We introduce the ACTIONFINDER, a deep learning algorithm designed to transform a sample of phase-space measurements along orbits in a static potential into action and angle coordinates. The algorithm finds the mapping from positions and velocities to actions and angles in an unsupervised way, by using the fact that points along the same orbit have identical actions. Here we present the workings of the method and test it on simple axisymmetric models, comparing the derived actions to those generated with the Torus Mapping technique. We show that it recovers the torus actions for halo-type orbits in a realistic model of the Milky Way to ~0.6% accuracy with as few as 1024 input phase-space measurements. These actions are much better conserved along orbits than those estimated with the Stäckel fudge. In our case, the reciprocal mapping from actions and angles to positions and velocities can also be learned. One of the advantages of the ACTIONFINDER is that it does not require the underlying potential to be known in advance-indeed it is designed to return the acceleration field. We expect the algorithm to be useful for analyzing the properties of dynamical systems in numerical simulations. However, our ultimate goal with this effort will be to apply it to real stellar streams to recover the Galactic acceleration field in a way that is relatively agnostic about the underlying dark matter properties or the behavior of gravity.