Active Learning during Federated Learning for Object Detection

Convolutional Neural Networks (CNNs) are currently among the most successful machine-learning techniques for object detection. One weakness of CNNs is that they require many labelled examples in order to train a model. This gives problems when training a model on decentralized data, such as in federated learning, where labels may not be available. Training on decentralized data is preferable, due to the benefits in privacy, and decreases in central data storage. Active learning can solve the unlabeled data problem, by selecting a portion of the unlabeled data and labelling it with an oracle. This paper explores, implements and evaluates several schemes which use active learning to label images locally and then use federated learning to train a global object detection model. Analysis shows the schemes maintain average precision close to centralized learning for homogeneous data. A novel approach based on a chain of devices allows for increased precision, while decreasing communication costs. The paper shows feasibility of training object detection models with active and federated learning, bringing the benefits of federated learning to the field of object detection.