AI applications based on Machine Learning are notoriously hard to verify because there is no definition of their expected behavior other than their input set. This makes it hard to put a ML based program in charge of a safety critical device such as a car. But it is possible to verify that the ML algorithm itself is correctly implemented. In the training stage, backpropagation has a defined effect on the parameters of the network. After training, it can be verified that the network produces the correct outcome (given knowledge of the network parameters) for a given stimulus.
We are interested to develop both the specification and a test suite for one of the well known ML libraries such as PyTorch and TensorFlow.
We offer:
- A pleasant working environment in Amsterdam
- Guidance when you need it
- An internship compensation