In this talk I will start by briefly reviewing the application of deep-learning to the study of jets at collider experiments. With these applications the deep-learning tools are in many cases used as jet observables, therefore they should be invariant to the symmetries of jets. These include rotational and translational symmetries, but I'll also discuss invariance to infrared and collinear effects within the jets. I'll then introduce a technique called contrastive-learning which allows us to construct highly-expressive invariant observables using deep-learning. To demonstrate the power of this method I'll benchmark the observables in an application to top-tagging, and show how we can visualise some of the symmetry properties of the observables. Finally, I'll discuss potential future applications of this work. This is all based on the paper 'Symmetries, Safety, and Self-Supervision', hep-ph/2108.04253.