Combination of Federated and Active Learning for Self-Driving Cars

A distributed learning approach to solving data privacy and many other training challenges in automotive applications

Centralized learning is an approach to train machine learning models at one place, usually in the cloud, using aggregated training sets from all devices utilizing that model. Centralized learning has the advantage of generalizing the data collected from all devices and providing an optimized model. This approach poses the following challenges:

• Data privacy: In some industries or applications, such as medicine or finance, data is sensitive and must not be moved to another location
• Limited network capacity: Some network properties such as bandwidth and latency between clients and centralized model affect the training process performance
• Non-i.i.d. (independent and identically distributed) data distribution: the data collected belong to different environments, times, and clients that do not meet the i.i.d. property of data that is considered in most statistical analysis and ML algorithms as a required assumption.
• Central storage capacity: High storage requirements due to large amounts of data in automotive applications

Federated Learning (FL) solves these centralized learning challenges or drawbacks mentioned above. A key feature of federated learning is the protection of user data. This means that the data should be located where it is generated and not moved to another location for the training process. In other words, the algorithm should move where data is being generated. But another aspect, such as limited network capacity, is also critical when you have many clients to send large amounts of data. We can replace recorded raw data with model parameters or weights for machine learning and reduce the communication overhead.

The goal, however, is to communicate machine learning model parameters instead of data that needs to be secure. The new challenge for machine learning is defining the proper loss function and training the model in a distributed manner to get a converged and accurate model ultimately.

Autonomous driving is a new application for FL

The development of self-driving cars requires collecting large amounts of environmental data from various points such as ego vehicle data, data from another vehicle on the road, infrastructure sensor data, map data, etc. Distributed data can all be relevant for a vehicle function implemented by the machine learning algorithm. Federated learning can offer a decentralized approach to learning when centralized learning is not possible or not allowed. Vehicles as an edge device are responsible for training the local model and either sharing it with other cars or a central server can aggregate all models and create a final and optimized model and update all vehicles immediately or regularly.

Active Learning

Active learning is a learning algorithm that queries the user data and labels relevant data for the algorithm. Active learning helps reduce training costs through automated data labeling. The most critical challenge in active learning is selecting a part of the data relevant to the training. The aim is to assess the unlabelled data's informative value and pick only the informative part of the data. We need to develop a strategy to avoid bias in each client's data. Each client may have a different amount of data or contains environment-specific data that causes data biasing.

FADNet a new approach to machine learning in autonomous driving

A new peer-to-peer approach called FADNet, which is described in this paper, is a solution to address centralized learning deficits. This method does not require a central server for training, and each client trains its data, and the result is delivered to the next client through an aggregation process in a chain. This approach requires more computing power than centralized or server-based federated learning.

This paper shows that the approach converges much faster than server-based federated learning. The learning architecture adapted to overcome non-i.i.d. data and manage the high variance between client's data and find a trade-off between accuracy and the ability to converge. This model proposed in the publication was evaluated by visualizing three image data sets from driving scenarios.

Outlook

Using this solution has advantages such as:

• We can at least pre-label data locally before centralized training of the model when there is a need to achieve the accuracy of the final model, which is very costly and challenging for distributed data.
• We train the model locally, so the vehicle always uses an updated and optimized model.
• The local storage of user data guarantees data privacy
• No central data storage required
• Reduction of network overhead
• Use of the computing power of the vehicle and real-world data to improve the ML model

Combining federated and active learning can train the model with distributed vehicle or infrastructure sensor data. The idea is fascinating because a large amount of data is collected from vehicles or infrastructures every second.

Choosing the right distributed machine learning architecture resolves the network performance limitations. V2X (Vehicle to Everything) communication technology is required to exchange data between vehicles (vehicle as an edge device) or between the car and the edge server, depending on the selected federated learning approach.

Active learning and federated learning enable us to access real-world driving and parking scenarios that are not always easy to validate through simulation. This approach allows accurate self-driving car validation, as Tesla and Waymo likely began many years ago. This method is still challenging as vehicles are in motion, and data collected is stamped on time and could soon be out of date. Labeling data is also challenging and requires an automated data labeling process. This approach reduces the time and cost of labeling and improves the overall CI/CD pipeline and deployment of the machine learning model for self-driving cars functions.