#
**Structure and Learning**

# Overview

The role of models is to identify or encode patterns in data to approximate some real-world process. These models vary in complexity. Models make mistakes if they are too simple or if they are too complex. If they are too simple, they introduce bias by making erroneous assumptions that do not approximate the real-world process well enough. If models are too complex, they can capture more complex processes but may overfit to training data and thereby suffer high error due to variance. We need to find a model that minimises bias and variance simultaneously.

The role of structure in robot learning is to decrease the variance or generalisation error of a complex model and find a suitable trade-off while not introducing too much bias. Check out the papers below for examples where we combine structure and learning!

# Algorithmic and Physics Priors for Learning

Kloss, A., Bohg, J. __On Learning Heteroscedastic Noise Models within Differentiable Bayes Filters__*OpenReview*. September 2018.

Kloss, A., Schaal, S., Bohg, J.
__ Combining learned and analytical models for predicting action effects__*arXiv*, September 2017.

# Exploiting Rigid Motion for Segmentation

Shao, L., Shah, P., Dwaracherla, V., Bohg, J. __Motion-based Object Segmentation based on Dense RGB-D Scene Flow__*IEEE Robotics and Automation Letters*, 3(4):3797-3804, IEEE, IEEE/RSJ International Conference on Intelligent Robots and Systems, October 2018

Shao, L., Tian, Y., Bohg, J. __ClusterNet: Instance Segmentation in RGB-D Images__*arXiv*. July 2018.