Beyond GANs: Transforming without a Target DistributionDownload PDF

25 Sept 2019 (modified: 05 May 2023)ICLR 2020 Conference Blind SubmissionReaders: Everyone
Original Pdf: pdf
Keywords: GAN, domain transfer, computational biology, latent space manipulations
TL;DR: A method for learning a transformation between one pair of source/target datasets and applying it a separate source dataset for which there is no target dataset
Abstract: While generative neural networks can learn to transform a specific input dataset into a specific target dataset, they require having just such a paired set of input/output datasets. For instance, to fool the discriminator, a generative adversarial network (GAN) exclusively trained to transform images of black-haired *men* to blond-haired *men* would need to change gender-related characteristics as well as hair color when given images of black-haired *women* as input. This is problematic, as often it is possible to obtain *a* pair of (source, target) distributions but then have a second source distribution where the target distribution is unknown. The computational challenge is that generative models are good at generation within the manifold of the data that they are trained on. However, generating new samples outside of the manifold or extrapolating "out-of-sample" is a much harder problem that has been less well studied. To address this, we introduce a technique called *neuron editing* that learns how neurons encode an edit for a particular transformation in a latent space. We use an autoencoder to decompose the variation within the dataset into activations of different neurons and generate transformed data by defining an editing transformation on those neurons. By performing the transformation in a latent trained space, we encode fairly complex and non-linear transformations to the data with much simpler distribution shifts to the neuron's activations. Our technique is general and works on a wide variety of data domains and applications. We first demonstrate it on image transformations and then move to our two main biological applications: removal of batch artifacts representing unwanted noise and modeling the effect of drug treatments to predict synergy between drugs.
7 Replies