In order to fully integrate deep learning into robotics, it is important that deep learning systems can reliably estimate the uncertainty in their predictions. This would allow robots to treat a deep neural network like any other sensor, and use the established Bayesian techniques to fuse the network’s predictions with prior knowledge or other sensor measurements or to accumulate information over time.
Deep learning systems, e.g. for classification or detection, typically return scores from their softmax layers that are proportional to the system’s confidence, but are not calibrated probabilities, and therefore not useable in a Bayesian sensor fusion framework.
Current approaches towards uncertainty estimation for deep learning are calibration techniques or Bayesian deep learning with approximations such as Monte Carlo Dropout or ensemble methods.
PhD topics in this area can focus on reliably extracting uncertainty using Bayesian Deep Learning approaches for the specific use case of object detection on a robot in open-set conditions, and using the uncertainty information to actively accumulate new knowledge about the environment, e.g. by asking a human for ground truth labels (active continuous learning)