Over the last decade, disruptive developments in the field of machine learning have revolutionized crucial tasks such as pattern classification and nonlinear prediction. Implementing neural networks in specific hardware is highly desired in a quest for better speed, cost, and adaptability. Photonic implementations of neural networks have recently been gaining considerable attention in this context. Realizing photonic networks with many nonlinear nodes in parallel and including learning abilities has been lacking so far. In this seminar, we present a network of up to 2500 diffractively coupled photonic nodes, creating a recurrent neural network.Additionally, we demonstrate reinforcement learning in our neural network using a digital micromirror device. Under these conditions the system efficiently reaches solutions providing low errors. Our scheme is fully parallel and offers attractive future perspectives.