New paper: The Incentives that Shape Behaviour
Jan 22 · 8 min read
Ryan Carey and Eric Langlois, introducing The Incentives that Shape Behaviour.
Machine learning algorithms are often highly effective, but it can be difficult to establish their safety and fairness. Typically, the properties of a machine learning system are established by testing. However, even if a system behaves safely in a testing environment, it may behave unsafely or unfairly when it is deployed. Alternatively, the properties of a model can be investigated by analysing input perturbations, individual decisions, or network activations, but this…

