Last couple of years have seen an explosion in neural models being used to generate all sorts of data, from images to text, molecules and many others. The generated data can be constrained to resemble an existing dataset ('image of a handbag that looks like a zebra') and/or to maximize an explicit scoring function. The variety of methods used for this (GANs, encoder-decoder, VAEs, different reinforcement learning flavors) can seem bewildering at first. The tutorial will walk the audience through choosing the right overall approach based on the properties of the problem at hand (continuous or discrete domain? Do we have a scoring function? Is there an existing dataset the generated data must be 'similar' to? ), and through the specific architectural choices within each approach (optimization method, type of network used for each component, how to decode network output, etc).

Overall, the talk will give the audience a joined-up understanding of the landscape of neural generative models. The target audience are data science professionals with limited prior exposure to generative neural networks.