dynamic learning in edge devices while guaranteeing the decentralisation of the devices.
Now that the problem of decentralised learning of edge devices is solved, what about the need for the entire network to be aware of the updates to these edge devices?
The solution to this is for the edge device to periodically contact the other nodes and the central node in the network to inform of the update. This ensures that all nodes in the network are kept in the same loop and updated without any delay.
The Challenges in FL
With all its advantage FL does have some shortcomings in its impplementations- at least in its current state of existence.
- Privacy- the obvious issue with decentralising data in FL is the availability of data to every edge network and the potential misuse of the same. There is a constant sharing of model updates happening across the network which is susceptible to hacks and implementing encryption and other security measure might cause more bandwidth wastage.
- Causes bottlenecks- while FL solves the issue of space in terms of communication it uses up a larger bandwidth through the communication of model updates back and forth. This can lead to bottlenecks in the bandwidth for communication and increase its expenses too.
- Heterogeneous devices- one disadvantage of having decentralsied edge devices is the heterogeneity in their basic properties such as computational and communications prices. To tackle this we need to look into IoT based personalisation to mitigate the heterogeneity.
Therefore, finding a way for the nodes to communicate the model updates int he shortest, least expensive and secure way is a priority in the advancement of FL in edge based IoT networks.
Federated learning will alter the face of AI and IoT soon enough and make Machine Learning services available at our fingertips. It is a revolutionary step in the field of IoT as the decentralised learning simplifies several complications for larger networks and reduce costs of physical communication techniques.
I for one am very excited to follow the growth of this revolutionary technology and use its benefits to the best, hope you are too!
Here are some sources if you want to read more about FL:
If you liked this blog and would like to read more, do leave claps and share the blog!
