The Scattering Transform for Texture Synthesis and Molecular Generation
The scattering transform is a wavelet-based feed-forward network originally introduced by S. Mallat to improve our theoretical understanding of convolutional neural networks (CNNs). Like the front end of a CNN, it produces a latent representation of input signal through an alternating sequence of convolutions and non-linearities. Following Mallat's original paper, subsequent work has shown that this latent representation can be used to synthesize new input signals such as textures. In a somewhat orthogonal extension, there has also been a number of papers which have shown how to adapt the scattering transform to graph-structured data.
In my talk, I will present a new network which combines these two ideas and uses the graph scattering transform to generate new molecules with the intended application being drug discovery. In order to ensure that the molecules produced by our network satisfy the laws of chemistry and resemble actual drugs, we use a regularized autoencoder to learn a compressed representation of the scattering coefficients of each graph and a generative adversarial network (GAN) to produce new molecules directly from this compressed representation.
Michael Perlmutter is a Hedrick Assistant Adjunct Professor in the department of mathematics at the University of California, Los Angeles. Previously he has held postdoctoral positions in the department of Statistics and Operations Research at the University of North Carolina at Chapel Hill and in the department of Computational Math., Science and Engineering at Michigan State University. He earned his PHD in Mathematics from Purdue University.