Machine Learning: New method enables accurate extrapolation
To ensure the safe operation of a robot, it is crucial to know how the robot reacts under different conditions. But how do you know what disturbs a robot without actually damaging it? The machine learning method developed by scientists from the Institute of Science and Technology Austria (IST Austria) and the Max Planck Institute for Intelligent Systems uses observations made under safe conditions to make accurate predictions for all possible conditions determined by the same physical dynamics. The method is specially developed for real situations and offers simple, interpretable descriptions of the underlying physics.
Traditionally, machine learning can only interpolate data – that is, make predictions about a situation that lies “between” other, known situations. Machine learning could not extrapolate – that is, it could not make predictions about situations outside the known situations, since it only learns to model known data locally as accurately as possible. Collecting enough data for effective interpolation is also time and resource intensive and requires data from extreme or dangerous situations. Georg Martius, former postdoc at IST Austria and group leader at the MPI for Intelligent Systems in Tübingen, Subham S. Sahoo, a PhD student at the MPI for Intelligent Systems, and Christoph Lampert, professor at IST Austria, have now developed a machine learning method that addresses these problems. This method can for the first time perform precise extrapolations for unknown situations.
The special thing about the new method is that it tries to find out the true dynamics of the situation: Based on the data, it draws conclusions and computes equations describing the underlying physics. “If you know these equations,” says Georg Martius, “you can say what will happen in all situations, even if you have not seen them.” This is what enables the method to extrapolate reliably and makes it unique among machine learning methods.
Next: What makes this method so unique
The method is unique in many ways. First, the solutions previously created by machine learning were far too complex for human beings to understand. The equations resulting from the new method are much simpler: “The equations of our method are something you would see in a textbook – simple and intuitive,” says Lampert. The latter is another advantage: other machine learning methods do not give any insight into the connection between inputs and results – and thus no insight into whether the model is plausible at all. “In all other research areas we expect models that make sense physically and tell us why,” adds Lampert. “We should expect that from machine learning and that is what our method offers.” Therefore, the team based its learning method on a simpler architecture than usual methods to ensure interpretability and to optimize it for physical situations. In practice, this means that less data is needed to achieve the same or even better results.
And it’s not all theory: “In my group we’re working on developing a robot that uses this kind of learning. In the future, the robot would experiment with different movements and then be able to find the equations that describe its body and its movement so that it can avoid dangerous actions or situations,” says Martius. While the main focus of research is on robotic applications, the method can be used with any type of data – from biological systems to X-ray transfer energies. It can also be integrated into larger machine learning networks.
The researchers present their findings at this year’s International Conference for Machine Learning (ICML) that currently (July 9 through July 15, 2018) takes place in Stockholm.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
